WO2022009301A1 - Image processing device, image processing method, and program - Google Patents

Image processing device, image processing method, and program Download PDF

Info

Publication number
WO2022009301A1
WO2022009301A1 PCT/JP2020/026534 JP2020026534W WO2022009301A1 WO 2022009301 A1 WO2022009301 A1 WO 2022009301A1 JP 2020026534 W JP2020026534 W JP 2020026534W WO 2022009301 A1 WO2022009301 A1 WO 2022009301A1
Authority
WO
WIPO (PCT)
Prior art keywords
query
person
image
posture
search
Prior art date
Application number
PCT/JP2020/026534
Other languages
French (fr)
Japanese (ja)
Inventor
雅冬 潘
登 吉田
諒 川合
健全 劉
祥治 西村
Original Assignee
日本電気株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 日本電気株式会社 filed Critical 日本電気株式会社
Priority to PCT/JP2020/026534 priority Critical patent/WO2022009301A1/en
Priority to JP2022534527A priority patent/JP7416252B2/en
Publication of WO2022009301A1 publication Critical patent/WO2022009301A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying

Definitions

  • the present invention relates to an image processing device, an image processing method, and a program.
  • Patent Documents 1 and 2 are known.
  • Patent Document 1 discloses a technique for searching the posture of a similar person based on key joints such as the head and limbs of the person included in the depth image.
  • Patent Document 2 discloses a technique for searching for a similar image by using posture information such as tilt added to the image, although it is not related to the posture of the person.
  • Patent Document 3 discloses a technique of acquiring posture information of a search target composed of a plurality of feature points from an image and searching for an image including a posture similar to the posture specified by the posture information.
  • Non-Patent Document 1 is known as a technique related to the estimation of the skeleton of a person.
  • Patent Documents 1 and 3 are techniques for searching for an image including a person in a predetermined state, but since the conditions that can be specified as a search query are limited, the searchable targets are also limited. That is, the techniques of Patent Documents 1 and 3 have room for improvement in the input method.
  • Patent Document 2 is not a technique for searching an image including a person in a predetermined state.
  • An object of the present invention is to expand the range of searchable objects in a system for searching an image including a person in a predetermined state.
  • a first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people
  • a second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
  • a search means for executing an image search based on the first query and the second query, and An image processing device comprising the above is provided.
  • the computer Get the first query that specifies the posture of each of the multiple people Obtain a second query that specifies at least one of the physical distance and the temporal distance between the plurality of people who take the specified posture.
  • An image processing method for executing an image search based on the first query and the second query is provided.
  • a first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people
  • a second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
  • a search means for performing an image search based on the first query and the second query.
  • a program is provided that functions as.
  • the range of searchable objects can be expanded.
  • FIG. 1 It is a block diagram which shows the outline of the image processing apparatus which concerns on embodiment. It is a block diagram which shows the structure of the image processing apparatus which concerns on Embodiment 1.
  • FIG. It is a flowchart which shows the image processing method which concerns on Embodiment 1. It is a flowchart which shows the classification method which concerns on Embodiment 1. It is a flowchart which shows the search method which concerns on Embodiment 1.
  • FIG. It is a figure which shows the human body model which concerns on Embodiment 1. It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1.
  • FIG. 1 It is a block diagram which shows the outline of the image processing apparatus which concerns on embodiment. It is a block diagram which shows the structure of the image processing apparatus which concerns on Embodiment 1.
  • FIG. It is a flowchart which shows the image processing method which concerns on Embodi
  • FIG. It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1.
  • FIG. It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1.
  • FIG. It is a graph which shows the specific example of the classification method which concerns on Embodiment 1.
  • FIG. It is a figure for demonstrating the search method which concerns on Embodiment 1.
  • the inventors examined a method of using a skeleton estimation technique such as Non-Patent Document 1 in order to recognize the state of a person desired by a user from an image on demand.
  • a related skeleton estimation technique such as OpenPose disclosed in Non-Patent Document 1
  • the skeleton of a person is estimated by learning image data with various patterns correctly answered.
  • the skeletal structure estimated by a skeletal estimation technique is composed of "key points” which are characteristic points of joints and the like and “bones (bone links)" which indicate links between key points. .. Therefore, in the following embodiments, the skeletal structure will be described using the terms “key point” and “bone”, but unless otherwise specified, the "key point” corresponds to the “joint” of a person, and " “Bone” corresponds to the "bone” of a person.
  • FIG. 1 shows an outline of the image processing apparatus 10 according to the embodiment.
  • the image processing device 10 includes a skeleton detection unit 11, a feature amount calculation unit 12, and a recognition unit 13.
  • the skeleton detection unit 11 detects a two-dimensional skeleton structure (hereinafter, may be simply referred to as “skeleton structure”) of a plurality of persons based on a two-dimensional image acquired from a camera or the like.
  • the feature amount calculation unit 12 calculates the feature amount of the plurality of two-dimensional skeleton structures detected by the skeleton detection unit 11.
  • the recognition unit 13 performs a recognition process for the states of a plurality of persons based on the similarity of the plurality of feature amounts calculated by the feature amount calculation unit 12.
  • the recognition process is a classification process of a person's state, a search process, or the like.
  • the two-dimensional skeleton structure of the person is detected from the two-dimensional image, and the recognition process such as classification and search of the state of the person is performed based on the feature amount calculated from the two-dimensional skeleton structure. ..
  • FIG. 2 shows the configuration of the image processing apparatus 100 according to the present embodiment.
  • the image processing device 100 constitutes an image processing system 1 together with a camera 200 and a storage means (database (DB) 110).
  • the image processing system 1 including the image processing device 100 is a system for classifying and searching states such as posture and behavior of a person based on the skeleton structure of the person estimated from the image.
  • the camera 200 is an image pickup unit such as a surveillance camera that generates a two-dimensional image.
  • the camera 200 is installed at a predetermined location and captures a person or the like in the imaging region from the installation location.
  • the camera 200 is directly connected by wire or wirelessly so that the captured image (video) can be output to the image processing device 100, or is connected via an arbitrary communication network or the like.
  • the camera 200 may be provided inside the image processing device 100.
  • the database 110 is a database that stores information (data), processing results, and the like necessary for processing of the image processing apparatus 100.
  • the database 110 includes an image acquired by the image acquisition unit 101, a detection result of the skeletal structure detection unit 102, data for machine learning, a feature amount calculated by the feature amount calculation unit 103, a classification result of the classification unit 104, and a search unit 105. The search results etc. of are memorized.
  • the database 110 is directly connected to the image processing device 100 by wire or wirelessly so that data can be input and output as needed, or is connected via an arbitrary communication network or the like.
  • the database 110 may be provided inside the image processing device 100 as a non-volatile memory such as a flash memory, a hard disk device, or the like.
  • the image processing apparatus 100 includes an image acquisition unit 101, a skeleton structure detection unit 102, a feature amount calculation unit 103, a classification unit 104, a search unit 105, an input unit 106, a display unit 107, and a first query.
  • the acquisition unit 109 and the second query acquisition unit 111 are provided.
  • the configuration of each part (block) is an example, and may be composed of other parts as long as the method (operation) described later is possible.
  • the image processing device 100 is realized by, for example, a computer device such as a personal computer or a server that executes a program, but it may be realized by one device or by a plurality of devices on a network. good.
  • the input unit 106, the display unit 107, and the like may be used as an external device.
  • both the classification unit 104 and the search unit 105 may be provided, or only one of them may be provided.
  • Both or one of the classification unit 104 and the search unit 105 is the recognition unit 13 that performs the recognition processing of the state of the person.
  • the search unit 105, the first query acquisition unit 109, and the second query acquisition unit 111 are functional units that execute the search process, and correspond to the recognition unit 13 in FIG.
  • the image processing device 100 executes data storage processing, classification processing, and search processing in this order. As will be described below, the image processing apparatus 100 does not have to execute the classification process.
  • the data storage process acquires an image to be analyzed (hereinafter referred to as "image to be analyzed”), detects a two-dimensional skeletal structure of a person from each of a plurality of images to be analyzed, and features the detected two-dimensional skeletal structure. This is a process of calculating the amount, associating the calculated feature amount with each analysis target image, and storing the calculated feature amount in the database 110.
  • image to be analyzed an image to be analyzed
  • This is a process of calculating the amount, associating the calculated feature amount with each analysis target image, and storing the calculated feature amount in the database 110.
  • the image acquisition unit 101 acquires the image to be analyzed.
  • acquisition means “acquisition of data stored in another device or storage medium by the own device” based on user input or program instruction (active acquisition). ) ”, For example, requesting or inquiring about another device to receive it, accessing and reading another device or storage medium, etc., and based on user input or program instructions,“ own device To input data output from other devices (passive acquisition) ", for example, to receive data to be delivered (or transmitted, push notification, etc.), and in the received data or information. Generate new data by selecting from and acquiring, and editing data (text conversion, data sorting, partial data extraction, file format change, etc.), and the new data is generated. Includes at least one of "getting”.
  • the image acquisition unit 101 acquires a two-dimensional image including a person captured by the camera 200 during a predetermined monitoring period as an analysis target image.
  • the image acquisition unit 101 may acquire a moving image. In this case, each of the plurality of frame images constituting the moving image becomes the analysis target image.
  • the image acquisition unit 101 may acquire a still image generated irregularly or at a time interval larger than the frame rate of the moving image as an analysis target image.
  • the image acquisition unit 101 may acquire a two-dimensional image including a person stored in a storage means such as a database 110 as an analysis target image.
  • the skeleton structure detection unit 102 detects the two-dimensional skeleton structure of a person from each of the acquired images to be analyzed.
  • the skeleton structure detection unit 102 can detect the skeleton structure of all the persons recognized in the analysis target image.
  • the skeleton structure detection unit 102 detects the skeleton structure of a person based on the characteristics such as the joints of the person to be recognized by using the skeleton estimation technique using machine learning.
  • the skeleton structure detection unit 102 uses, for example, a skeleton estimation technique such as OpenPose of Non-Patent Document 1 to extract key points that are characteristic points of joints and the like.
  • the feature amount calculation unit 103 calculates the feature amount of the detected two-dimensional skeletal structure, associates the calculated feature amount with the analysis target image in which the two-dimensional skeletal structure is detected, and stores it in the database 110. ..
  • the feature amount of the skeletal structure indicates the characteristics of the skeleton of the person, and is an element for classifying or searching the state of the person based on the skeleton of the person. Usually, this feature quantity includes a plurality of parameters (for example, classification elements described later).
  • the feature amount may be an entire feature amount of the skeletal structure, a partial feature amount of the skeletal structure, or a plurality of feature amounts such as each part of the skeletal structure.
  • the feature amount may be calculated by any method such as machine learning or normalization, and the minimum value or the maximum value may be obtained as the normalization.
  • the feature amount is a feature amount obtained by machine learning the skeletal structure, a size on an image of the skeletal structure from the head to the foot, and the like.
  • the size of the skeleton structure is the vertical height and area of the skeleton region including the skeleton structure on the image.
  • the vertical direction (height direction or vertical direction) is a vertical direction (Y-axis direction) in the image, and is, for example, a direction perpendicular to the ground (reference plane).
  • the left-right direction (horizontal direction) is a left-right direction (X-axis direction) in the image, and is, for example, a direction parallel to the ground.
  • a feature amount having robustness for the classification and search processing it is preferable to use a feature amount having robustness for the classification and search processing.
  • a robust feature amount may be used for the orientation or body shape of the person. It depends on the orientation and body shape of the person by learning the skeleton of a person who is facing in various directions with the same posture and the skeleton of a person with various body shapes in the same posture, and by extracting the characteristics of the skeleton only in the vertical direction. It is possible to obtain a feature amount that does not.
  • the classification process is based on the data stored in the database 110 in the data storage process (data in which the image to be analyzed is associated with the feature amount of the two-dimensional skeletal structure detected from each image to be analyzed).
  • This is a process of collectively classifying (grouping) a plurality of two-dimensional skeletal structures detected from the above by those having similar features.
  • the analysis target image and the two-dimensional skeleton structure detected from each analysis target image are associated with each other. Therefore, the classification of a plurality of two-dimensional skeletal structures by the classification process also becomes the classification of a plurality of images to be analyzed.
  • the classification process a plurality of images to be analyzed are grouped together with images containing similar two-dimensional skeletal structures.
  • the configuration of the functional unit related to the classification process will be described.
  • the classification unit 104 classifies (clusters) a plurality of skeletal structures stored in the database 110 based on the similarity of the feature amounts of the skeletal structures. It can be said that the classification unit 104 classifies the states of a plurality of persons based on the feature amount of the skeletal structure as the process of recognizing the state of the person. Similarity is the distance between features of the skeletal structure.
  • the classification unit 104 may be classified according to the similarity of the features of the whole skeleton structure, or may be classified according to the similarity of the features of a part of the skeleton structure, and the first part of the skeleton structure (for example, for example). It may be classified according to the similarity of the features of both hands) and the second part (for example, both feet).
  • the posture of the person may be classified based on the feature amount of the skeletal structure of the person in each image, or the behavior of the person based on the change in the feature amount of the skeletal structure of the person in a plurality of consecutive images in time series. May be classified. That is, the classification unit 104 can classify the state of the person including the posture and behavior of the person based on the feature amount of the skeletal structure. For example, the classification unit 104 targets a plurality of skeletal structures in a plurality of images captured during a predetermined monitoring period. The classification unit 104 obtains the degree of similarity between the features to be classified, and classifies the skeletal structures having a high degree of similarity into the same cluster (group with similar postures). As with the search, the user may be able to specify the classification conditions. The classification unit 104 can store the classification result of the skeletal structure in the database 110 and display it on the display unit 107.
  • the search process is based on the data stored in the database 110 in the data storage process (data in which the image to be analyzed is associated with the feature amount of the two-dimensional skeletal structure detected from each image to be analyzed). It is a process of searching for a predetermined scene from the inside.
  • a predetermined scene is specified by the posture of each of the plurality of persons, the time distance of the timing at which each of the plurality of persons takes a designated posture, the physical distance of the plurality of persons, and the like.
  • the configuration of the functional unit related to the search process will be described.
  • the first query acquisition unit 109 acquires the first query in which the postures of each of the plurality of persons are specified.
  • the postures of the plurality of persons may be the same posture or different postures. Further, the plurality of persons may be two persons or three or more persons. The posture of each of the plurality of persons is specified by user input.
  • First query acquisition example 1 In this example, the first query acquisition unit 109 displays a plurality of options (postures of a person) prepared in advance toward the user. Then, the first query acquisition unit 109 acquires the posture selected by the user from the options as the first query. The first query acquisition unit 109 can accept the selection corresponding to each of the plurality of persons.
  • multiple options may be presented with keywords such as “sitting with crossed legs”, “seiza”, and “lying on the back”.
  • a plurality of options may be presented with an image showing each of the plurality of postures.
  • the image showing the posture may be an image of a model of a two-dimensional skeleton structure as shown in FIG. 7, or an image including a person and a model of the two-dimensional skeleton structure as shown in FIG.
  • the image may include a person and do not include a model of a two-dimensional skeleton structure.
  • the first query acquisition unit 109 displays an image of a model having a two-dimensional skeleton structure as shown in FIG. 7. Then, the first query acquisition unit 109 accepts an input for changing the posture of the model and a user input for determining the changed posture as the first query.
  • the change of the posture of the model is realized by, for example, an input for moving the position of each of a plurality of key points.
  • the first query acquisition unit 109 acquires the first query in cooperation with the image acquisition unit 101, the skeleton structure detection unit 102, and the feature amount calculation unit 103.
  • the image acquisition unit 101 acquires a query image.
  • the skeleton structure detection unit 102 detects the two-dimensional skeleton structure of the person included in the query image.
  • the feature amount calculation unit 103 calculates the feature amount of the detected two-dimensional skeleton structure.
  • the first query acquisition unit 109 acquires the feature amount of the two-dimensional skeleton structure calculated from the query image as the first query.
  • the image acquisition unit 101 can acquire a query image by, for example, any of the following acquisition examples.
  • the image acquisition unit 101 acquires a query image prepared by the user and input to the image processing device 100.
  • FIG. 42 shows an example of a UI (user interface) screen used in the search.
  • An input field for keywords and a display field for search results are shown.
  • the image selected by the user from the images displayed as the search result is the query image.
  • the keywords are assumed to be related to the state (posture, behavior, etc.) of the person, such as "sitting" and "standing". Keyword input can be realized using a well-known GUI such as a text box, a dropdown, or a check box.
  • an image prepared in advance for use as a query image (hereinafter, “image for query”) and a keyword (a word indicating the state of a person included in each image).
  • image for query a query image
  • keyword a word indicating the state of a person included in each image.
  • the information associated with may be registered in the database 110.
  • the image acquisition unit 101 searches for a query image associated with the input keyword from the information, and acquires a part or all of the query image included in the search result as a query image. You may.
  • information associated with a part of the image to be analyzed and a keyword may be registered in the database 110.
  • the image acquisition unit 101 may search the analysis target image associated with the input keyword from the information and acquire a part or all of the analysis target image included in the search result as a query image. good.
  • the image acquisition unit 101 may send the input keyword to a search engine that searches for images related to the keyword, and acquire the search result from the search engine. Then, the image acquisition unit 101 may acquire a part or all of the images included in the search result as a query image.
  • the acquired query image may include one person or a plurality of persons.
  • the first query acquisition unit 109 selects at least one person from the plurality of persons included in the query image based on the user input. Then, the feature amount of the two-dimensional skeleton structure of the selected person is acquired as the first query.
  • the first query acquisition unit 109 can realize the selection by, for example, one of the following two methods.
  • the first query acquisition unit 109 freely accepts user input for designating a part of the image. Then, the first query acquisition unit 109 selects a person detected in the designated partial area.
  • Person detected in a specified partial area is a person whose body is completely contained in the specified partial area, or at least a part of the body is included in the specified partial area. Is a person who is
  • the first query acquisition unit 109 causes the display unit 107 to display the query image P, and accepts user input that specifies at least a part of the query image P.
  • the first query acquisition unit 109 may accept user input for designating a partial area by user input surrounding a partial area with a frame W. The position and size of the frame W in the image can be freely specified by the user.
  • the first query acquisition unit 109 selects a person detected in the designated part of the area.
  • the process of detecting a person for the entire query image is executed before accepting the user input for selecting the person. Then, as shown in FIG. 45, the first query acquisition unit 109 displays a processed image in which the person detected in the query image is identifiable as a selectable person on the query image. The person surrounded by the frame Y is a selectable person. Then, the first query acquisition unit 109 accepts a user input for selecting at least one of the selectable persons.
  • a process of detecting a person and a process of determining a person to be displayed as a selectable person will be described.
  • the first query acquisition unit 109 executes a process of detecting a person for the entire query image (the entire area in the image) based on the process of detecting a well-known person. Then, the first query acquisition unit 109 sets all the detected persons as selectable persons. In addition, the first query acquisition unit 109 may select a person whose area occupied in the image (meaning the size of the person in the image) is equal to or larger than the reference value among the detected persons as a selectable person. .. For a person whose area occupied in the image is smaller than a predetermined level, it becomes difficult to detect the two-dimensional skeleton structure and calculate the feature amount. If such a posture of a person is used as a query for a search, it becomes difficult to obtain a desired search result. Therefore, such a person may be excluded from the selectable persons.
  • the skeleton structure detection unit 102 executes a process of detecting the two-dimensional skeleton structure of a person for the entire query image (the entire area in the image). Then, the first query acquisition unit 109 sets a person whose detection result of the two-dimensional skeleton structure by the skeleton structure detection unit 102 satisfies a predetermined condition as a selectable person.
  • a person who meets a predetermined condition is calculated based on at least one of "a person whose number of extracted key points is equal to or more than a reference value" or "the number of extracted key points and the reliability of each of the extracted key points".
  • the detailed algorithm for calculating the evaluation value is not particularly limited, but it is designed to satisfy the following contents. -The larger the number of extracted key points, the higher the evaluation value. -The higher the reliability of the extracted key points, the higher the evaluation value. -The larger the person in the image, the higher the evaluation value.
  • each evaluation value may be displayed in association with each of the selectably displayed persons. Further, in the search result of the query image as shown in FIG. 42, each evaluation value may be displayed in association with each of the plurality of query images included in the search result.
  • the statistical values maximum value, minimum value, average value, mode value, median value, etc.
  • the second query acquisition unit 111 acquires the second query in which at least one of the physical distance and the temporal distance between the plurality of persons taking the posture specified in the first query is specified. ..
  • the second query acquisition unit 111 is the physical distance and time between the two persons. Gets a second query that specifies at least one of the target distances. It is not always necessary to specify the physical distance and the time distance corresponding to all the pairs, and the physical distance and the time distance may be specified only for some pairs.
  • the "physical distance” is the physical distance between the first person who takes the posture specified by the first query and the second person who takes the posture specified by the first query.
  • the physical distance may be expressed as an actual distance using a unit such as "m (meter)” or as a distance in each image in terms of the number of pixels.
  • part of the person's body is used as the basis for calculating the physical distance between people.
  • the physical distance between the heads, the physical distance between the feet, and the like may be used.
  • the shortest physical distance between the region in the image in which the body of the first person exists and the region in the image in which the body of the second person exists is defined as the physical distance between the two. May be good.
  • the example here is just an example, and the physical distance between people may be calculated by another method.
  • the physical distance between a plurality of persons taking the posture specified in the first query is specified as, for example, "X1 or more”, “X2 or less”, “X3 or more and X4 or less", and the like.
  • the "time distance” is the time difference between the timing when the first person who takes the posture specified by the first query appears and the timing when the second person who takes the posture specified by the first query appears. Is. In the second query, the temporal distance may be expressed in units such as "seconds" or in the number of frames.
  • the temporal distance between a plurality of persons taking the posture specified in the first query is specified as, for example, "Y1 or more”, “Y2 or less”, “Y3 or more and Y4 or less", and the like.
  • “0 (zero)” may be specified for the time distance.
  • a condition is specified that a plurality of people taking the posture specified in the first query appear in the same image at the same time.
  • the second query acquisition unit 111 displays a plurality of options (physical distance and temporal distance) prepared in advance toward the user. Then, the second query acquisition unit 111 acquires the physical distance or the temporal distance selected by the user from the options as the second query.
  • the second query acquisition unit 111 accepts a user input that specifies a physical distance or a temporal distance numerically. Then, the second query acquisition unit 111 acquires the numerical value input by the user as the second query.
  • the second query acquisition unit 111 calculates at least one of the physical distance and the temporal distance between the two parties included in the query image acquired by the image acquisition unit 101, and the calculated value and its value.
  • the numerical range including the periphery is acquired as the second query. An example will be described below.
  • the second query acquisition unit 111 calculates the physical distance P between the selected two parties, and sets a numerical range (P- ⁇ or more, P + ⁇ or less) including the calculated value P and its surroundings. Determined as the physical distance between the two in the second query.
  • the image acquisition unit 101 may acquire a moving image as a query image. Then, a person is selected from each of the two different frame images in the moving image by user input, and the posture (feature amount of the two-dimensional skeleton structure) of each of the selected two people is acquired as the first query.
  • the posture feature amount of the two-dimensional skeleton structure
  • the second query acquisition unit 111 calculates the temporal distance Q between the two frame images selected by the two persons, and the numerical range (Q- ⁇ ) including the calculated value Q and its surroundings. As described above, Q + ⁇ or less) is determined as the time distance between the two parties in the second query.
  • the second query acquisition unit 111 calculates the physical distance between the two frame images selected by the two persons and the frame images sandwiched between the two persons. Then, the second query acquisition unit 111 calculates the statistical value R (maximum value, minimum value, mean value, mode value, median value, etc.) of the calculated value, and a numerical range including the calculated value R and its surroundings. (R- ⁇ or more, R + ⁇ or less) is the physical between the two in the second query.
  • the search unit 105 executes an image search based on the first query and the second query. For example, the search unit 105 can search for a scene that satisfies all of the following conditions from a moving image composed of a plurality of analysis target images (frame images).
  • first person who takes the first posture specified by the first query and a second person who takes the second posture specified by the first query appear.
  • the time difference between the first timing in which the first person takes the first posture and the second timing in which the second person takes the second posture is a condition of the time distance specified by the second query.
  • the physical distance between the first person and the second person in at least part of the time zone between the first timing and the second timing is the physical distance specified in the second query. Meet the distance conditions.
  • FIG. 47 shows an example of the searched scene.
  • the posture of the first person A included in the analysis target image F1 is designated as the first posture
  • the posture of the second person B included in the analysis target image F2 is the first. It was designated as the posture of 2.
  • the time difference T between these two frame images satisfies the condition of the time distance specified by the second query.
  • the physical distance L between the first person A and the second person B is the second query. Satisfy the conditions of the physical distance specified in.
  • a well-known technique eg, object tracking technique, person recognition using appearance features, etc.
  • the search unit 105 can search for still images that satisfy all of the following conditions from among a plurality of analysis target images (still images).
  • FIG. 46 shows an example of the searched scene.
  • the posture of the first person A included in the analysis target image F3 is designated as the first posture
  • the posture of the second person B included in the analysis target image F3 is the first. It was designated as the posture of 2.
  • the physical distance L between the first person A and the second person B calculated from the analysis target image F3 satisfies the condition of the physical distance specified by the second query.
  • the search unit 105 has, for example, a skeleton having a high degree of similarity to the feature amount of the two-dimensional skeleton structure specified in the first query (query state) from among a plurality of skeleton structures stored in the database 110 in the data storage process. By searching the structure, it is possible to search for a person who takes the posture specified by the first query.
  • the search unit 105 searches for a skeleton structure having a high degree of similarity to the feature amount of the search query by collating the feature amount of the search query with the feature amount of the skeleton structure detected from each of the plurality of analysis target images. You may. In the case of this configuration, the above-mentioned classification process becomes unnecessary. However, since the collation target is all of the plurality of analysis target images, the processing load of the computer in collation becomes large.
  • the search unit 105 determines a representative of the feature amount of the two-dimensional skeletal structure for each group obtained by the classification process by an arbitrary means, and collates the representative with the feature amount of the above search query to obtain a search query. You may search for a skeletal structure having a high degree of similarity to the feature amount. In the case of this configuration, since the number of collation targets is reduced, the processing load of the computer in collation is reduced.
  • the image to be analyzed and the two-dimensional skeleton structure detected from each image to be analyzed are associated with each other. Therefore, by the above-mentioned "process of searching for a predetermined skeleton structure from a plurality of two-dimensional skeleton structures detected from the image to be analyzed", the predetermined skeleton structure (the skeleton structure having a high degree of similarity to the feature amount of the search query). ) Can be searched for the image to be analyzed. That is, it is possible to search the analysis target image including the person who takes the posture specified in the first query from the analysis target images.
  • the degree of similarity is the distance between the features of the skeletal structure.
  • the search unit 105 may search by the similarity of the features of the whole skeleton structure, or may search by the similarity of the features of a part of the skeleton structure, and may search by the similarity of the first part of the skeleton structure (for example,). You may search by the similarity of the features of both hands) and the second part (for example, both feet).
  • the posture of the person may be searched based on the feature amount of the skeletal structure of the person in each image, or the behavior of the person may be searched based on the change of the feature amount of the skeletal structure of the person in a plurality of consecutive images in time series. You may search for.
  • the search unit 105 can search the state of the person including the posture and behavior of the person based on the feature amount of the skeletal structure. For example, the search unit 105 searches for feature quantities of a plurality of skeletal structures in a plurality of analysis target images captured during a predetermined monitoring period.
  • the estimation method is not limited thereto.
  • the search unit 105 calculates the height t in the image of the reference person or a person located around the person. Here, for example, it is represented by the number of pixels.
  • the search unit 105 calculates the physical distance d in the image from the reference person to the person located in the surroundings. Here, d is expressed in the same unit as t (for example, the number of pixels).
  • the search unit 105 calculates d / t and multiplies this value by a preset reference height to calculate the physical distance between the reference person and the person located around the reference person.
  • the search unit 105 may change the reference height according to this attribute.
  • the search unit 105 When calculating the physical distance, the search unit 105 preferably performs a process of correcting this distortion.
  • the search unit 105 can perform distortion correction processing according to the position of a person in the image.
  • image distortion is due, for example, to the optical system of the camera 200 (eg, a lens) and the vertical orientation of the camera 200 (eg, an angle with respect to a horizontal plane). Therefore, the content of the strain correction process according to the position of the person in the image is set according to the optical system (for example, the lens) of the camera 200 and the vertical orientation of the camera 200.
  • the input unit 106 is an input interface for acquiring information input from a user who operates the image processing device 100.
  • the user is a monitor who monitors a person in a suspicious state from an image of a surveillance camera.
  • the input unit 106 is, for example, a GUI (Graphical User Interface), and information according to a user's operation is input from an input device such as a keyboard, a mouse, a touch panel, a microphone, and a physical button.
  • GUI Graphic User Interface
  • the display unit 107 is a display unit that displays the result of the operation (processing) of the image processing device 100, and is, for example, a display device such as a liquid crystal display or an organic EL (ElectroLuminescence) display.
  • the display unit 107 displays the classification result of the classification unit 104, the search result of the search unit 105, and the like.
  • Each functional unit of the image processing device 100 is stored in a storage unit (from the stage of shipping the device in advance) such as a CPU (Central Processing Unit) of an arbitrary computer, a memory, a program loaded in the memory, and a hard disk for storing the program. It can store programs downloaded from storage media such as CDs (Compact Discs) and servers on the Internet), and can be realized by any combination of hardware and software centered on the network connection interface. Will be done. And, it is understood by those skilled in the art that there are various variations in the method of realizing the device and the device.
  • FIG. 41 is a block diagram illustrating a hardware configuration of the image processing device 100.
  • the image processing device 100 includes a processor 1A, a memory 2A, an input / output interface 3A, a peripheral circuit 4A, and a bus 5A.
  • the peripheral circuit 4A includes various modules.
  • the image processing device 100 does not have to have the peripheral circuit 4A.
  • the image processing device 100 may be composed of a plurality of physically and / or logically separated devices, or may be composed of one physically and / or logically integrated device. When the image processing device 100 is composed of a plurality of physically and / or logically separated devices, each of the plurality of devices can be provided with the above hardware configuration.
  • the bus 5A is a data transmission path for the processor 1A, the memory 2A, the peripheral circuit 4A, and the input / output interface 3A to transmit and receive data to each other.
  • the processor 1A is, for example, an arithmetic processing unit such as a CPU or a GPU (Graphics Processing Unit).
  • the memory 2A is, for example, a memory such as a RAM (RandomAccessMemory) or a ROM (ReadOnlyMemory).
  • the input / output interface 3A includes an interface for acquiring information from an input device, an external device, an external server, an external sensor, a camera, etc., an interface for outputting information to an output device, an external device, an external server, etc. ..
  • the input device is, for example, a keyboard, a mouse, a microphone, a physical button, a touch panel, or the like.
  • the output device is, for example, a display, a speaker, a printer, a mailer, or the like.
  • the processor 1A can issue a command to each module and perform a calculation based on the calculation result thereof.
  • FIG. 3 to 5 show the processing flow of the image processing apparatus 100 according to the present embodiment.
  • FIG. 3 shows a flow from image acquisition to search processing in the image processing apparatus 100
  • FIG. 4 shows a flow of classification processing (S104) in FIG. 3
  • FIG. 5 shows a flow of search processing (S105) in FIG. It shows the flow.
  • the image acquisition unit 101 acquires a plurality of analysis target images (S101). Subsequently, the skeleton structure detection unit 102 detects the two-dimensional skeleton structure of the person from each of the acquired plurality of analysis target images (S102).
  • FIG. 6 shows an example of detecting a skeletal structure. As shown in FIG. 6, the analysis target image may include a plurality of people. In this case, the skeleton structure detection unit 102 detects the skeleton structure for each person included in the analysis target image.
  • FIG. 7 shows the skeleton structure of the human body model 300 detected at this time
  • FIGS. 8 to 10 show an example of detecting the skeleton structure.
  • the skeleton structure detection unit 102 detects the skeleton structure of the human body model (two-dimensional skeleton model) 300 as shown in FIG. 7 from the two-dimensional image by using a skeleton estimation technique such as OpenPose.
  • the human body model 300 is a two-dimensional model composed of key points such as joints of a person and bones connecting the key points.
  • the skeleton structure detection unit 102 extracts feature points that can be key points from an image, and detects each key point of a person by referring to information obtained by machine learning the image of the key points.
  • the key points of the person are head A1, neck A2, right shoulder A31, left shoulder A32, right elbow A41, left elbow A42, right hand A51, left hand A52, right hip A61, left hip A62, and right knee A71.
  • Left knee A72, right foot A81, left foot A82 are detected.
  • Bone B31 and B32 connecting the elbow A41 and the left elbow A42, respectively, connecting the right elbow A41 and the left elbow A42 to the right hand A51 and the left hand A52, respectively, and connecting the neck A2 to the right waist A61 and the left waist A62, respectively.
  • the skeleton structure detection unit 102 stores the detected skeleton structure of the person in the database 110.
  • FIG. 8 is an example of detecting a person in an upright position.
  • a standing person is imaged from the front, and bones B1, bone B51 and bone B52, bones B61 and bone B62, bones B71 and bone B72 viewed from the front are detected without overlapping, and the right foot is detected.
  • Bone B61 and Bone B71 are slightly bent more than Bone B62 and Bone B72 of the left foot.
  • FIG. 9 is an example of detecting a person in a crouched state.
  • a crouching person is imaged from the right side, bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72, respectively, viewed from the right side, and bone B61 on the right foot.
  • the bone B71 and the bone B62 and the bone B72 of the left foot are greatly bent and overlapped.
  • FIG. 10 is an example of detecting a sleeping person.
  • a sleeping person is imaged from diagonally left front, and bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 viewed from diagonally left front are detected, respectively, and the right foot.
  • the bones B61 and B71 of the left foot and the bones B62 and B72 of the left foot are bent and overlapped.
  • the feature amount calculation unit 103 calculates the feature amount of the detected skeletal structure (S103). For example, when the height or area of the skeleton region is used as the feature amount, the feature amount calculation unit 103 extracts a region including the skeleton structure and obtains the height (number of pixels) or area (pixel area) of the region. The height and area of the skeletal region can be obtained from the coordinates of the end of the extracted skeleton region and the coordinates of the key points at the ends. The feature amount calculation unit 103 stores the obtained feature amount of the skeletal structure in the database 110. The feature amount of this skeletal structure is also used as information indicating the state of a person.
  • the skeletal region including all the bones is extracted from the skeletal structure of the standing person.
  • the upper end of the skeletal region is the key point A1 of the head
  • the lower end of the skeletal region is the key point A82 of the left foot
  • the left end of the skeletal region is the key point A41 of the right elbow
  • the right end of the skeletal region is the key point A52 of the left hand. .. Therefore, the height of the skeleton region is obtained from the difference between the Y coordinates of the key point A1 and the key point A82.
  • the width of the skeleton region is obtained from the difference between the X coordinates of the key point A41 and the key point A52, and the area is obtained from the height and width of the skeleton region.
  • the skeletal region including all bones is extracted from the skeletal structure of a crouched person.
  • the upper end of the skeletal region is the key point A1 of the head
  • the lower end of the skeletal region is the key point A81 of the right foot
  • the left end of the skeletal region is the key point A61 of the right hip
  • the right end of the skeletal region is the key point A51 of the right hand. .. Therefore, the height of the skeleton region is obtained from the difference between the Y coordinates of the key point A1 and the key point A81.
  • the width of the skeleton region is obtained from the difference between the X coordinates of the key point A61 and the key point A51, and the area is obtained from the height and width of the skeleton region.
  • a skeletal region including all bones is extracted from the skeletal structure of a sleeping person.
  • the upper end of the skeleton region is the key point A32 of the left shoulder
  • the lower end of the skeleton region is the key point A52 of the left hand
  • the left end of the skeleton region is the key point A51 of the right hand
  • the right end of the skeleton region is the key point A82 of the left foot. Therefore, the height of the skeleton region is obtained from the difference between the Y coordinates of the key point A32 and the key point A52.
  • the width of the skeleton region is obtained from the difference between the X coordinates of the key point A51 and the key point A82, and the area is obtained from the height and width of the skeleton region.
  • the classification unit 104 performs a classification process (S104).
  • the classification unit 104 calculates the similarity of the calculated feature amount of the skeletal structure (S111), and classifies the skeletal structure based on the calculated feature amount (S112). ..
  • the classification unit 104 obtains the similarity of the feature quantities between all the skeletal structures stored in the database 110 to be classified, and classifies (clusters) the skeletal structures (postures) having the highest similarity into the same cluster. .. Further, the similarity between the classified clusters is obtained and classified, and the classification is repeated until a predetermined number of clusters are obtained.
  • FIG. 11 shows an image of the classification result of the feature amount of the skeletal structure.
  • FIG. 11 is an image of cluster analysis using a two-dimensional classification element, and the two classification elements are, for example, the height of the skeletal region and the area of the skeletal region.
  • the feature quantities of the plurality of skeletal structures are classified into three clusters C1 to C3.
  • the clusters C1 to C3 correspond to each posture such as a standing posture, a sitting posture, and a sleeping posture, and the skeletal structure (person) is classified for each similar posture.
  • various classification methods can be used by classifying based on the feature amount of the skeletal structure of a person.
  • the classification method may be set in advance or may be arbitrarily set by the user. Further, the classification may be performed by the same method as the search method described later. That is, it may be classified according to the same classification conditions as the search conditions.
  • the classification unit 104 classifies by the following classification method. Any classification method may be used, or an arbitrarily selected classification method may be combined.
  • the feature quantities may be stacked in the time series direction and classified based on the cumulative value. Further, it may be classified based on the change (change amount) of the feature amount of the skeletal structure in a plurality of continuous images.
  • ⁇ Classification method 3> Classification ignoring the left and right of the skeletal structure Classify the skeletal structures whose right and left sides are opposite to each other as the same skeletal structure.
  • the classification unit 104 displays the classification result of the skeletal structure (S113).
  • the classification unit 104 acquires images of necessary skeleton structures and people from the database 110, and displays the skeleton structure and people on the display unit 107 for each posture (cluster) similar as a classification result.
  • FIG. 12 shows a display example when the postures are classified into three. For example, as shown in FIG. 12, the posture regions WA1 to WA3 for each posture are displayed in the display window W1, and the skeletal structure and the person (image) of the posture corresponding to each of the posture regions WA1 to WA3 are displayed.
  • the posture area WA1 is, for example, a display area of a standing posture, and displays a skeletal structure and a person similar to a standing posture classified into cluster C1.
  • the posture area WA2 is, for example, a sitting posture display area, and displays a skeletal structure and a person similar to the sitting posture classified into cluster C2.
  • the posture area WA3 is, for example, a display area of a sleeping posture, and displays a skeletal structure and a person similar to the sleeping posture classified into cluster C2.
  • the image processing apparatus 100 performs a search process (S105).
  • the image processing device 100 accepts the input of the search condition (S121).
  • the image processing device 100 displays a UI screen as shown in FIG. 17, and accepts input of a first query and a second query from the UI screen.
  • the illustrated UI screen accepts the input of the first query in the posture designation item.
  • the posture of each of a plurality of persons can be specified.
  • "select from choices”, "input query image”, and "use skeleton model” can be selected. These methods correspond to the first query acquisition examples 1 to 3 described above.
  • the first query acquisition unit 109 acquires the first query by the selected method.
  • the illustrated UI screen accepts the input of the second query in the items of physical distance and temporal distance.
  • “select from choices”, “numerical input”, and “input query image” can be selected. These methods correspond to the above-mentioned second query acquisition examples 1 to 3.
  • the second query acquisition unit 111 acquires the second query by the selected method.
  • the search unit 105 executes an image search based on the first query and the second query (S122).
  • the search unit 105 searches for a skeleton structure having a high degree of similarity in feature quantities from the skeleton structures stored in the database 110 to be searched, using the skeleton structure specified in the first query as a search query.
  • the search result is used to search for a scene that satisfies the condition of the second query.
  • the search unit 105 has a feature amount of the skeleton structure of the search query and a feature amount of the skeleton structure of the search target (detected from the image to be analyzed). The degree of similarity with the feature amount of the skeleton structure) is calculated, and the skeleton structure whose calculated degree of similarity is higher than a predetermined threshold is extracted.
  • the feature amount of the skeleton structure of the search query the feature amount calculated in advance may be used, or the feature amount obtained at the time of search may be used.
  • search method can be used by searching based on the feature amount of the skeletal structure of the person.
  • the search method may be set in advance or may be arbitrarily set by the user.
  • the search unit 105 searches by the following search method. Either search method may be used, or an arbitrarily selected search method may be combined.
  • a plurality of search methods search conditions may be combined and searched by a logical expression (for example, AND (logical product), OR (logical sum), NOT (negation)).
  • search condition may be searched as "(posture in which the right hand is raised) AND (posture in which the left foot is raised)".
  • ⁇ Search method 1> Search by only the feature amount in the height direction By searching using only the feature amount in the height direction of the person, the influence of the lateral change of the person can be suppressed, and the change of the direction of the person and the body shape of the person can be suppressed. On the other hand, the robustness is improved. For example, as in the skeletal structures 501 to 503 of FIG. 13, even if the orientation and body shape of the person are different, the feature amount in the height direction does not change significantly. Therefore, in the skeletal structures 501 to 503, it can be determined that the postures are the same at the time of searching (at the time of classification).
  • ⁇ Search method 2> When a part of the body of a person is hidden in the partial search image, the search is performed using only the information of the recognizable part. For example, as in the skeletal structures 511 and 512 of FIG. 14, even if the key point of the left foot cannot be detected due to the hiding of the left foot, the feature amount of other detected key points can be used for the search. Therefore, in the skeletal structures 511 and 512, it can be determined that the postures are the same at the time of searching (at the time of classification). That is, it is possible to perform classification and search using the features of some key points instead of all the key points. In the examples of the skeletal structures 521 and 522 of FIG.
  • the feature quantities of the key points of the upper body (A1, A2, A31, A32, A41, A42, A51, A52) are used as the search query. Therefore, it can be determined that the posture is the same. Further, the portion (feature point) to be searched may be weighted and searched, or the threshold value for determining the similarity may be changed. When a part of the body is hidden, the hidden part may be ignored and the search may be performed, or the hidden part may be added to the search. By searching including hidden parts, it is possible to search for postures in which the same part is hidden.
  • ⁇ Search method 3> Search ignoring the left and right of the skeletal structure Search for the skeletal structure with the opposite right and left sides of the person as the same skeletal structure.
  • the posture in which the right hand is raised and the posture in which the left hand is raised can be searched (classified) as the same posture.
  • the skeletal structure 531 and the skeletal structure 532 have different positions of the right hand key point A51, the right elbow key point A41, the left hand key point A52, and the left elbow key point A42, but other key points. The position of is the same.
  • ⁇ Search method 4> Search by features in the vertical and horizontal directions After searching only with the features in the vertical direction (Y-axis direction) of the person, the obtained results are further obtained using the features in the horizontal direction (X-axis direction) of the person. Search.
  • Search by multiple images in chronological order Search is performed based on the feature quantity of the skeletal structure in a plurality of images consecutive in chronological order.
  • the feature quantities may be stacked in the time series direction and searched based on the cumulative value.
  • the search may be performed based on the change (change amount) of the feature amount of the skeletal structure in a plurality of consecutive images.
  • the search unit 105 displays the search result of the skeletal structure (S123).
  • the search unit 105 displays still images included in the search results, thumbnail images of scenes, and the like on the display unit 107.
  • the present embodiment it is possible to detect the skeletal structure of a person from a two-dimensional image and perform classification and search based on the feature amount of the detected skeletal structure. As a result, it is possible to classify by similar postures having a high degree of similarity, and it is possible to search for similar postures having a high degree of similarity with a search query (search key).
  • search key search key
  • the posture of the person in the image can be grasped without the user specifying the posture or the like. Since the user can specify the posture of the search query from the classification results, the desired posture can be searched even if the user does not know the posture to be searched in detail in advance. For example, since it is possible to perform classification and search on the condition of the whole or part of the skeleton structure of a person, flexible classification and search is possible.
  • the postures of each of the plurality of persons are specified, the physical distance between the persons who take those postures, the time distance of the timing when each of the plurality of persons takes the designated postures, and the like. You can specify and perform an image search. Therefore, it is possible to efficiently search for a desired scene.
  • a predetermined posture appears with a time lag, such as one person throwing an object and another person receiving the object.
  • a time lag such as one person throwing an object and another person receiving the object.
  • "throwing posture”, “receiving thing”, “time difference between these two postures”, “physical distance between two people taking those two postures”, etc. should be set appropriately. Therefore, the desired scene can be efficiently searched.
  • FIG. 18 shows the configuration of the image processing apparatus 100 according to the present embodiment.
  • the image processing apparatus 100 further includes a height calculation unit 108 in addition to the configuration of the first embodiment.
  • the feature amount calculation unit 103 and the height calculation unit 108 may be combined into one processing unit.
  • the height calculation unit (height estimation unit) 108 calculates the height (called the number of height pixels) of a person in a two-dimensional image when standing up based on the two-dimensional skeleton structure detected by the skeleton structure detection unit 102 (referred to as the number of height pixels). presume. It can also be said that the number of height pixels is the height of the person in the two-dimensional image (the length of the whole body of the person in the two-dimensional image space). The height calculation unit 108 obtains the number of height pixels (number of pixels) from the length (length on the two-dimensional image space) of each bone of the detected skeleton structure.
  • specific examples 1 to 3 are used as a method for obtaining the number of height pixels.
  • any of the methods of Specific Examples 1 to 3 may be used, or a plurality of arbitrarily selected methods may be used in combination.
  • the number of height pixels is obtained by totaling the lengths of the bones from the head to the foot among the bones of the skeletal structure. If the skeleton structure detection unit 102 (skeleton estimation technique) does not output the crown and feet, it can be corrected by multiplying by a constant if necessary.
  • the number of height pixels is calculated using a human body model showing the relationship between the length of each bone and the length of the whole body (height in the two-dimensional image space).
  • the number of height pixels is calculated by fitting (fitting) a three-dimensional human body model to a two-dimensional skeleton structure.
  • the feature amount calculation unit 103 of the present embodiment is a normalization unit that normalizes the skeletal structure (skeleton information) of a person based on the calculated number of height pixels of the person.
  • the feature amount calculation unit 103 stores the feature amount (normalized value) of the normalized skeletal structure in the database 110.
  • the feature amount calculation unit 103 normalizes the height of each key point (feature point) included in the skeleton structure on the image by the number of height pixels.
  • the height direction is the vertical direction (Y-axis direction) in the two-dimensional coordinate (XY coordinate) space of the image. In this case, the height of the key point can be obtained from the value (number of pixels) of the Y coordinate of the key point.
  • the height direction may be the direction of the vertical axis perpendicular to the ground (reference plane) in the three-dimensional coordinate space in the real world, and the direction of the vertical projection axis projected onto the two-dimensional coordinate space (vertical projection direction).
  • the height of the key point is the vertical projection axis obtained by projecting the axis perpendicular to the ground in the real world onto the two-dimensional coordinate space based on the camera parameters, and the value along this vertical projection axis (number of pixels). ) Can be obtained.
  • the camera parameters are image imaging parameters, and for example, the camera parameters are the posture, position, imaging angle, focal length, and the like of the camera 200.
  • the camera 200 can take an image of an object whose length and position are known in advance, and obtain camera parameters from the image. Distortion occurs at both ends of the captured image, and the vertical direction of the real world may not match the vertical direction of the image. On the other hand, by using the parameters of the camera that took the image, you can see how much the vertical direction in the real world is tilted in the image. Therefore, by normalizing the value of the key points along the vertical projection axis projected in the image based on the camera parameters by height, it is possible to feature the key points in consideration of the deviation between the real world and the image. can.
  • the left-right direction is the left-right direction (X-axis direction) in the two-dimensional coordinate (XY coordinates) space of the image, or the direction parallel to the ground in the three-dimensional coordinate space in the real world. Is the direction projected onto the two-dimensional coordinate space.
  • 19 to 23 show the processing flow of the image processing apparatus 100 according to the present embodiment.
  • 19 shows a flow from image acquisition to search processing in the image processing apparatus 100
  • FIGS. 20 to 22 show the flow of specific examples 1 to 3 of the height pixel number calculation process (S201) of FIG. 23 shows the flow of the normalization process (S202) of FIG.
  • the height pixel number calculation process (S201) and the normalization process (S202) are performed as the feature amount calculation process (S103) in the first embodiment. Others are the same as those in the first embodiment.
  • the image processing apparatus 100 performs height pixel number calculation processing based on the detected skeleton structure (S201) following image acquisition (S101) and skeleton structure detection (S102).
  • the height of the skeleton structure of the person standing upright in the image is the height pixel number (h)
  • the height of each key point of the skeleton structure in the state of the person in the image is the key point. Let it be the height (yi).
  • specific examples 1 to 3 of the height pixel number calculation process will be described.
  • the number of height pixels is obtained using the length of the bone from the head to the foot.
  • the height calculation unit 108 acquires the length of each bone (S211) and totals the lengths of the acquired bones (S212).
  • the height calculation unit 108 acquires the length of the bone on the two-dimensional image of the foot from the head of the person, and obtains the number of height pixels. That is, from the image in which the skeletal structure is detected, among the bones of FIG. 24, bone B1 (length L1), bone B51 (length L21), bone B61 (length L31) and bone B71 (length L41), or , Bone B1 (length L1), bone B52 (length L22), bone B62 (length L32), and bone B72 (length L42) are acquired. The length of each bone can be obtained from the coordinates of each key point in the two-dimensional image.
  • the longer value is taken as the number of height pixels. That is, each bone has the longest length in the image when it is imaged from the front, and it is displayed short when it is tilted in the depth direction with respect to the camera. Therefore, it is more likely that the longer bone is imaged from the front, which is considered to be closer to the true value. Therefore, it is preferable to select the longer value.
  • bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 are detected without overlapping.
  • the total of these bones, L1 + L21 + L31 + L41 and L1 + L22 + L32 + L42, is obtained, and for example, the value obtained by multiplying L1 + L22 + L32 + L42 on the left foot side where the detected bone length is long by a correction constant is taken as the height pixel number.
  • bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 are detected, respectively, and the right foot bone B61 and bone B71 and the left foot bone B62 and bone B72 overlap each other. ..
  • the total of these bones, L1 + L21 + L31 + L41 and L1 + L22 + L32 + L42, is obtained, and for example, the value obtained by multiplying L1 + L21 + L31 + L41 on the right foot side where the detected bone length is long by a correction constant is taken as the height pixel number.
  • bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 are detected, respectively, and the right foot bone B61 and bone B71 and the left foot bone B62 and bone B72 overlap each other. ..
  • the total of these bones, L1 + L21 + L31 + L41 and L1 + L22 + L32 + L42, is obtained, and for example, the value obtained by multiplying L1 + L22 + L32 + L42 on the left foot side where the detected bone length is long by a correction constant is taken as the height pixel number.
  • the height can be calculated by summing the lengths of the bones from the head to the feet, so the number of height pixels can be calculated by a simple method.
  • the number of height pixels can be accurately calculated even when the entire person is not always shown in the image, such as when crouching down. Can be estimated.
  • the number of height pixels is obtained using a two-dimensional skeleton model showing the relationship between the length of the bone included in the two-dimensional skeleton structure and the length of the whole body of the person in the two-dimensional image space.
  • FIG. 28 is a human body model (two-dimensional skeleton model) 301 showing the relationship between the length of each bone in the two-dimensional image space and the length of the whole body in the two-dimensional image space used in the second embodiment.
  • the relationship between the length of each bone of an average person and the length of the whole body is associated with each bone of the human body model 301.
  • the length of the bone B1 of the head is the length of the whole body ⁇ 0.2 (20%)
  • the length of the bone B41 of the right hand is the length of the whole body ⁇ 0.15 (15%)
  • the length of the right foot is the length of the whole body ⁇ 0.25 (25%).
  • the average whole body length can be obtained from the length of each bone.
  • a human body model may be prepared for each attribute of the person such as age, gender, and nationality. As a result, the length (height) of the whole body can be appropriately obtained according to the attributes of the person.
  • the height calculation unit 108 acquires the length of each bone (S221).
  • the height calculation unit 108 acquires the lengths (lengths in the two-dimensional image space) of all the bones in the detected skeletal structure.
  • FIG. 29 is an example in which a person in a crouched state is imaged from diagonally right behind and the skeletal structure is detected. In this example, since the face and left side of the person are not shown, the bones of the head, the left arm, and the bones of the left hand cannot be detected. Therefore, the lengths of the detected bones B21, B22, B31, B41, B51, B52, B61, B62, B71, and B72 are acquired.
  • the height calculation unit 108 calculates the number of height pixels from the length of each bone based on the human body model (S222).
  • the height calculation unit 108 refers to the human body model 301 showing the relationship between each bone and the length of the whole body as shown in FIG. 28, and obtains the number of height pixels from the length of each bone. For example, since the length of the bone B41 on the right hand is the length of the whole body ⁇ 0.15, the number of height pixels based on the bone B41 is obtained from the length of the bone B41 / 0.15. Further, since the length of the bone B71 of the right foot is the length of the whole body ⁇ 0.25, the number of height pixels based on the bone B71 is obtained from the length of the bone B71 / 0.25.
  • the human body model referred to at this time is, for example, a human body model of an average person, but a human body model may be selected according to the attributes of the person such as age, gender, and nationality. For example, when a person's face is shown in the captured image, the attribute of the person is identified based on the face, and the human body model corresponding to the identified attribute is referred to. It is possible to recognize a person's attributes from the facial features of the image by referring to the information obtained by machine learning the face for each attribute. Further, when the attribute of the person cannot be identified from the image, the human body model of the average person may be used.
  • the number of height pixels calculated from the length of the bone may be corrected by the camera parameter. For example, when the camera is taken at a high position and looking down at a person, the horizontal length of the shoulder-width bones, etc. is not affected by the depression angle of the camera in the two-dimensional skeletal structure, but the vertical length of the neck-waist bones, etc. The length decreases as the depression angle of the camera increases. Then, the number of height pixels calculated from the horizontal length of the shoulder-width bones and the like tends to be larger than the actual number.
  • the height calculation unit 108 calculates the optimum value of the number of height pixels as shown in FIG. 21 (S223).
  • the height calculation unit 108 calculates the optimum value of the number of height pixels from the number of height pixels obtained for each bone. For example, as shown in FIG. 30, a histogram of the number of height pixels obtained for each bone is generated, and a large number of height pixels is selected from the histogram. That is, the number of height pixels longer than the others is selected from the plurality of height pixels obtained based on the plurality of bones. For example, the upper 30% is set as a valid value, and in FIG. 30, the number of height pixels by bones B71, B61, and B51 is selected.
  • the average number of selected height pixels may be obtained as the optimum value, or the largest number of height pixels may be used as the optimum value. Since the height is calculated from the length of the bone in the 2D image, the length of the bone is taken from the front when the bone is not made from the front, that is, when the bone is tilted in the depth direction when viewed from the camera. It will be shorter than the case. Then, a value having a large number of height pixels is more likely to be captured from the front than a value having a small number of height pixels, and is a more plausible value. Therefore, a larger value is set as the optimum value.
  • the number of height pixels is calculated based on the detected bones of the skeleton structure using a human body model showing the relationship between the bones in the two-dimensional image space and the length of the whole body, so that all the skeletons from the head to the feet are obtained. Even if is not obtained, the number of height pixels can be obtained from some bones. In particular, the number of height pixels can be estimated accurately by adopting a larger value among the values obtained from a plurality of bones.
  • the height calculation unit 108 first calculates the camera parameters based on the image captured by the camera 200 (S231).
  • the height calculation unit 108 extracts an object whose length is known in advance from a plurality of images captured by the camera 200, and obtains a camera parameter from the size (number of pixels) of the extracted object.
  • the camera parameters may be obtained in advance, and the obtained camera parameters may be acquired as needed.
  • the height calculation unit 108 adjusts the arrangement and height of the three-dimensional human body model (S232).
  • the height calculation unit 108 prepares a three-dimensional human body model for calculating the number of height pixels for the detected two-dimensional skeleton structure, and arranges the three-dimensional human body model in the same two-dimensional image based on the camera parameters.
  • the "relative positional relationship between the camera and the person in the real world" is specified from the camera parameters and the two-dimensional skeleton structure. For example, assuming that the position of the camera is the coordinates (0, 0, 0), the coordinates (x, y, z) of the position where the person is standing (or sitting) are specified. Then, by assuming an image in which a three-dimensional human body model is placed at the same position (x, y, z) as the specified person and captured, the two-dimensional skeleton structure and the three-dimensional human body model are superimposed.
  • FIG. 31 is an example in which a crouching person is imaged diagonally from the front left and the two-dimensional skeleton structure 401 is detected.
  • the two-dimensional skeleton structure 401 has two-dimensional coordinate information. It is preferable that all bones are detected, but some bones may not be detected.
  • a three-dimensional human body model 402 as shown in FIG. 32 is prepared.
  • the three-dimensional human body model (three-dimensional skeleton model) 402 has three-dimensional coordinate information and is a model of a skeleton having the same shape as the two-dimensional skeleton structure 401.
  • the prepared three-dimensional human body model 402 is arranged and superimposed on the detected two-dimensional skeleton structure 401.
  • the height of the three-dimensional human body model 402 is adjusted so as to match the two-dimensional skeleton structure 401.
  • the three-dimensional human body model 402 prepared at this time may be a model in a state close to the posture of the two-dimensional skeleton structure 401 as shown in FIG. 33, or may be a model in an upright state.
  • a 3D human body model 402 of the estimated posture may be generated by using a technique of estimating the posture of the 3D space from the 2D image using machine learning. By learning the information of the joints in the 2D image and the joints in the 3D space, the 3D posture can be estimated from the 2D image.
  • the height calculation unit 108 fits the three-dimensional human body model into the two-dimensional skeletal structure as shown in FIG. 22 (S233). As shown in FIG. 34, the height calculation unit 108 superimposes the three-dimensional human body model 402 on the two-dimensional skeletal structure 401 so that the postures of the three-dimensional human body model 402 and the two-dimensional skeletal structure 401 match. Transform the dimensional human body model 402. That is, the height, body orientation, and joint angle of the three-dimensional human body model 402 are adjusted and optimized so that there is no difference from the two-dimensional skeletal structure 401.
  • the joints of the three-dimensional human body model 402 are rotated within the range of movement of the person, the entire three-dimensional human body model 402 is rotated, and the overall size is adjusted.
  • the fitting of the three-dimensional human body model and the two-dimensional skeletal structure is performed in the two-dimensional space (two-dimensional coordinates). That is, the 3D human body model is mapped to the 2D space, and the 3D human body model is converted into a 2D skeletal structure in consideration of how the deformed 3D human body model changes in the 2D space (image). Optimize.
  • the height calculation unit 108 calculates the number of height pixels of the fitted three-dimensional human body model as shown in FIG. 22 (S234). As shown in FIG. 35, the height calculation unit 108 obtains the number of height pixels of the three-dimensional human body model 402 in that state when the difference between the three-dimensional human body model 402 and the two-dimensional skeleton structure 401 disappears and the postures match. With the optimized 3D human body model 402 upright, the length of the whole body in 2D space is obtained based on the camera parameters. For example, the height pixel number is calculated from the bone length (number of pixels) from the head to the foot when the three-dimensional human body model 402 is upright. Similar to the first embodiment, the lengths of the bones from the head to the foot of the three-dimensional human body model 402 may be totaled.
  • the image processing apparatus 100 performs a normalization process (S202) following the height pixel number calculation process.
  • the feature amount calculation unit 103 calculates the key point height (S241).
  • the feature amount calculation unit 103 calculates the key point height (number of pixels) of all the key points included in the detected skeleton structure.
  • the key point height is the length (number of pixels) in the height direction from the lowest end of the skeleton structure (for example, the key point of any foot) to the key point.
  • the height of the key point is obtained from the Y coordinate of the key point in the image.
  • the key point height may be obtained from the length in the direction along the vertical projection axis based on the camera parameters.
  • the height (y) of the key point A2 of the neck is a value obtained by subtracting the Y coordinate of the key point A81 of the right foot or the Y coordinate of the key point A82 of the left foot from the Y coordinate of the key point A2.
  • the reference point is a reference point for expressing the relative height of the key point.
  • the reference point may be preset or may be selectable by the user.
  • the reference point is preferably at the center of the skeletal structure or higher than the center (upper in the vertical direction of the image), and for example, the coordinates of the key point of the neck are used as the reference point.
  • the coordinates of the head and other key points, not limited to the neck, may be used as the reference point.
  • any coordinate for example, the center coordinate of the skeleton structure may be used as a reference point.
  • the feature amount calculation unit 103 normalizes the key point height (yi) by the number of height pixels (S243).
  • the feature amount calculation unit 103 normalizes each key point by using the key point height, the reference point, and the number of height pixels of each key point. Specifically, the feature amount calculation unit 103 normalizes the relative height of the key point with respect to the reference point by the number of height pixels.
  • the feature amount (normalized value) is obtained by using the following equation (1) with the Y coordinate of the reference point (key point of the neck) as (yc).
  • (yi) and (yc) are converted into values in the direction along the vertical projection axis.
  • the coordinates (x0, y0), (x1, y1), ... (X17, y17) of the 18 points of each key point are set to the following using the above equation (1). It is converted into an 18-dimensional feature amount as in.
  • FIG. 36 shows an example of the feature amount of each key point obtained by the feature amount calculation unit 103.
  • the feature amount of the key point A2 is 0.0
  • the feature amount of the key point A31 on the right shoulder and the key point A32 on the left shoulder at the same height as the neck are also. It is 0.0.
  • the feature amount of the key point A1 of the head higher than the neck is -0.2.
  • the feature amount of the right hand key point A51 and the left hand key point A52 lower than the neck is 0.4, and the feature amount of the right foot key point A81 and the left foot key point A82 is 0.9.
  • the feature amount of the key point A52 of the left hand is ⁇ 0.4.
  • the feature amount (normalized value) of the present embodiment shows the characteristics of the skeleton structure (key point) in the height direction (Y direction), and affects the change of the skeleton structure in the lateral direction (X direction). Do not receive.
  • the skeleton structure of a person is detected from the two-dimensional image, and the number of height pixels (height when standing upright on the two-dimensional image space) obtained from the detected skeleton structure is used. Normalize each key point in the skeletal structure. By using this normalized feature amount, it is possible to improve the robustness when classification, search, or the like is performed. That is, since the feature amount of the present embodiment is not affected by the lateral change of the person as described above, it is highly robust to the change of the direction of the person and the body shape of the person.
  • a first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people
  • a second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
  • a search means for executing an image search based on the first query and the second query, and Image processing device with. 2.
  • the second query acquisition means describes in any one of 1 to 3 for acquiring a value selected by the user from a plurality of options prepared in advance or a numerical value input by the user as the second query.
  • the search means is From the moving image A first person in the first posture specified in the first query and a second person in the second posture specified in the first query appear. The time difference between the first timing in which the first person takes the first posture and the second timing in which the second person takes the second posture is the time specified in the second query. Satisfy the conditions of target distance and The physical distance between the first person and the second person is specified in the second query in at least a portion of the time zone between the first timing and the second timing.
  • the image processing apparatus according to any one of 1 to 5, which searches for a scene satisfying the condition of the physical distance. 7.
  • the search means is From multiple still images A first person in the first posture specified in the first query and a second person in the second posture specified in the first query appear.
  • Image processing device 8.
  • the image processing apparatus according to any one of 1 to 7, wherein the postures of the plurality of persons specified in the first query are the same posture or different postures. 9.
  • the computer Get the first query that specifies the posture of each of the multiple people Obtain a second query that specifies at least one of the physical distance and the temporal distance between the plurality of people who take the specified posture.
  • An image processing method for executing an image search based on the first query and the second query. 10.
  • a first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people
  • a second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
  • a search means for performing an image search based on the first query and the second query.
  • a program that functions as.
  • Image processing system 10 Image processing device 11 Skeleton detection unit 12 Feature amount calculation unit 13 Recognition unit 100 Image processing device 101 Image acquisition unit 102 Skeleton structure detection unit 103 Feature amount calculation unit 104 Classification unit 105 Search unit 106 Input unit 107 Display unit 108 Height calculation unit 109 First query acquisition unit 110 Database 111 Second query acquisition unit 200 Camera 300, 301 Human body model 401 Two-dimensional skeleton structure 402 Three-dimensional human body model

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Image Analysis (AREA)

Abstract

The present invention provides an image processing device (100) having: a first query acquisition unit (109) for acquiring a first query in which the attitude of each of a plurality of persons is designated; a second query acquisition unit (111) for acquiring a second query in which the physical distance and/or the temporal distance between the plurality of persons assuming the designated attitude is designated; and a search unit (105) for executing an image search based on the first and second queries.

Description

画像処理装置、画像処理方法、及びプログラムImage processing equipment, image processing methods, and programs
　本発明は、画像処理装置、画像処理方法、及びプログラムに関する。 The present invention relates to an image processing device, an image processing method, and a program.
　近年、監視システム等において、監視カメラの画像から人物の姿勢や行動等の状態の検出や検索を行う技術が利用されている。関連する技術として、例えば、特許文献１及び２が知られている。特許文献１には、深さ映像に含まれる人物の頭や手足等のキージョイントに基づいて、類似する人物の姿勢を検索する技術が開示されている。特許文献２には、人物の姿勢と関連しないが、画像に付加された傾き等の姿勢情報を利用して類似画像を検索する技術が開示されている。特許文献３には、画像から複数の特徴点で構成される検索対象の姿勢情報を取得し、その姿勢情報で特定される姿勢と類似した姿勢を含む画像を検索する技術が開示されている。なお、その他に、人物の骨格推定に関連する技術として、非特許文献１が知られている。 In recent years, in surveillance systems and the like, technology for detecting and searching states such as postures and behaviors of people from images of surveillance cameras has been used. As related techniques, for example, Patent Documents 1 and 2 are known. Patent Document 1 discloses a technique for searching the posture of a similar person based on key joints such as the head and limbs of the person included in the depth image. Patent Document 2 discloses a technique for searching for a similar image by using posture information such as tilt added to the image, although it is not related to the posture of the person. Patent Document 3 discloses a technique of acquiring posture information of a search target composed of a plurality of feature points from an image and searching for an image including a posture similar to the posture specified by the posture information. In addition, Non-Patent Document 1 is known as a technique related to the estimation of the skeleton of a person.
特表２０１４－５２２０３５号公報Special Table 2014-52035 Gazette 特開２００６－２６０４０５号公報Japanese Unexamined Patent Publication No. 2006-260405 特開２０１９－９１１３８号公報Japanese Unexamined Patent Publication No. 2019-9138
　特許文献１及び３は、所定の状態の人物を含む画像を検索する技術であるが、検索クエリとして指定できる条件が限られているので、検索できる対象も限られる。すなわち、特許文献１及び３の技術は、入力方法に改善の余地があった。特許文献２は、所定の状態の人物を含む画像を検索する技術でない。 Patent Documents 1 and 3 are techniques for searching for an image including a person in a predetermined state, but since the conditions that can be specified as a search query are limited, the searchable targets are also limited. That is, the techniques of Patent Documents 1 and 3 have room for improvement in the input method. Patent Document 2 is not a technique for searching an image including a person in a predetermined state.
　本発明の目的は、所定の状態の人物を含む画像を検索するシステムにおいて、検索できる対象の幅を拡張させることである。 An object of the present invention is to expand the range of searchable objects in a system for searching an image including a person in a predetermined state.
　本発明によれば、
　複数の人物各々の姿勢を指定した第１のクエリを取得する第１のクエリ取得手段と、
　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する第２のクエリ取得手段と、
　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する検索手段と、
を備える画像処理装置が提供される。 According to the present invention
A first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people,
A second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
A search means for executing an image search based on the first query and the second query, and
An image processing device comprising the above is provided.
　また、本発明によれば、
　コンピュータが、
　　複数の人物各々の姿勢を指定した第１のクエリを取得し、
　　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得し、
　　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する画像処理方法が提供される。 Further, according to the present invention,
The computer
Get the first query that specifies the posture of each of the multiple people
Obtain a second query that specifies at least one of the physical distance and the temporal distance between the plurality of people who take the specified posture.
An image processing method for executing an image search based on the first query and the second query is provided.
　また、本発明によれば、
　コンピュータを、
　　複数の人物各々の姿勢を指定した第１のクエリを取得する第１のクエリ取得手段、
　　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する第２のクエリ取得手段、
　　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する検索手段、
として機能させるプログラムが提供される。 Further, according to the present invention,
Computer,
A first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people,
A second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
A search means for performing an image search based on the first query and the second query.
A program is provided that functions as.
　本発明によれば、所定の状態の人物を含む画像を検索するシステムにおいて、検索できる対象の幅を拡張させることができる。 According to the present invention, in a system for searching an image including a person in a predetermined state, the range of searchable objects can be expanded.
実施の形態に係る画像処理装置の概要を示す構成図である。It is a block diagram which shows the outline of the image processing apparatus which concerns on embodiment. 実施の形態１に係る画像処理装置の構成を示す構成図である。It is a block diagram which shows the structure of the image processing apparatus which concerns on Embodiment 1. FIG. 実施の形態１に係る画像処理方法を示すフローチャートである。It is a flowchart which shows the image processing method which concerns on Embodiment 1. 実施の形態１に係る分類方法を示すフローチャートである。It is a flowchart which shows the classification method which concerns on Embodiment 1. 実施の形態１に係る検索方法を示すフローチャートである。It is a flowchart which shows the search method which concerns on Embodiment 1. 実施の形態１に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1. FIG. 実施の形態１に係る人体モデルを示す図である。It is a figure which shows the human body model which concerns on Embodiment 1. 実施の形態１に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1. FIG. 実施の形態１に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1. FIG. 実施の形態１に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 1. FIG. 実施の形態１に係る分類方法の具体例を示すグラフである。It is a graph which shows the specific example of the classification method which concerns on Embodiment 1. 実施の形態１に係る分類結果の表示例を示す図である。It is a figure which shows the display example of the classification result which concerns on Embodiment 1. FIG. 実施の形態１に係る検索方法を説明するための図である。It is a figure for demonstrating the search method which concerns on Embodiment 1. 実施の形態１に係る検索方法を説明するための図である。It is a figure for demonstrating the search method which concerns on Embodiment 1. 実施の形態１に係る検索方法を説明するための図である。It is a figure for demonstrating the search method which concerns on Embodiment 1. 実施の形態１に係る検索方法を説明するための図である。It is a figure for demonstrating the search method which concerns on Embodiment 1. 画像処理装置が出力する画面の一例を模式的に示す図である。It is a figure which shows typically an example of the screen which an image processing apparatus outputs. 実施の形態２に係る画像処理装置の構成を示す構成図である。It is a block diagram which shows the structure of the image processing apparatus which concerns on Embodiment 2. 実施の形態２に係る画像処理方法を示すフローチャートである。It is a flowchart which shows the image processing method which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法の具体例１を示すフローチャートである。It is a flowchart which shows the specific example 1 of the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法の具体例２を示すフローチャートである。It is a flowchart which shows the specific example 2 of the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法の具体例２を示すフローチャートである。It is a flowchart which shows the specific example 2 of the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る正規化方法を示すフローチャートである。It is a flowchart which shows the normalization method which concerns on Embodiment 2. 実施の形態２に係る人体モデルを示す図である。It is a figure which shows the human body model which concerns on Embodiment 2. 実施の形態２に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 2. 実施の形態２に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 2. 実施の形態２に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 2. 実施の形態２に係る人体モデルを示す図である。It is a figure which shows the human body model which concerns on Embodiment 2. 実施の形態２に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法を説明するためのヒストグラムである。It is a histogram for demonstrating the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る骨格構造の検出例を示す図である。It is a figure which shows the detection example of the skeleton structure which concerns on Embodiment 2. 実施の形態２に係る３次元人体モデルを示す図である。It is a figure which shows the 3D human body model which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法を説明するための図である。It is a figure for demonstrating the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法を説明するための図である。It is a figure for demonstrating the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る身長画素数算出方法を説明するための図である。It is a figure for demonstrating the height pixel number calculation method which concerns on Embodiment 2. 実施の形態２に係る正規化方法を説明するための図である。It is a figure for demonstrating the normalization method which concerns on Embodiment 2. 実施の形態２に係る正規化方法を説明するための図である。It is a figure for demonstrating the normalization method which concerns on Embodiment 2. 実施の形態２に係る正規化方法を説明するための図である。It is a figure for demonstrating the normalization method which concerns on Embodiment 2. 画像処理装置が処理する情報の一例を模式的に示す図である。It is a figure which shows an example of the information which an image processing apparatus processes. 画像処理装置が処理する情報の一例を模式的に示す図である。It is a figure which shows an example of the information which an image processing apparatus processes. 画像処理装置のハードウエア構成例を示す図である。It is a figure which shows the hardware configuration example of an image processing apparatus. 画像処理装置が出力する画面の一例を模式的に示す図である。It is a figure which shows typically an example of the screen which an image processing apparatus outputs. 画像処理装置が出力する画面の一例を模式的に示す図である。It is a figure which shows typically an example of the screen which an image processing apparatus outputs. 画像処理装置が出力する画面の一例を模式的に示す図である。It is a figure which shows typically an example of the screen which an image processing apparatus outputs. 画像処理装置が出力する画面の一例を模式的に示す図である。It is a figure which shows typically an example of the screen which an image processing apparatus outputs. 画像処理装置が検索できる対象の一例を示す図である。It is a figure which shows an example of the object which an image processing apparatus can search. 画像処理装置が検索できる対象の一例を示す図である。It is a figure which shows an example of the object which an image processing apparatus can search.
　以下、本発明の実施の形態について、図面を用いて説明する。尚、すべての図面において、同様な構成要素には同様の符号を付し、適宜説明を省略する。 Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all drawings, similar components are designated by the same reference numerals, and the description thereof will be omitted as appropriate.
（実施の形態に至る検討）
　近年、ディープラーニング等の機械学習を活用した画像認識技術が様々なシステムに応用されている。例えば、監視カメラの画像により監視を行う監視システムへの適用が進められている。監視システムに機械学習を活用することで、画像から人物の姿勢や行動等の状態をある程度把握することが可能とされつつある。 (Examination leading to the embodiment)
In recent years, image recognition technology utilizing machine learning such as deep learning has been applied to various systems. For example, it is being applied to a surveillance system that monitors images from a surveillance camera. By utilizing machine learning for monitoring systems, it is becoming possible to grasp the state of a person's posture, behavior, etc. to some extent from images.
　しかしながら、このような関連する技術では、必ずしもオンデマンドにユーザが望む人物の状態を把握できない場合がある。例えば、ユーザが検索し把握したい人物の状態を事前に特定できている場合もあれば、未知の状態のように具体的に特定できていない場合もある。そうすると、場合によっては、ユーザが検索したい人物の状態を詳細に指定することができない。また、人物の体の一部が隠れているような場合には検索等を行うことができない。関連する技術では、特定の検索条件のみからしか人物の状態を検索できないため、所望の人物の状態を柔軟に検索や分類することが困難である。 However, with such related technologies, it may not always be possible to grasp the state of the person desired by the user on demand. For example, there are cases where the state of a person that the user wants to search and grasp can be specified in advance, and there are cases where the state cannot be specifically specified, such as an unknown state. Then, in some cases, it is not possible to specify in detail the state of the person that the user wants to search. In addition, when a part of the body of a person is hidden, it is not possible to perform a search or the like. With related technology, it is difficult to flexibly search and classify a desired person's state because the person's state can be searched only from a specific search condition.
　発明者らは、オンデマンドに画像からユーザ所望の人物の状態を認識するため、非特許文献１などの骨格推定技術を利用する方法を検討した。非特許文献１に開示されたＯｐｅｎＰｏｓｅ等のように、関連する骨格推定技術では、様々なパターンの正解付けされた画像データを学習することで、人物の骨格を推定する。以下の実施の形態では、このような骨格推定技術を活用することで、人物の状態を柔軟に認識することを可能とする。 The inventors examined a method of using a skeleton estimation technique such as Non-Patent Document 1 in order to recognize the state of a person desired by a user from an image on demand. In a related skeleton estimation technique such as OpenPose disclosed in Non-Patent Document 1, the skeleton of a person is estimated by learning image data with various patterns correctly answered. In the following embodiments, it is possible to flexibly recognize the state of a person by utilizing such a skeleton estimation technique.
　なお、ＯｐｅｎＰｏｓｅ等の骨格推定技術により推定される骨格構造は、関節等の特徴的な点である「キーポイント」と、キーポイント間のリンクを示す「ボーン（ボーンリンク）」とから構成される。このため、以下の実施の形態では、骨格構造について「キーポイント」と「ボーン」という用語を用いて説明するが、特に限定されない限り、「キーポイント」は人物の「関節」に対応し、「ボーン」は人物の「骨」に対応している。 The skeletal structure estimated by a skeletal estimation technique such as OpenPose is composed of "key points" which are characteristic points of joints and the like and "bones (bone links)" which indicate links between key points. .. Therefore, in the following embodiments, the skeletal structure will be described using the terms "key point" and "bone", but unless otherwise specified, the "key point" corresponds to the "joint" of a person, and " "Bone" corresponds to the "bone" of a person.
（実施の形態の概要）
　図１は、実施の形態に係る画像処理装置１０の概要を示している。図１に示すように、画像処理装置１０は、骨格検出部１１、特徴量算出部１２、及び認識部１３を備えている。骨格検出部１１は、カメラ等から取得される２次元画像に基づいて、複数の人物の２次元骨格構造（以下、単に「骨格構造」という場合がある）を検出する。特徴量算出部１２は、骨格検出部１１により検出された複数の２次元骨格構造の特徴量を算出する。認識部１３は、特徴量算出部１２により算出された複数の特徴量の類似度に基づいて、複数の人物の状態の認識処理を行う。認識処理は、人物の状態の分類処理や検索処理等である。 (Outline of embodiment)
FIG. 1 shows an outline of the image processing apparatus 10 according to the embodiment. As shown in FIG. 1, the image processing device 10 includes a skeleton detection unit 11, a feature amount calculation unit 12, and a recognition unit 13. The skeleton detection unit 11 detects a two-dimensional skeleton structure (hereinafter, may be simply referred to as “skeleton structure”) of a plurality of persons based on a two-dimensional image acquired from a camera or the like. The feature amount calculation unit 12 calculates the feature amount of the plurality of two-dimensional skeleton structures detected by the skeleton detection unit 11. The recognition unit 13 performs a recognition process for the states of a plurality of persons based on the similarity of the plurality of feature amounts calculated by the feature amount calculation unit 12. The recognition process is a classification process of a person's state, a search process, or the like.
　このように、実施の形態では、２次元画像から人物の２次元骨格構造を検出し、この２次元骨格構造から算出される特徴量に基づいて人物の状態の分類や検索等の認識処理を行う。 As described above, in the embodiment, the two-dimensional skeleton structure of the person is detected from the two-dimensional image, and the recognition process such as classification and search of the state of the person is performed based on the feature amount calculated from the two-dimensional skeleton structure. ..
（実施の形態１）
　以下、図面を参照して実施の形態１について説明する。図２は、本実施の形態に係る画像処理装置１００の構成を示している。画像処理装置１００は、カメラ２００及び記憶手段（データベース（ＤＢ）１１０）、とともに画像処理システム１を構成する。画像処理装置１００を含む画像処理システム１は、画像から推定される人物の骨格構造に基づき、人物の姿勢や行動等の状態を分類及び検索するシステムである。 (Embodiment 1)
Hereinafter, the first embodiment will be described with reference to the drawings. FIG. 2 shows the configuration of the image processing apparatus 100 according to the present embodiment. The image processing device 100 constitutes an image processing system 1 together with a camera 200 and a storage means (database (DB) 110). The image processing system 1 including the image processing device 100 is a system for classifying and searching states such as posture and behavior of a person based on the skeleton structure of the person estimated from the image.
　カメラ２００は、２次元の画像を生成する監視カメラ等の撮像部である。カメラ２００は、所定の箇所に設置されて、その設置箇所から撮像領域における人物等を撮像する。カメラ２００は、撮像した画像（映像）を画像処理装置１００へ出力可能に有線又は無線で直接接続、もしくは任意の通信ネットワーク等を介して接続されている。なお、カメラ２００を画像処理装置１００の内部に設けてもよい。 The camera 200 is an image pickup unit such as a surveillance camera that generates a two-dimensional image. The camera 200 is installed at a predetermined location and captures a person or the like in the imaging region from the installation location. The camera 200 is directly connected by wire or wirelessly so that the captured image (video) can be output to the image processing device 100, or is connected via an arbitrary communication network or the like. The camera 200 may be provided inside the image processing device 100.
　データベース１１０は、画像処理装置１００の処理に必要な情報（データ）や処理結果等を格納するデータベースである。データベース１１０は、画像取得部１０１が取得した画像や、骨格構造検出部１０２の検出結果、機械学習用のデータ、特徴量算出部１０３が算出した特徴量、分類部１０４の分類結果、検索部１０５の検索結果等を記憶する。データベース１１０は、画像処理装置１００と必要に応じてデータを入出力可能に有線又は無線で直接接続、もしくは任意の通信ネットワーク等を介して接続されている。なお、データベース１１０をフラッシュメモリなどの不揮発性メモリやハードディスク装置等として、画像処理装置１００の内部に設けてもよい。 The database 110 is a database that stores information (data), processing results, and the like necessary for processing of the image processing apparatus 100. The database 110 includes an image acquired by the image acquisition unit 101, a detection result of the skeletal structure detection unit 102, data for machine learning, a feature amount calculated by the feature amount calculation unit 103, a classification result of the classification unit 104, and a search unit 105. The search results etc. of are memorized. The database 110 is directly connected to the image processing device 100 by wire or wirelessly so that data can be input and output as needed, or is connected via an arbitrary communication network or the like. The database 110 may be provided inside the image processing device 100 as a non-volatile memory such as a flash memory, a hard disk device, or the like.
　図２に示すように、画像処理装置１００は、画像取得部１０１、骨格構造検出部１０２、特徴量算出部１０３、分類部１０４、検索部１０５、入力部１０６、表示部１０７、第１のクエリ取得部１０９及び第２のクエリ取得部１１１を備えている。なお、各部（ブロック）の構成は一例であり、後述の方法（動作）が可能であれば、その他の各部で構成されてもよい。また、画像処理装置１００は、例えば、プログラムを実行するパーソナルコンピュータやサーバ等のコンピュータ装置で実現されるが、１つの装置で実現してもよいし、ネットワーク上の複数の装置で実現してもよい。例えば、入力部１０６や表示部１０７等を外部の装置としてもよい。また、分類部１０４及び検索部１０５の両方を備えていてもよいし、いずれか一方のみを備えていてもよい。分類部１０４及び検索部１０５の両方、もしくは一方は、人物の状態の認識処理を行う認識部１３である。また、検索部１０５、第１のクエリ取得部１０９及び第２のクエリ取得部１１１は、検索処理を実行する機能部であり、図１の認識部１３に該当する。 As shown in FIG. 2, the image processing apparatus 100 includes an image acquisition unit 101, a skeleton structure detection unit 102, a feature amount calculation unit 103, a classification unit 104, a search unit 105, an input unit 106, a display unit 107, and a first query. The acquisition unit 109 and the second query acquisition unit 111 are provided. The configuration of each part (block) is an example, and may be composed of other parts as long as the method (operation) described later is possible. Further, the image processing device 100 is realized by, for example, a computer device such as a personal computer or a server that executes a program, but it may be realized by one device or by a plurality of devices on a network. good. For example, the input unit 106, the display unit 107, and the like may be used as an external device. Further, both the classification unit 104 and the search unit 105 may be provided, or only one of them may be provided. Both or one of the classification unit 104 and the search unit 105 is the recognition unit 13 that performs the recognition processing of the state of the person. Further, the search unit 105, the first query acquisition unit 109, and the second query acquisition unit 111 are functional units that execute the search process, and correspond to the recognition unit 13 in FIG.
　画像処理装置１００は、データ蓄積処理と、分類処理と、検索処理とをこの順に実行する。なお、以下で説明するように、画像処理装置１００は分類処理を実行しなくてもよい。 The image processing device 100 executes data storage processing, classification processing, and search processing in this order. As will be described below, the image processing apparatus 100 does not have to execute the classification process.
＜データ蓄積処理＞
　データ蓄積処理は、解析対象の画像（以下、「解析対象画像」）を取得し、複数の解析対象画像各々から人物の２次元の骨格構造を検出し、検出された２次元の骨格構造の特徴量を算出し、算出した特徴量を各解析対象画像に紐付けてデータベース１１０に格納する処理である。以下、データ蓄積処理に関わる機能部の構成を説明する。 <Data storage processing>
The data storage process acquires an image to be analyzed (hereinafter referred to as "image to be analyzed"), detects a two-dimensional skeletal structure of a person from each of a plurality of images to be analyzed, and features the detected two-dimensional skeletal structure. This is a process of calculating the amount, associating the calculated feature amount with each analysis target image, and storing the calculated feature amount in the database 110. Hereinafter, the configuration of the functional unit related to the data storage process will be described.
　画像取得部１０１は、解析対象画像を取得する。本明細書において、「取得」とは、ユーザ入力に基づき、又は、プログラムの指示に基づき、「自装置が他の装置や記憶媒体に格納されているデータを取りに行くこと（能動的な取得）」、たとえば、他の装置にリクエストまたは問い合わせして受信すること、他の装置や記憶媒体にアクセスして読み出すこと等、および、ユーザ入力に基づき、又は、プログラムの指示に基づき、「自装置に他の装置から出力されるデータを入力すること（受動的な取得）」、たとえば、配信（または、送信、プッシュ通知等）されるデータを受信すること、また、受信したデータまたは情報の中から選択して取得すること、及び、「データを編集（テキスト化、データの並び替え、一部データの抽出、ファイル形式の変更等）などして新たなデータを生成し、当該新たなデータを取得すること」の少なくともいずれか一方を含む。 The image acquisition unit 101 acquires the image to be analyzed. In the present specification, "acquisition" means "acquisition of data stored in another device or storage medium by the own device" based on user input or program instruction (active acquisition). ) ”, For example, requesting or inquiring about another device to receive it, accessing and reading another device or storage medium, etc., and based on user input or program instructions,“ own device To input data output from other devices (passive acquisition) ", for example, to receive data to be delivered (or transmitted, push notification, etc.), and in the received data or information. Generate new data by selecting from and acquiring, and editing data (text conversion, data sorting, partial data extraction, file format change, etc.), and the new data is generated. Includes at least one of "getting".
　例えば、画像取得部１０１は、所定の監視期間にカメラ２００が撮像した人物を含む２次元の画像を解析対象画像として取得する。画像取得部１０１は、動画像を取得してもよい。この場合、動画像を構成する複数のフレーム画像各々が、解析対象画像となる。その他、画像取得部１０１は、不定期に、又は、動画像のフレームレートよりも大きい時間間隔で定期的に生成された静止画像を解析対象画像として取得してもよい。その他、画像取得部１０１は、データベース１１０等の記憶手段に記憶されている人物を含む２次元の画像を解析対象画像として取得してもよい。 For example, the image acquisition unit 101 acquires a two-dimensional image including a person captured by the camera 200 during a predetermined monitoring period as an analysis target image. The image acquisition unit 101 may acquire a moving image. In this case, each of the plurality of frame images constituting the moving image becomes the analysis target image. In addition, the image acquisition unit 101 may acquire a still image generated irregularly or at a time interval larger than the frame rate of the moving image as an analysis target image. In addition, the image acquisition unit 101 may acquire a two-dimensional image including a person stored in a storage means such as a database 110 as an analysis target image.
　骨格構造検出部１０２は、取得された解析対象画像各々から人物の２次元の骨格構造を検出する。骨格構造検出部１０２は、解析対象画像の中で認識される全ての人物について、骨格構造を検出することができる。骨格構造検出部１０２は、機械学習を用いた骨格推定技術を用いて、認識される人物の関節等の特徴に基づき人物の骨格構造を検出する。骨格構造検出部１０２は、例えば、非特許文献１のＯｐｅｎＰｏｓｅ等の骨格推定技術を用い、関節等の特徴的な点であるキーポイントを抽出したりする。 The skeleton structure detection unit 102 detects the two-dimensional skeleton structure of a person from each of the acquired images to be analyzed. The skeleton structure detection unit 102 can detect the skeleton structure of all the persons recognized in the analysis target image. The skeleton structure detection unit 102 detects the skeleton structure of a person based on the characteristics such as the joints of the person to be recognized by using the skeleton estimation technique using machine learning. The skeleton structure detection unit 102 uses, for example, a skeleton estimation technique such as OpenPose of Non-Patent Document 1 to extract key points that are characteristic points of joints and the like.
　特徴量算出部１０３は、検出された２次元の骨格構造の特徴量を算出し、算出した特徴量を、その２次元の骨格構造が検出された解析対象画像に紐付けてデータベース１１０に格納する。骨格構造の特徴量は、人物の骨格の特徴を示しており、人物の骨格に基づいて人物の状態を分類や検索するための要素となる。通常、この特徴量は、複数のパラメータ（例えば後述する分類要素）を含んでいる。特徴量は、骨格構造の全体の特徴量でもよいし、骨格構造の一部の特徴量でもよく、骨格構造の各部のように複数の特徴量を含んでもよい。特徴量の算出方法は、機械学習や正規化等の任意の方法でよく、正規化として最小値や最大値を求めてもよい。一例として、特徴量は、骨格構造を機械学習することで得られた特徴量や、骨格構造の頭部から足部までの画像上の大きさ等である。骨格構造の大きさは、画像上の骨格構造を含む骨格領域の上下方向の高さや面積等である。上下方向（高さ方向または縦方向）は、画像における上下の方向（Ｙ軸方向）であり、例えば、地面（基準面）に対し垂直な方向である。また、左右方向（横方向）は、画像における左右の方向（Ｘ軸方向）であり、例えば、地面に対し平行な方向である。 The feature amount calculation unit 103 calculates the feature amount of the detected two-dimensional skeletal structure, associates the calculated feature amount with the analysis target image in which the two-dimensional skeletal structure is detected, and stores it in the database 110. .. The feature amount of the skeletal structure indicates the characteristics of the skeleton of the person, and is an element for classifying or searching the state of the person based on the skeleton of the person. Usually, this feature quantity includes a plurality of parameters (for example, classification elements described later). The feature amount may be an entire feature amount of the skeletal structure, a partial feature amount of the skeletal structure, or a plurality of feature amounts such as each part of the skeletal structure. The feature amount may be calculated by any method such as machine learning or normalization, and the minimum value or the maximum value may be obtained as the normalization. As an example, the feature amount is a feature amount obtained by machine learning the skeletal structure, a size on an image of the skeletal structure from the head to the foot, and the like. The size of the skeleton structure is the vertical height and area of the skeleton region including the skeleton structure on the image. The vertical direction (height direction or vertical direction) is a vertical direction (Y-axis direction) in the image, and is, for example, a direction perpendicular to the ground (reference plane). Further, the left-right direction (horizontal direction) is a left-right direction (X-axis direction) in the image, and is, for example, a direction parallel to the ground.
　なお、ユーザが望む分類や検索を行うためには、分類や検索処理に対しロバスト性を有する特徴量を用いることが好ましい。例えば、ユーザが、人物の向きや体型に依存しない分類や検索を望む場合、人物の向きや体型にロバストな特徴量を使用してもよい。同じ姿勢で様々な方向に向いている人物の骨格や同じ姿勢で様々な体型の人物の骨格を学習することや、骨格の上下方向のみの特徴を抽出することで、人物の向きや体型に依存しない特徴量を得ることができる。 In addition, in order to perform the classification and search desired by the user, it is preferable to use a feature amount having robustness for the classification and search processing. For example, if the user desires a classification or search that does not depend on the orientation or body shape of the person, a robust feature amount may be used for the orientation or body shape of the person. It depends on the orientation and body shape of the person by learning the skeleton of a person who is facing in various directions with the same posture and the skeleton of a person with various body shapes in the same posture, and by extracting the characteristics of the skeleton only in the vertical direction. It is possible to obtain a feature amount that does not.
＜分類処理＞
　分類処理は、データ蓄積処理でデータベース１１０に格納されたデータ（解析対象画像と、各解析対象画像から検出された２次元の骨格構造の特徴量とを紐付けたデータ）に基づき、解析対象画像から検出された複数の２次元の骨格構造をその特徴量が類似するもの同士でまとめて分類（グループ分け）する処理である。なお、解析対象画像と、各解析対象画像から検出された２次元の骨格構造とは互いに紐付けられている。このため、分類処理による複数の２次元の骨格構造の分類は、複数の解析対象画像の分類にもなる。分類処理により、複数の解析対象画像は、類似する２次元の骨格構造を含むもの同士でまとめられる。以下、分類処理に関わる機能部の構成を説明する。 <Classification process>
The classification process is based on the data stored in the database 110 in the data storage process (data in which the image to be analyzed is associated with the feature amount of the two-dimensional skeletal structure detected from each image to be analyzed). This is a process of collectively classifying (grouping) a plurality of two-dimensional skeletal structures detected from the above by those having similar features. The analysis target image and the two-dimensional skeleton structure detected from each analysis target image are associated with each other. Therefore, the classification of a plurality of two-dimensional skeletal structures by the classification process also becomes the classification of a plurality of images to be analyzed. By the classification process, a plurality of images to be analyzed are grouped together with images containing similar two-dimensional skeletal structures. Hereinafter, the configuration of the functional unit related to the classification process will be described.
　分類部１０４は、データベース１１０に格納された複数の骨格構造を、骨格構造の特徴量の類似度に基づいて分類する（クラスタリングする）。分類部１０４は、人物の状態の認識処理として、骨格構造の特徴量に基づいて複数の人物の状態を分類しているとも言える。類似度は、骨格構造の特徴量間の距離である。分類部１０４は、骨格構造の全体の特徴量の類似度により分類してもよいし、骨格構造の一部の特徴量の類似度により分類してもよく、骨格構造の第１の部分（例えば両手）及び第２の部分（例えば両足）の特徴量の類似度により分類してもよい。なお、各画像における人物の骨格構造の特徴量に基づいて人物の姿勢を分類してもよいし、時系列に連続する複数の画像における人物の骨格構造の特徴量の変化に基づいて人物の行動を分類してもよい。すなわち、分類部１０４は、骨格構造の特徴量に基づいて人物の姿勢や行動を含む人物の状態を分類できる。例えば、分類部１０４は、所定の監視期間に撮像された複数の画像における複数の骨格構造を分類対象とする。分類部１０４は、分類対象の特徴量間の類似度を求め、類似度の高い骨格構造が同じクラスタ（似た姿勢のグループ）となるように分類する。なお、検索と同様に、分類条件をユーザが指定できるようにしてもよい。分類部１０４は、骨格構造の分類結果をデータベース１１０に格納するとともに、表示部１０７に表示することができる。 The classification unit 104 classifies (clusters) a plurality of skeletal structures stored in the database 110 based on the similarity of the feature amounts of the skeletal structures. It can be said that the classification unit 104 classifies the states of a plurality of persons based on the feature amount of the skeletal structure as the process of recognizing the state of the person. Similarity is the distance between features of the skeletal structure. The classification unit 104 may be classified according to the similarity of the features of the whole skeleton structure, or may be classified according to the similarity of the features of a part of the skeleton structure, and the first part of the skeleton structure (for example, for example). It may be classified according to the similarity of the features of both hands) and the second part (for example, both feet). The posture of the person may be classified based on the feature amount of the skeletal structure of the person in each image, or the behavior of the person based on the change in the feature amount of the skeletal structure of the person in a plurality of consecutive images in time series. May be classified. That is, the classification unit 104 can classify the state of the person including the posture and behavior of the person based on the feature amount of the skeletal structure. For example, the classification unit 104 targets a plurality of skeletal structures in a plurality of images captured during a predetermined monitoring period. The classification unit 104 obtains the degree of similarity between the features to be classified, and classifies the skeletal structures having a high degree of similarity into the same cluster (group with similar postures). As with the search, the user may be able to specify the classification conditions. The classification unit 104 can store the classification result of the skeletal structure in the database 110 and display it on the display unit 107.
＜検索処理＞
　検索処理は、データ蓄積処理でデータベース１１０に格納されたデータ（解析対象画像と、各解析対象画像から検出された２次元の骨格構造の特徴量とを紐付けたデータ）に基づき、解析対象画像の中から所定のシーンを検索する処理である。所定のシーンは、複数の人物各々の姿勢、複数の人物各々が指定された姿勢をとるタイミングの時間的距離、複数の人物の物理的距離等により特定される。以下、検索処理に関わる機能部の構成を説明する。 <Search process>
The search process is based on the data stored in the database 110 in the data storage process (data in which the image to be analyzed is associated with the feature amount of the two-dimensional skeletal structure detected from each image to be analyzed). It is a process of searching for a predetermined scene from the inside. A predetermined scene is specified by the posture of each of the plurality of persons, the time distance of the timing at which each of the plurality of persons takes a designated posture, the physical distance of the plurality of persons, and the like. Hereinafter, the configuration of the functional unit related to the search process will be described.
　第１のクエリ取得部１０９は、複数の人物各々の姿勢を指定した第１のクエリを取得する。複数の人物の姿勢は、同じ姿勢であってもよいし、異なる姿勢であってもよい。また、複数の人物は、２人であってもよいし、３人以上であってもよい。複数の人物各々の姿勢は、ユーザ入力により指定される。 The first query acquisition unit 109 acquires the first query in which the postures of each of the plurality of persons are specified. The postures of the plurality of persons may be the same posture or different postures. Further, the plurality of persons may be two persons or three or more persons. The posture of each of the plurality of persons is specified by user input.
　ここで、第１のクエリの取得方法の例を説明する。 Here, an example of the acquisition method of the first query will be described.
「第１のクエリ取得例１」
　当該例では、第１のクエリ取得部１０９は、予め用意された複数の選択肢（人物の姿勢）を、ユーザに向けて表示する。そして、第１のクエリ取得部１０９は、当該選択肢の中からユーザが選択した姿勢を、第１のクエリとして取得する。第１のクエリ取得部１０９は、複数の人物各々に対応して、当該選択を受付けることができる。 "First query acquisition example 1"
In this example, the first query acquisition unit 109 displays a plurality of options (postures of a person) prepared in advance toward the user. Then, the first query acquisition unit 109 acquires the posture selected by the user from the options as the first query. The first query acquisition unit 109 can accept the selection corresponding to each of the plurality of persons.
　例えば、「胡坐をかいて座る」、「正座」、「仰向けで寝そべる」等のようなキーワードで複数の選択肢が提示されてもよい。その他、複数の姿勢各々を示した画像で、複数の選択肢が提示されてもよい。姿勢を示した画像は、図７に示すような２次元骨格構造のモデルの画像であってもよいし、図８に示すような人物と２次元骨格構造のモデルを含む画像であってもよいし、人物を含み、２次元骨格構造のモデルを含まない画像であってもよい。 For example, multiple options may be presented with keywords such as "sitting with crossed legs", "seiza", and "lying on the back". In addition, a plurality of options may be presented with an image showing each of the plurality of postures. The image showing the posture may be an image of a model of a two-dimensional skeleton structure as shown in FIG. 7, or an image including a person and a model of the two-dimensional skeleton structure as shown in FIG. However, the image may include a person and do not include a model of a two-dimensional skeleton structure.
「第１のクエリ取得例２」
　当該例では、第１のクエリ取得部１０９は、図７に示すような２次元骨格構造のモデルの画像を表示する。そして、第１のクエリ取得部１０９は、当該モデルの姿勢を変更する入力、及び、変更後の姿勢を第１のクエリとして決定するユーザ入力を受付ける。当該モデルの姿勢の変更は、例えば、複数のキーポイント各々の位置を移動させる入力で実現される。 "First query acquisition example 2"
In this example, the first query acquisition unit 109 displays an image of a model having a two-dimensional skeleton structure as shown in FIG. 7. Then, the first query acquisition unit 109 accepts an input for changing the posture of the model and a user input for determining the changed posture as the first query. The change of the posture of the model is realized by, for example, an input for moving the position of each of a plurality of key points.
「第１のクエリ取得例３」
　当該例では、第１のクエリ取得部１０９は、画像取得部１０１、骨格構造検出部１０２及び特徴量算出部１０３と協働して、第１のクエリを取得する。画像取得部１０１は、クエリ画像を取得する。骨格構造検出部１０２は、クエリ画像に含まれる人物の２次元骨格構造を検出する。特徴量算出部１０３は、検出された２次元骨格構造の特徴量を算出する。そして、第１のクエリ取得部１０９は、クエリ画像から算出された２次元骨格構造の特徴量を第１のクエリとして取得する。 "First query acquisition example 3"
In this example, the first query acquisition unit 109 acquires the first query in cooperation with the image acquisition unit 101, the skeleton structure detection unit 102, and the feature amount calculation unit 103. The image acquisition unit 101 acquires a query image. The skeleton structure detection unit 102 detects the two-dimensional skeleton structure of the person included in the query image. The feature amount calculation unit 103 calculates the feature amount of the detected two-dimensional skeleton structure. Then, the first query acquisition unit 109 acquires the feature amount of the two-dimensional skeleton structure calculated from the query image as the first query.
　画像取得部１０１は、例えば以下の取得例のいずれかにより、クエリ画像を取得することができる。 The image acquisition unit 101 can acquire a query image by, for example, any of the following acquisition examples.
「クエリ画像取得例１」
　当該例では、画像取得部１０１は、ユーザが用意し、画像処理装置１００に入力したクエリ画像を取得する。 "Query image acquisition example 1"
In this example, the image acquisition unit 101 acquires a query image prepared by the user and input to the image processing device 100.
「クエリ画像取得例２」
　当該例では、画像取得部１０１は、ユーザが指定したキーワードで検索した画像をクエリ画像として取得する。図４２に、当該検索で利用されるＵＩ（user interface）画面の一例を示す。キーワードの入力欄、及び、検索結果の表示欄が示されている。当該ＵＩ画面の場合、検索結果として表示された画像の中からユーザが選択した画像が、クエリ画像となる。キーワードは、「座る」、「立つ」等のように、人物の状態（姿勢、行動等）に関する内容が想定される。キーワードの入力は、例えばテキストボックス、ドロップダウン、チェックボックス等の周知のＧＵＩを利用して実現できる。 "Query image acquisition example 2"
In this example, the image acquisition unit 101 acquires an image searched by a keyword specified by the user as a query image. FIG. 42 shows an example of a UI (user interface) screen used in the search. An input field for keywords and a display field for search results are shown. In the case of the UI screen, the image selected by the user from the images displayed as the search result is the query image. The keywords are assumed to be related to the state (posture, behavior, etc.) of the person, such as "sitting" and "standing". Keyword input can be realized using a well-known GUI such as a text box, a dropdown, or a check box.
　例えば、予め、図３９に示すように、クエリ画像として利用されるために用意された画像（以下、「クエリ用の画像」）と、キーワード（各画像に含まれる人物の状態を示すワード）とを紐付けた情報が、データベース１１０に登録されてもよい。そして、画像取得部１０１は、入力されたキーワードが紐付けられたクエリ用の画像を当該情報の中から検索し、検索結果に含まれるクエリ用の画像の一部又は全部をクエリ画像として取得してもよい。 For example, as shown in FIG. 39, an image prepared in advance for use as a query image (hereinafter, “image for query”) and a keyword (a word indicating the state of a person included in each image). The information associated with may be registered in the database 110. Then, the image acquisition unit 101 searches for a query image associated with the input keyword from the information, and acquires a part or all of the query image included in the search result as a query image. You may.
　その他、図４０に示すように、解析対象画像の一部と、キーワード（各画像に含まれる人物の状態を示すワード）とを紐付けた情報が、データベース１１０に登録されてもよい。そして、画像取得部１０１は、入力されたキーワードが紐付けられた解析対象画像を当該情報の中から検索し、検索結果に含まれる解析対象画像の一部又は全部をクエリ画像として取得してもよい。 In addition, as shown in FIG. 40, information associated with a part of the image to be analyzed and a keyword (a word indicating the state of a person included in each image) may be registered in the database 110. Then, the image acquisition unit 101 may search the analysis target image associated with the input keyword from the information and acquire a part or all of the analysis target image included in the search result as a query image. good.
　その他、画像取得部１０１は、キーワードに関連する画像を検索する検索エンジンに、入力されたキーワードを送信し、当該検索エンジンから検索結果を取得してもよい。そして、画像取得部１０１は、検索結果に含まれる画像の一部又は全部をクエリ画像として取得してもよい。 In addition, the image acquisition unit 101 may send the input keyword to a search engine that searches for images related to the keyword, and acquire the search result from the search engine. Then, the image acquisition unit 101 may acquire a part or all of the images included in the search result as a query image.
　なお、上述したいずれのクエリ画像取得例においても、取得されたクエリ画像は、１人の人物を含む場合もあれば、複数の人物を含む場合もある。クエリ画像が複数の人物を含む場合、第１のクエリ取得部１０９は、ユーザ入力に基づき、クエリ画像に含まれる複数の人物の中から少なくとも１人の人物を選択する。そして、選択された人物の２次元骨格構造の特徴量を第１のクエリとして取得する。第１のクエリ取得部１０９は、例えば以下の２つの手法のいずれかで、当該選択を実現することができる。 In any of the above-mentioned query image acquisition examples, the acquired query image may include one person or a plurality of persons. When the query image includes a plurality of persons, the first query acquisition unit 109 selects at least one person from the plurality of persons included in the query image based on the user input. Then, the feature amount of the two-dimensional skeleton structure of the selected person is acquired as the first query. The first query acquisition unit 109 can realize the selection by, for example, one of the following two methods.
「クエリ画像に含まれる少なくとも１人の人物を選択する処理例１」
　当該例では、第１のクエリ取得部１０９は、自由に画像内の一部領域を指定するユーザ入力を受付ける。そして、第１のクエリ取得部１０９は、指定された一部領域内で検出された人物を選択する。「指定された一部領域内で検出された人物」は、指定された一部領域内に身体が完全に包含される人物、または、指定された一部領域内に身体の少なくとも一部が含まれる人物である。 "Processing example 1 to select at least one person included in the query image"
In this example, the first query acquisition unit 109 freely accepts user input for designating a part of the image. Then, the first query acquisition unit 109 selects a person detected in the designated partial area. "Person detected in a specified partial area" is a person whose body is completely contained in the specified partial area, or at least a part of the body is included in the specified partial area. Is a person who is
　例えば、画像取得部１０１が、図４３に示すようなクエリ画像Ｐを取得したとする。第１のクエリ取得部１０９は、当該クエリ画像Ｐを表示部１０７に表示させ、当該クエリ画像Ｐの中で少なくとも一部領域を指定するユーザ入力を受付ける。例えば、第１のクエリ取得部１０９は、図４４に示すように枠Ｗで一部領域を囲むユーザ入力により、一部領域を指定するユーザ入力を受付けてもよい。枠Ｗの画像内の位置及び大きさはユーザが自由に指定可能となっている。クエリ画像Ｐの中で少なくとも一部領域を指定するユーザ入力を受付けた後、第１のクエリ取得部１０９は、指定された一部領域内で検出された人物を選択する。 For example, it is assumed that the image acquisition unit 101 has acquired the query image P as shown in FIG. 43. The first query acquisition unit 109 causes the display unit 107 to display the query image P, and accepts user input that specifies at least a part of the query image P. For example, as shown in FIG. 44, the first query acquisition unit 109 may accept user input for designating a partial area by user input surrounding a partial area with a frame W. The position and size of the frame W in the image can be freely specified by the user. After accepting the user input for designating at least a part of the query image P, the first query acquisition unit 109 selects a person detected in the designated part of the area.
「クエリ画像に含まれる少なくとも１人の人物を選択する処理例２」
　当該例では、人物を選択するユーザ入力を受付ける前に、クエリ画像全体（画像内の全領域）に対して人物を検出する処理が実行される。そして、第１のクエリ取得部１０９は、図４５に示すように、クエリ画像上に、当該クエリ画像内で検出された人物を選択可能な人物として識別可能に表示した加工画像を表示する。枠Ｙで囲まれた人物が、選択可能な人物である。そして、第１のクエリ取得部１０９は、選択可能な人物の中から少なくとも１人を選択するユーザ入力を受付ける。ここで、人物を検出する処理、及び、選択可能な人物として表示する人物を決定する処理の一例を説明する。 "Processing example 2 for selecting at least one person included in the query image"
In this example, the process of detecting a person for the entire query image (the entire area in the image) is executed before accepting the user input for selecting the person. Then, as shown in FIG. 45, the first query acquisition unit 109 displays a processed image in which the person detected in the query image is identifiable as a selectable person on the query image. The person surrounded by the frame Y is a selectable person. Then, the first query acquisition unit 109 accepts a user input for selecting at least one of the selectable persons. Here, an example of a process of detecting a person and a process of determining a person to be displayed as a selectable person will be described.
－人物を検出する処理、及び、選択可能な人物として表示する人物を決定する処理例１－
　第１のクエリ取得部１０９は、周知の人物を検出する処理に基づき、クエリ画像全体（画像内の全領域）に対して人物を検出する処理を実行する。そして、第１のクエリ取得部１０９は、検出したすべての人物を、選択可能な人物とする。その他、第１のクエリ取得部１０９は、検出した人物の中の、画像内で占める領域（画像内の人物の大きさを意味する）が基準値以上の人物を、選択可能な人物としてもよい。画像内で占める領域が所定レベルより小さい人物は、２次元骨格構造の検出や特徴量の算出が困難になる。このような人物の姿勢を検索のためのクエリとすると、所望の検索結果が得られにくくなる。そこで、このような人物は選択可能な人物から除外してもよい。 -Process to detect a person and process to determine the person to be displayed as a selectable person 1-
The first query acquisition unit 109 executes a process of detecting a person for the entire query image (the entire area in the image) based on the process of detecting a well-known person. Then, the first query acquisition unit 109 sets all the detected persons as selectable persons. In addition, the first query acquisition unit 109 may select a person whose area occupied in the image (meaning the size of the person in the image) is equal to or larger than the reference value among the detected persons as a selectable person. .. For a person whose area occupied in the image is smaller than a predetermined level, it becomes difficult to detect the two-dimensional skeleton structure and calculate the feature amount. If such a posture of a person is used as a query for a search, it becomes difficult to obtain a desired search result. Therefore, such a person may be excluded from the selectable persons.
－人物を検出する処理、及び、選択可能な人物として表示する人物を決定する処理例２－
　骨格構造検出部１０２は、クエリ画像全体（画像内の全領域）に対し人物の２次元骨格構造を検出する処理を実行する。そして、第１のクエリ取得部１０９は、骨格構造検出部１０２による２次元骨格構造の検出結果が所定条件を満たす人物を、選択可能な人物とする。 -Process to detect a person and process example to determine the person to be displayed as a selectable person 2-
The skeleton structure detection unit 102 executes a process of detecting the two-dimensional skeleton structure of a person for the entire query image (the entire area in the image). Then, the first query acquisition unit 109 sets a person whose detection result of the two-dimensional skeleton structure by the skeleton structure detection unit 102 satisfies a predetermined condition as a selectable person.
　所定条件を満たす人物は、「抽出されたキーポイントの数が基準値以上の人物」、又は、「抽出されたキーポイントの数、及び抽出されたキーポイント各々の信頼度の少なくとも一方に基づき算出された評価値が基準値以上の人物」である。 A person who meets a predetermined condition is calculated based on at least one of "a person whose number of extracted key points is equal to or more than a reference value" or "the number of extracted key points and the reliability of each of the extracted key points". A person whose evaluation value is equal to or higher than the standard value. "
　評価値算出の詳細なアルゴリズムは特段制限されないが、以下の内容を満たすように設計される。
・抽出されたキーポイントの数が多いほど、評価値が高い。
・抽出されたキーポイントの信頼度が高いほど、評価値が高い。
・画像内における人物が大きいほど評価値が高い。 The detailed algorithm for calculating the evaluation value is not particularly limited, but it is designed to satisfy the following contents.
-The larger the number of extracted key points, the higher the evaluation value.
-The higher the reliability of the extracted key points, the higher the evaluation value.
-The larger the person in the image, the higher the evaluation value.
　この例の場合、図４５に示す画面において、選択可能に表示された人物各々に対応付けて、各々の評価値を表示してもよい。また、図４２に示すようなクエリ画像の検索結果において、検索結果に含まれる複数のクエリ画像各々に対応付けて各々の評価値を表示してもよい。クエリ画像が複数の人物を含む場合、複数の人物各々の評価値の統計値（最大値、最小値、平均値、最頻値、中央値等）を表示してもよい。 In the case of this example, on the screen shown in FIG. 45, each evaluation value may be displayed in association with each of the selectably displayed persons. Further, in the search result of the query image as shown in FIG. 42, each evaluation value may be displayed in association with each of the plurality of query images included in the search result. When the query image contains a plurality of people, the statistical values (maximum value, minimum value, average value, mode value, median value, etc.) of the evaluation values of each of the plurality of people may be displayed.
　次に、第２のクエリ取得部１１１は、第１のクエリで指定された姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する。第１のクエリ取得部１０９が３人以上の人物の姿勢を指定する第１のクエリを取得している場合、第２のクエリ取得部１１１は、その中の２者間における物理的距離及び時間的距離の少なくとも一方を指定した第２のクエリを取得する。なお、必ずしもすべてのペアに対応して物理的距離や時間的距離を指定する必要はなく、一部のペアに対してのみ物理的距離や時間的距離が指定できてもよい。 Next, the second query acquisition unit 111 acquires the second query in which at least one of the physical distance and the temporal distance between the plurality of persons taking the posture specified in the first query is specified. .. When the first query acquisition unit 109 acquires the first query that specifies the postures of three or more people, the second query acquisition unit 111 is the physical distance and time between the two persons. Gets a second query that specifies at least one of the target distances. It is not always necessary to specify the physical distance and the time distance corresponding to all the pairs, and the physical distance and the time distance may be specified only for some pairs.
　「物理的距離」は、第１のクエリで指定された姿勢をとる第１の人物と、第１のクエリで指定された姿勢をとる第２の人物との間の物理的な距離である。第２のクエリにおいて、物理的距離は、「ｍ（メートル）」等の単位を用いて実際の距離で示されてもよいし、画素数で各画像内における距離で示されてもよい。 The "physical distance" is the physical distance between the first person who takes the posture specified by the first query and the second person who takes the posture specified by the first query. In the second query, the physical distance may be expressed as an actual distance using a unit such as "m (meter)" or as a distance in each image in terms of the number of pixels.
　人物間の物理的距離を人物の身体のどの箇所を基準にして算出するかは設計的事項である。例えば、頭部間の物理的距離や足元間の物理的距離等としてもよい。その他、第１の人物の身体が存在する画像内の領域と、第２の人物の身体が存在する画像内の領域との間の最短の物理的距離を、その２者間の物理的距離としてもよい。なお、ここでの例示はあくまで一例であり、その他の手法で人物間の物理的距離を算出してもよい。 It is a design matter which part of the person's body is used as the basis for calculating the physical distance between people. For example, the physical distance between the heads, the physical distance between the feet, and the like may be used. In addition, the shortest physical distance between the region in the image in which the body of the first person exists and the region in the image in which the body of the second person exists is defined as the physical distance between the two. May be good. The example here is just an example, and the physical distance between people may be calculated by another method.
　第２のクエリにおいて、第１のクエリで指定された姿勢をとる複数の人物間の物理的距離は、例えば「Ｘ１以上」、「Ｘ２以下」、「Ｘ３以上Ｘ４以下」等のように指定される。 In the second query, the physical distance between a plurality of persons taking the posture specified in the first query is specified as, for example, "X1 or more", "X2 or less", "X3 or more and X4 or less", and the like. To.
　「時間的距離」は、第１のクエリで指定された姿勢をとる第１の人物が現れるタイミングと、第１のクエリで指定された姿勢をとる第２の人物が現れるタイミングとの間の時間差である。第２のクエリにおいて、時間的距離は、「秒」等の単位を用いて示されてもよいし、フレーム数で示されてもよい。 The "time distance" is the time difference between the timing when the first person who takes the posture specified by the first query appears and the timing when the second person who takes the posture specified by the first query appears. Is. In the second query, the temporal distance may be expressed in units such as "seconds" or in the number of frames.
　第２のクエリにおいて、第１のクエリで指定された姿勢をとる複数の人物間の時間的距離は、例えば「Ｙ１以上」、「Ｙ２以下」、「Ｙ３以上Ｙ４以下」等のように指定される。その他、当該時間的距離は「０（ゼロ）」を指定できてもよい。時間的距離として「０（ゼロ）」が指定された場合、第１のクエリで指定された姿勢をとる複数の人物が同時に同じ画像内に現れるという条件が指定される。 In the second query, the temporal distance between a plurality of persons taking the posture specified in the first query is specified as, for example, "Y1 or more", "Y2 or less", "Y3 or more and Y4 or less", and the like. To. In addition, "0 (zero)" may be specified for the time distance. When "0 (zero)" is specified as the time distance, a condition is specified that a plurality of people taking the posture specified in the first query appear in the same image at the same time.
　ここで、第２のクエリの取得方法の例を説明する。 Here, an example of the acquisition method of the second query will be described.
「第２のクエリ取得例１」
　当該例では、第２のクエリ取得部１１１は、予め用意された複数の選択肢（物理的距離及び時間的距離）を、ユーザに向けて表示する。そして、第２のクエリ取得部１１１は、当該選択肢の中からユーザが選択した物理的距離や時間的距離を、第２のクエリとして取得する。 "Second query acquisition example 1"
In this example, the second query acquisition unit 111 displays a plurality of options (physical distance and temporal distance) prepared in advance toward the user. Then, the second query acquisition unit 111 acquires the physical distance or the temporal distance selected by the user from the options as the second query.
「第２のクエリ取得例２」
　当該例では、第２のクエリ取得部１１１は、数値で、物理的距離や時間的距離を指定するユーザ入力を受付ける。そして、第２のクエリ取得部１１１は、ユーザが入力した数値を第２のクエリとして取得する。 "Second query acquisition example 2"
In this example, the second query acquisition unit 111 accepts a user input that specifies a physical distance or a temporal distance numerically. Then, the second query acquisition unit 111 acquires the numerical value input by the user as the second query.
「第２のクエリ取得例３」
　当該例では、第２のクエリ取得部１１１は、画像取得部１０１が取得したクエリ画像に含まれる２者間の物理的距離、及び、時間的距離の少なくとも一方を算出し、算出した値及びその周辺を含む数値範囲を、第２のクエリとして取得する。以下、一例を説明する。 "Second query acquisition example 3"
In this example, the second query acquisition unit 111 calculates at least one of the physical distance and the temporal distance between the two parties included in the query image acquired by the image acquisition unit 101, and the calculated value and its value. The numerical range including the periphery is acquired as the second query. An example will be described below.
　例えば、ユーザ入力により複数の人物を含む１枚のクエリ画像の中から２者が選択され、選択された２人の人物各々の姿勢（２次元骨格構造の特徴量）が第１のクエリとして取得されたとする。この場合、第２のクエリ取得部１１１は、この選択された２者間の物理的距離Ｐを算出し、算出した値Ｐ及びその周辺を含む数値範囲（Ｐ－α以上、Ｐ＋α以下）を、第２のクエリにおける２者間の物理的距離として決定する。 For example, two people are selected from one query image including a plurality of people by user input, and the postures (features of the two-dimensional skeleton structure) of each of the selected two people are acquired as the first query. Suppose it was done. In this case, the second query acquisition unit 111 calculates the physical distance P between the selected two parties, and sets a numerical range (P-α or more, P + α or less) including the calculated value P and its surroundings. Determined as the physical distance between the two in the second query.
　その他の例として、例えば、画像取得部１０１は、クエリ画像として、動画像を取得してもよい。そして、ユーザ入力により動画像の中の互いに異なる２つのフレーム画像各々から人物が選択され、選択された２人の人物各々の姿勢（２次元骨格構造の特徴量）が第１のクエリとして取得されたとする。 As another example, for example, the image acquisition unit 101 may acquire a moving image as a query image. Then, a person is selected from each of the two different frame images in the moving image by user input, and the posture (feature amount of the two-dimensional skeleton structure) of each of the selected two people is acquired as the first query. Suppose.
　この場合、第２のクエリ取得部１１１は、上記２人の人物が選択された２つのフレーム画像間の時間的距離Ｑを算出し、算出した値Ｑ及びその周辺を含む数値範囲（Ｑ－β以上、Ｑ＋β以下）を、第２のクエリにおける２者間の時間的距離として決定する。 In this case, the second query acquisition unit 111 calculates the temporal distance Q between the two frame images selected by the two persons, and the numerical range (Q-β) including the calculated value Q and its surroundings. As described above, Q + β or less) is determined as the time distance between the two parties in the second query.
　また、第２のクエリ取得部１１１は、上記２人の人物が選択された２つのフレーム画像、及び、これらに挟まれるフレーム画像各々における当該２人の人物間の物理的距離を算出する。そして第２のクエリ取得部１１１は、算出した値の統計値Ｒ（最大値、最小値、平均値、最頻値、中央値等）を算出し、算出した値Ｒ及びその周辺を含む数値範囲（Ｒ－γ以上、Ｒ＋γ以下）を、第２のクエリにおける２者間の物理的 Further, the second query acquisition unit 111 calculates the physical distance between the two frame images selected by the two persons and the frame images sandwiched between the two persons. Then, the second query acquisition unit 111 calculates the statistical value R (maximum value, minimum value, mean value, mode value, median value, etc.) of the calculated value, and a numerical range including the calculated value R and its surroundings. (R-γ or more, R + γ or less) is the physical between the two in the second query.
　検索部１０５は、第１のクエリ及び第２のクエリに基づく画像検索を実行する。例えば、検索部１０５は、複数の解析対象画像（フレーム画像）で構成される動画像の中から、以下の条件のすべてを満たすシーンを検索することができる。 The search unit 105 executes an image search based on the first query and the second query. For example, the search unit 105 can search for a scene that satisfies all of the following conditions from a moving image composed of a plurality of analysis target images (frame images).
・第１のクエリで指定された第１の姿勢をとる第１の人物と、第１のクエリで指定された第２の姿勢をとる第２の人物が現れる。
・第１の人物が第１の姿勢をとる第１のタイミングと、第２の人物が第２の姿勢をとる第２のタイミングの時間差は、第２のクエリで指定された時間的距離の条件を満たす。
・第１のタイミングと第２のタイミングとの間の時間帯の少なくとも一部において、第１の人物と第２の人物との間の物理的距離が、第２のクエリで指定された物理的距離の条件を満たす。 -A first person who takes the first posture specified by the first query and a second person who takes the second posture specified by the first query appear.
-The time difference between the first timing in which the first person takes the first posture and the second timing in which the second person takes the second posture is a condition of the time distance specified by the second query. Meet.
The physical distance between the first person and the second person in at least part of the time zone between the first timing and the second timing is the physical distance specified in the second query. Meet the distance conditions.
　図４７は、検索されたシーンの一例を示す。この例の場合、第１のクエリにおいて、解析対象画像Ｆ１に含まれる第１の人物Ａの姿勢が第１の姿勢として指定され、解析対象画像Ｆ２に含まれる第２の人物Ｂの姿勢が第２の姿勢として指定されていた。そして、これら２つのフレーム画像（解析対象画像Ｆ１及びＦ２）間の時間差Ｔは、第２のクエリで指定された時間的距離の条件を満たす。また、これら２つのフレーム画像（解析対象画像Ｆ１及びＦ２）間の時間帯の少なくとも一部において、第１の人物Ａと第２の人物Ｂとの間の物理的距離Ｌは、第２のクエリで指定された物理的距離の条件を満たす。なお、周知の技術（例：物体追跡技術、外観の特徴量を利用した人物認識等）を利用して、複数のフレーム画像にまたがる同一人物を特定することができる。 FIG. 47 shows an example of the searched scene. In the case of this example, in the first query, the posture of the first person A included in the analysis target image F1 is designated as the first posture, and the posture of the second person B included in the analysis target image F2 is the first. It was designated as the posture of 2. Then, the time difference T between these two frame images (analysis target images F1 and F2) satisfies the condition of the time distance specified by the second query. Further, in at least a part of the time zone between these two frame images (analysis target images F1 and F2), the physical distance L between the first person A and the second person B is the second query. Satisfy the conditions of the physical distance specified in. It should be noted that a well-known technique (eg, object tracking technique, person recognition using appearance features, etc.) can be used to identify the same person across a plurality of frame images.
　その他、検索部１０５は、複数の解析対象画像（静止画像）の中から、以下の条件のすべてを満たす静止画像を検索することができる。 In addition, the search unit 105 can search for still images that satisfy all of the following conditions from among a plurality of analysis target images (still images).
・第１のクエリで指定された第１の姿勢をとる第１の人物と、第１のクエリで指定された第２の姿勢をとる第２の人物が現れる。
・第１の人物と第２の人物との間の物理的距離が、第２のクエリで指定された物理的距離の条件を満たす。 -A first person who takes the first posture specified by the first query and a second person who takes the second posture specified by the first query appear.
-The physical distance between the first person and the second person satisfies the condition of the physical distance specified by the second query.
　図４６は、検索されたシーンの一例を示す。この例の場合、第１のクエリにおいて、解析対象画像Ｆ３に含まれる第１の人物Ａの姿勢が第１の姿勢として指定され、解析対象画像Ｆ３に含まれる第２の人物Ｂの姿勢が第２の姿勢として指定されていた。そして、解析対象画像Ｆ３から算出される第１の人物Ａと第２の人物Ｂとの間の物理的距離Ｌは、第２のクエリで指定された物理的距離の条件を満たす。 FIG. 46 shows an example of the searched scene. In the case of this example, in the first query, the posture of the first person A included in the analysis target image F3 is designated as the first posture, and the posture of the second person B included in the analysis target image F3 is the first. It was designated as the posture of 2. Then, the physical distance L between the first person A and the second person B calculated from the analysis target image F3 satisfies the condition of the physical distance specified by the second query.
　検索部１０５は、例えば、データ蓄積処理でデータベース１１０に格納された複数の骨格構造の中から、第１のクエリ（クエリ状態）で指定された２次元骨格構造の特徴量と類似度の高い骨格構造を検索することで、第１のクエリで指定された姿勢をとる人物を検索することができる。 The search unit 105 has, for example, a skeleton having a high degree of similarity to the feature amount of the two-dimensional skeleton structure specified in the first query (query state) from among a plurality of skeleton structures stored in the database 110 in the data storage process. By searching the structure, it is possible to search for a person who takes the posture specified by the first query.
　例えば、検索部１０５は、検索クエリの特徴量と、複数の解析対象画像各々から検出された骨格構造の特徴量とを照合することで、検索クエリの特徴量と類似度の高い骨格構造を検索してもよい。この構成の場合、上述した分類処理は不要となる。しかし、照合対象が複数の解析対象画像の全てとなるので、照合におけるコンピュータの処理負担が大きくなる。 For example, the search unit 105 searches for a skeleton structure having a high degree of similarity to the feature amount of the search query by collating the feature amount of the search query with the feature amount of the skeleton structure detected from each of the plurality of analysis target images. You may. In the case of this configuration, the above-mentioned classification process becomes unnecessary. However, since the collation target is all of the plurality of analysis target images, the processing load of the computer in collation becomes large.
　そこで、検索部１０５は、分類処理で得られたグループ毎に２次元の骨格構造の特徴量の代表を任意の手段で決定し、代表と上記検索クエリの特徴量との照合により、検索クエリの特徴量と類似度の高い骨格構造を検索してもよい。この構成の場合、照合対象の数が少なくなるので、照合におけるコンピュータの処理負担が小さくなる。 Therefore, the search unit 105 determines a representative of the feature amount of the two-dimensional skeletal structure for each group obtained by the classification process by an arbitrary means, and collates the representative with the feature amount of the above search query to obtain a search query. You may search for a skeletal structure having a high degree of similarity to the feature amount. In the case of this configuration, since the number of collation targets is reduced, the processing load of the computer in collation is reduced.
　なお、解析対象画像と、各解析対象画像から検出された２次元の骨格構造とは互いに紐付けられている。このため、上記「解析対象画像から検出された複数の２次元の骨格構造の中から所定の骨格構造を検索する処理」により、所定の骨格構造（検索クエリの特徴量と類似度の高い骨格構造）を含む解析対象画像を検索することができる。すなわち、解析対象画像の中から、第１のクエリで指定された姿勢をとる人物を含む解析対象画像を検索することができる。 The image to be analyzed and the two-dimensional skeleton structure detected from each image to be analyzed are associated with each other. Therefore, by the above-mentioned "process of searching for a predetermined skeleton structure from a plurality of two-dimensional skeleton structures detected from the image to be analyzed", the predetermined skeleton structure (the skeleton structure having a high degree of similarity to the feature amount of the search query). ) Can be searched for the image to be analyzed. That is, it is possible to search the analysis target image including the person who takes the posture specified in the first query from the analysis target images.
　類似度は、骨格構造の特徴量間の距離である。検索部１０５は、骨格構造の全体の特徴量の類似度により検索してもよいし、骨格構造の一部の特徴量の類似度により検索してもよく、骨格構造の第１の部分（例えば両手）及び第２の部分（例えば両足）の特徴量の類似度により検索してもよい。なお、各画像における人物の骨格構造の特徴量に基づいて人物の姿勢を検索してもよいし、時系列に連続する複数の画像における人物の骨格構造の特徴量の変化に基づいて人物の行動を検索してもよい。すなわち、検索部１０５は、骨格構造の特徴量に基づいて人物の姿勢や行動を含む人物の状態を検索できる。例えば、検索部１０５は、所定の監視期間に撮像された複数の解析対象画像における複数の骨格構造の特徴量を検索対象とする。 The degree of similarity is the distance between the features of the skeletal structure. The search unit 105 may search by the similarity of the features of the whole skeleton structure, or may search by the similarity of the features of a part of the skeleton structure, and may search by the similarity of the first part of the skeleton structure (for example,). You may search by the similarity of the features of both hands) and the second part (for example, both feet). The posture of the person may be searched based on the feature amount of the skeletal structure of the person in each image, or the behavior of the person may be searched based on the change of the feature amount of the skeletal structure of the person in a plurality of consecutive images in time series. You may search for. That is, the search unit 105 can search the state of the person including the posture and behavior of the person based on the feature amount of the skeletal structure. For example, the search unit 105 searches for feature quantities of a plurality of skeletal structures in a plurality of analysis target images captured during a predetermined monitoring period.
　ところで、第２のクエリにおいて、物理的距離が「ｍ（メートル）」等の単位を用いて実際の距離で示されている場合、上記検索を実現するためには、画像内における２者間の実際の物理的距離を推定する必要がある。以下、当該推定方法の一例を説明するが、これに限定されない。 By the way, in the second query, when the physical distance is indicated by the actual distance using a unit such as "m (meter)", in order to realize the above search, the two parties in the image are used. It is necessary to estimate the actual physical distance. Hereinafter, an example of the estimation method will be described, but the estimation method is not limited thereto.
　まず、検索部１０５は、基準となる人、又はその人の周囲に位置する人の画像内における高さｔを算出する。ここでは、例えば画素数で表される。次いで検索部１０５は、基準となる人から周囲に位置する人までの画像内における物理的距離ｄを算出する。ここでｄは、ｔと同じ単位（例えば画素数）で表される。次いで、検索部１０５は、ｄ／ｔを算出し、この値に予め設定された基準身長を乗ずることにより、基準となる人とその周囲に位置する人の物理的距離を算出する。検索部１０５は、画像処理によって高さｔの算出対象となった人の属性（例えば性別及び年齢層の少なくとも一方）が推定できた場合、この属性によって基準身長を変更してもよい。 First, the search unit 105 calculates the height t in the image of the reference person or a person located around the person. Here, for example, it is represented by the number of pixels. Next, the search unit 105 calculates the physical distance d in the image from the reference person to the person located in the surroundings. Here, d is expressed in the same unit as t (for example, the number of pixels). Next, the search unit 105 calculates d / t and multiplies this value by a preset reference height to calculate the physical distance between the reference person and the person located around the reference person. When the search unit 105 can estimate the attribute (for example, at least one of the gender and the age group) of the person whose height t is calculated by image processing, the search unit 105 may change the reference height according to this attribute.
　なお、ほとんどの画像には、その画像を生成したカメラ２００に固有の歪が生じている。検索部１０５は、上記物理的距離を算出する際、この歪を補正する処理を行うのが好ましい。検索部１０５は、画像内における人の位置に応じたひずみ補正処理を行うことができる。一般的に、画像の歪は、例えばカメラ２００が有する光学系（例えばレンズ）、及びそのカメラ２００の上下方向の向き（例えば水平面に対する角度）に起因している。そこで画像内における人の位置に応じたひずみ補正処理の内容は、カメラ２００が有する光学系（例えばレンズ）、及びカメラ２００の上下方向の向きに応じて設定される。 Most of the images have distortion peculiar to the camera 200 that generated the image. When calculating the physical distance, the search unit 105 preferably performs a process of correcting this distortion. The search unit 105 can perform distortion correction processing according to the position of a person in the image. In general, image distortion is due, for example, to the optical system of the camera 200 (eg, a lens) and the vertical orientation of the camera 200 (eg, an angle with respect to a horizontal plane). Therefore, the content of the strain correction process according to the position of the person in the image is set according to the optical system (for example, the lens) of the camera 200 and the vertical orientation of the camera 200.
　入力部１０６は、画像処理装置１００を操作するユーザから入力された情報を取得する入力インターフェイスである。例えば、ユーザは、監視カメラの画像から不審な状態の人物を監視する監視者である。入力部１０６は、例えば、ＧＵＩ（Graphical User Interface）であり、キーボード、マウス、タッチパネル、マイク、物理ボタン等の入力装置から、ユーザの操作に応じた情報が入力される。 The input unit 106 is an input interface for acquiring information input from a user who operates the image processing device 100. For example, the user is a monitor who monitors a person in a suspicious state from an image of a surveillance camera. The input unit 106 is, for example, a GUI (Graphical User Interface), and information according to a user's operation is input from an input device such as a keyboard, a mouse, a touch panel, a microphone, and a physical button.
　表示部１０７は、画像処理装置１００の動作（処理）の結果等を表示する表示部であり、例えば、液晶ディスプレイや有機ＥＬ（Electro Luminescence）ディスプレイ等のディスプレイ装置である。表示部１０７は、分類部１０４の分類結果、検索部１０５の検索結果等を表示する。 The display unit 107 is a display unit that displays the result of the operation (processing) of the image processing device 100, and is, for example, a display device such as a liquid crystal display or an organic EL (ElectroLuminescence) display. The display unit 107 displays the classification result of the classification unit 104, the search result of the search unit 105, and the like.
　次に、画像処理装置１００のハードウエア構成の一例を説明する。画像処理装置１００の各機能部は、任意のコンピュータのＣＰＵ（Central Processing Unit）、メモリ、メモリにロードされるプログラム、そのプログラムを格納するハードディスク等の記憶ユニット（あらかじめ装置を出荷する段階から格納されているプログラムのほか、ＣＤ（Compact Disc）等の記憶媒体やインターネット上のサーバ等からダウンロードされたプログラムをも格納できる）、ネットワーク接続用インターフェイスを中心にハードウエアとソフトウエアの任意の組合せによって実現される。そして、その実現方法、装置にはいろいろな変形例があることは、当業者には理解されるところである。 Next, an example of the hardware configuration of the image processing device 100 will be described. Each functional unit of the image processing device 100 is stored in a storage unit (from the stage of shipping the device in advance) such as a CPU (Central Processing Unit) of an arbitrary computer, a memory, a program loaded in the memory, and a hard disk for storing the program. It can store programs downloaded from storage media such as CDs (Compact Discs) and servers on the Internet), and can be realized by any combination of hardware and software centered on the network connection interface. Will be done. And, it is understood by those skilled in the art that there are various variations in the method of realizing the device and the device.
　図４１は、画像処理装置１００のハードウエア構成を例示するブロック図である。図４１に示すように、画像処理装置１００は、プロセッサ１Ａ、メモリ２Ａ、入出力インターフェイス３Ａ、周辺回路４Ａ、バス５Ａを有する。周辺回路４Ａには、様々なモジュールが含まれる。画像処理装置１００は周辺回路４Ａを有さなくてもよい。なお、画像処理装置１００は物理的及び／又は論理的に分かれた複数の装置で構成されてもよいし、物理的及び／又は論理的に一体となった１つの装置で構成されてもよい。画像処理装置１００が物理的及び／又は論理的に分かれた複数の装置で構成される場合、複数の装置各々が上記ハードウエア構成を備えることができる。 FIG. 41 is a block diagram illustrating a hardware configuration of the image processing device 100. As shown in FIG. 41, the image processing device 100 includes a processor 1A, a memory 2A, an input / output interface 3A, a peripheral circuit 4A, and a bus 5A. The peripheral circuit 4A includes various modules. The image processing device 100 does not have to have the peripheral circuit 4A. The image processing device 100 may be composed of a plurality of physically and / or logically separated devices, or may be composed of one physically and / or logically integrated device. When the image processing device 100 is composed of a plurality of physically and / or logically separated devices, each of the plurality of devices can be provided with the above hardware configuration.
　バス５Ａは、プロセッサ１Ａ、メモリ２Ａ、周辺回路４Ａ及び入出力インターフェイス３Ａが相互にデータを送受信するためのデータ伝送路である。プロセッサ１Ａは、例えばＣＰＵ、ＧＰＵ（Graphics Processing Unit）などの演算処理装置である。メモリ２Ａは、例えばＲＡＭ（Random Access Memory）やＲＯＭ（Read Only Memory）などのメモリである。入出力インターフェイス３Ａは、入力装置、外部装置、外部サーバ、外部センサー、カメラ等から情報を取得するためのインターフェイスや、出力装置、外部装置、外部サーバ等に情報を出力するためのインターフェイスなどを含む。入力装置は、例えばキーボード、マウス、マイク、物理ボタン、タッチパネル等である。出力装置は、例えばディスプレイ、スピーカ、プリンター、メーラ等である。プロセッサ１Ａは、各モジュールに指令を出し、それらの演算結果をもとに演算を行うことができる。 The bus 5A is a data transmission path for the processor 1A, the memory 2A, the peripheral circuit 4A, and the input / output interface 3A to transmit and receive data to each other. The processor 1A is, for example, an arithmetic processing unit such as a CPU or a GPU (Graphics Processing Unit). The memory 2A is, for example, a memory such as a RAM (RandomAccessMemory) or a ROM (ReadOnlyMemory). The input / output interface 3A includes an interface for acquiring information from an input device, an external device, an external server, an external sensor, a camera, etc., an interface for outputting information to an output device, an external device, an external server, etc. .. The input device is, for example, a keyboard, a mouse, a microphone, a physical button, a touch panel, or the like. The output device is, for example, a display, a speaker, a printer, a mailer, or the like. The processor 1A can issue a command to each module and perform a calculation based on the calculation result thereof.
　図３～図５は、本実施の形態に係る画像処理装置１００の処理の流れを示している。図３は、画像処理装置１００における画像取得から検索処理までの流れを示し、図４は、図３の分類処理（Ｓ１０４）の流れを示し、図５は、図３の検索処理（Ｓ１０５）の流れを示している。 3 to 5 show the processing flow of the image processing apparatus 100 according to the present embodiment. FIG. 3 shows a flow from image acquisition to search processing in the image processing apparatus 100, FIG. 4 shows a flow of classification processing (S104) in FIG. 3, and FIG. 5 shows a flow of search processing (S105) in FIG. It shows the flow.
　図３に示すように、画像取得部１０１は、複数の解析対象画像を取得する（Ｓ１０１）。続いて、骨格構造検出部１０２は、取得された複数の解析対象画像各々から人物の２次元骨格構造を検出する（Ｓ１０２）。図６は、骨格構造の検出例を示している。図６に示すように、解析対象画像には複数の人物が含まれている場合がある。この場合、骨格構造検出部１０２は、解析対象画像に含まれる各人物について骨格構造を検出する。 As shown in FIG. 3, the image acquisition unit 101 acquires a plurality of analysis target images (S101). Subsequently, the skeleton structure detection unit 102 detects the two-dimensional skeleton structure of the person from each of the acquired plurality of analysis target images (S102). FIG. 6 shows an example of detecting a skeletal structure. As shown in FIG. 6, the analysis target image may include a plurality of people. In this case, the skeleton structure detection unit 102 detects the skeleton structure for each person included in the analysis target image.
　図７は、このとき検出する人体モデル３００の骨格構造を示しており、図８～図１０は、骨格構造の検出例を示している。骨格構造検出部１０２は、ＯｐｅｎＰｏｓｅ等の骨格推定技術を用いて、２次元の画像から図７のような人体モデル（２次元骨格モデル）３００の骨格構造を検出する。人体モデル３００は、人物の関節等のキーポイントと、各キーポイントを結ぶボーンから構成された２次元モデルである。 FIG. 7 shows the skeleton structure of the human body model 300 detected at this time, and FIGS. 8 to 10 show an example of detecting the skeleton structure. The skeleton structure detection unit 102 detects the skeleton structure of the human body model (two-dimensional skeleton model) 300 as shown in FIG. 7 from the two-dimensional image by using a skeleton estimation technique such as OpenPose. The human body model 300 is a two-dimensional model composed of key points such as joints of a person and bones connecting the key points.
　骨格構造検出部１０２は、例えば、画像の中からキーポイントとなり得る特徴点を抽出し、キーポイントの画像を機械学習した情報を参照して、人物の各キーポイントを検出する。図７の例では、人物のキーポイントとして、頭Ａ１、首Ａ２、右肩Ａ３１、左肩Ａ３２、右肘Ａ４１、左肘Ａ４２、右手Ａ５１、左手Ａ５２、右腰Ａ６１、左腰Ａ６２、右膝Ａ７１、左膝Ａ７２、右足Ａ８１、左足Ａ８２を検出する。さらに、これらのキーポイントを連結した人物の骨として、頭Ａ１と首Ａ２を結ぶボーンＢ１、首Ａ２と右肩Ａ３１及び左肩Ａ３２をそれぞれ結ぶボーンＢ２１及びボーンＢ２２、右肩Ａ３１及び左肩Ａ３２と右肘Ａ４１及び左肘Ａ４２をそれぞれ結ぶボーンＢ３１及びボーンＢ３２、右肘Ａ４１及び左肘Ａ４２と右手Ａ５１及び左手Ａ５２をそれぞれ結ぶボーンＢ４１及びボーンＢ４２、首Ａ２と右腰Ａ６１及び左腰Ａ６２をそれぞれ結ぶボーンＢ５１及びボーンＢ５２、右腰Ａ６１及び左腰Ａ６２と右膝Ａ７１及び左膝Ａ７２をそれぞれ結ぶボーンＢ６１及びボーンＢ６２、右膝Ａ７１及び左膝Ａ７２と右足Ａ８１及び左足Ａ８２をそれぞれ結ぶボーンＢ７１及びボーンＢ７２を検出する。骨格構造検出部１０２は、検出した人物の骨格構造をデータベース１１０に格納する。 The skeleton structure detection unit 102, for example, extracts feature points that can be key points from an image, and detects each key point of a person by referring to information obtained by machine learning the image of the key points. In the example of FIG. 7, the key points of the person are head A1, neck A2, right shoulder A31, left shoulder A32, right elbow A41, left elbow A42, right hand A51, left hand A52, right hip A61, left hip A62, and right knee A71. , Left knee A72, right foot A81, left foot A82 are detected. Further, as the bones of the person connecting these key points, the bone B1 connecting the head A1 and the neck A2, the bones B21 and the bone B22 connecting the neck A2 and the right shoulder A31 and the left shoulder A32, respectively, and the right shoulder A31 and the left shoulder A32 and the right. Bone B31 and B32 connecting the elbow A41 and the left elbow A42, respectively, connecting the right elbow A41 and the left elbow A42 to the right hand A51 and the left hand A52, respectively, and connecting the neck A2 to the right waist A61 and the left waist A62, respectively. Bones B51 and B52, right waist A61 and left waist A62, right knee A71 and left knee A72, respectively, bone B61 and bone B62, right knee A71 and left knee A72, right foot A81 and left foot A82, respectively. B72 is detected. The skeleton structure detection unit 102 stores the detected skeleton structure of the person in the database 110.
　図８は、起立した状態の人物を検出する例である。図８では、起立した人物が正面から撮像されており、正面から見たボーンＢ１、ボーンＢ５１及びボーンＢ５２、ボーンＢ６１及びボーンＢ６２、ボーンＢ７１及びボーンＢ７２がそれぞれ重ならずに検出され、右足のボーンＢ６１及びボーンＢ７１は左足のボーンＢ６２及びボーンＢ７２よりも多少折れ曲がっている。 FIG. 8 is an example of detecting a person in an upright position. In FIG. 8, a standing person is imaged from the front, and bones B1, bone B51 and bone B52, bones B61 and bone B62, bones B71 and bone B72 viewed from the front are detected without overlapping, and the right foot is detected. Bone B61 and Bone B71 are slightly bent more than Bone B62 and Bone B72 of the left foot.
　図９は、しゃがみ込んでいる状態の人物を検出する例である。図９では、しゃがみ込んでいる人物が右側から撮像されており、右側から見たボーンＢ１、ボーンＢ５１及びボーンＢ５２、ボーンＢ６１及びボーンＢ６２、ボーンＢ７１及びボーンＢ７２がそれぞれ検出され、右足のボーンＢ６１及びボーンＢ７１と左足のボーンＢ６２及びボーンＢ７２は大きく折れ曲がり、かつ、重なっている。 FIG. 9 is an example of detecting a person in a crouched state. In FIG. 9, a crouching person is imaged from the right side, bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72, respectively, viewed from the right side, and bone B61 on the right foot. And the bone B71 and the bone B62 and the bone B72 of the left foot are greatly bent and overlapped.
　図１０は、寝ている状態の人物を検出する例である。図１０では、寝ている人物が左斜め前から撮像されており、左斜め前から見たボーンＢ１、ボーンＢ５１及びボーンＢ５２、ボーンＢ６１及びボーンＢ６２、ボーンＢ７１及びボーンＢ７２がそれぞれ検出され、右足のボーンＢ６１及びボーンＢ７１と左足のボーンＢ６２及びボーンＢ７２は折れ曲がり、かつ、重なっている。 FIG. 10 is an example of detecting a sleeping person. In FIG. 10, a sleeping person is imaged from diagonally left front, and bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 viewed from diagonally left front are detected, respectively, and the right foot. The bones B61 and B71 of the left foot and the bones B62 and B72 of the left foot are bent and overlapped.
　続いて、図３に示すように、特徴量算出部１０３は、検出された骨格構造の特徴量を算出する（Ｓ１０３）。例えば、骨格領域の高さや面積を特徴量とする場合、特徴量算出部１０３は、骨格構造を含む領域を抽出し、その領域の高さ（画素数）や面積（画素面積）を求める。骨格領域の高さや面積は、抽出される骨格領域の端部の座標や端部のキーポイントの座標から求められる。特徴量算出部１０３は、求めた骨格構造の特徴量をデータベース１１０に格納する。なお、この骨格構造の特徴量は、人物の状態を示す情報としても用いられる。 Subsequently, as shown in FIG. 3, the feature amount calculation unit 103 calculates the feature amount of the detected skeletal structure (S103). For example, when the height or area of the skeleton region is used as the feature amount, the feature amount calculation unit 103 extracts a region including the skeleton structure and obtains the height (number of pixels) or area (pixel area) of the region. The height and area of the skeletal region can be obtained from the coordinates of the end of the extracted skeleton region and the coordinates of the key points at the ends. The feature amount calculation unit 103 stores the obtained feature amount of the skeletal structure in the database 110. The feature amount of this skeletal structure is also used as information indicating the state of a person.
　図８の例では、起立した人物の骨格構造から全てのボーンを含む骨格領域を抽出する。この場合、骨格領域の上端は頭部のキーポイントＡ１、骨格領域の下端は左足のキーポイントＡ８２、骨格領域の左端は右肘のキーポイントＡ４１、骨格領域の右端は左手のキーポイントＡ５２となる。このため、キーポイントＡ１とキーポイントＡ８２のＹ座標の差分から骨格領域の高さを求める。また、キーポイントＡ４１とキーポイントＡ５２のＸ座標の差分から骨格領域の幅を求め、骨格領域の高さと幅から面積を求める。 In the example of FIG. 8, the skeletal region including all the bones is extracted from the skeletal structure of the standing person. In this case, the upper end of the skeletal region is the key point A1 of the head, the lower end of the skeletal region is the key point A82 of the left foot, the left end of the skeletal region is the key point A41 of the right elbow, and the right end of the skeletal region is the key point A52 of the left hand. .. Therefore, the height of the skeleton region is obtained from the difference between the Y coordinates of the key point A1 and the key point A82. Further, the width of the skeleton region is obtained from the difference between the X coordinates of the key point A41 and the key point A52, and the area is obtained from the height and width of the skeleton region.
　図９の例では、しゃがみ込んだ人物の骨格構造から全てのボーンを含む骨格領域を抽出する。この場合、骨格領域の上端は頭部のキーポイントＡ１、骨格領域の下端は右足のキーポイントＡ８１、骨格領域の左端は右腰のキーポイントＡ６１、骨格領域の右端は右手のキーポイントＡ５１となる。このため、キーポイントＡ１とキーポイントＡ８１のＹ座標の差分から骨格領域の高さを求める。また、キーポイントＡ６１とキーポイントＡ５１のＸ座標の差分から骨格領域の幅を求め、骨格領域の高さと幅から面積を求める。 In the example of FIG. 9, the skeletal region including all bones is extracted from the skeletal structure of a crouched person. In this case, the upper end of the skeletal region is the key point A1 of the head, the lower end of the skeletal region is the key point A81 of the right foot, the left end of the skeletal region is the key point A61 of the right hip, and the right end of the skeletal region is the key point A51 of the right hand. .. Therefore, the height of the skeleton region is obtained from the difference between the Y coordinates of the key point A1 and the key point A81. Further, the width of the skeleton region is obtained from the difference between the X coordinates of the key point A61 and the key point A51, and the area is obtained from the height and width of the skeleton region.
　図１０の例では、寝ている人物の骨格構造から全てのボーンを含む骨格領域を抽出する。この場合、骨格領域の上端は左肩のキーポイントＡ３２、骨格領域の下端は左手のキーポイントＡ５２、骨格領域の左端は右手のキーポイントＡ５１、骨格領域の右端は左足のキーポイントＡ８２となる。このため、キーポイントＡ３２とキーポイントＡ５２のＹ座標の差分から骨格領域の高さを求める。また、キーポイントＡ５１とキーポイントＡ８２のＸ座標の差分から骨格領域の幅を求め、骨格領域の高さと幅から面積を求める。 In the example of FIG. 10, a skeletal region including all bones is extracted from the skeletal structure of a sleeping person. In this case, the upper end of the skeleton region is the key point A32 of the left shoulder, the lower end of the skeleton region is the key point A52 of the left hand, the left end of the skeleton region is the key point A51 of the right hand, and the right end of the skeleton region is the key point A82 of the left foot. Therefore, the height of the skeleton region is obtained from the difference between the Y coordinates of the key point A32 and the key point A52. Further, the width of the skeleton region is obtained from the difference between the X coordinates of the key point A51 and the key point A82, and the area is obtained from the height and width of the skeleton region.
　続いて、図３に示すように、分類部１０４は、分類処理を行う（Ｓ１０４）。分類処理では、図４に示すように、分類部１０４は、算出された骨格構造の特徴量の類似度を算出し（Ｓ１１１）、算出された特徴量に基づいて骨格構造を分類する（Ｓ１１２）。分類部１０４は、分類対象であるデータベース１１０に格納されている全ての骨格構造間の特徴量の類似度を求め、最も類似度が高い骨格構造（姿勢）を同じクラスタに分類する（クラスタリングする）。さらに、分類したクラスタ間の類似度を求めて分類し、所定の数のクラスタとなるまで分類を繰り返す。図１１は、骨格構造の特徴量の分類結果のイメージを示している。図１１は、２次元の分類要素によるクラスタ分析のイメージであり、２つ分類要素は、例えば、骨格領域の高さと骨格領域の面積等である。図１１では、分類の結果、複数の骨格構造の特徴量が３つのクラスタＣ１～Ｃ３に分類されている。クラスタＣ１～Ｃ３は、例えば、立っている姿勢、座っている姿勢、寝ている姿勢のように各姿勢に対応し、似ている姿勢ごとに骨格構造（人物）が分類される。 Subsequently, as shown in FIG. 3, the classification unit 104 performs a classification process (S104). In the classification process, as shown in FIG. 4, the classification unit 104 calculates the similarity of the calculated feature amount of the skeletal structure (S111), and classifies the skeletal structure based on the calculated feature amount (S112). .. The classification unit 104 obtains the similarity of the feature quantities between all the skeletal structures stored in the database 110 to be classified, and classifies (clusters) the skeletal structures (postures) having the highest similarity into the same cluster. .. Further, the similarity between the classified clusters is obtained and classified, and the classification is repeated until a predetermined number of clusters are obtained. FIG. 11 shows an image of the classification result of the feature amount of the skeletal structure. FIG. 11 is an image of cluster analysis using a two-dimensional classification element, and the two classification elements are, for example, the height of the skeletal region and the area of the skeletal region. In FIG. 11, as a result of classification, the feature quantities of the plurality of skeletal structures are classified into three clusters C1 to C3. The clusters C1 to C3 correspond to each posture such as a standing posture, a sitting posture, and a sleeping posture, and the skeletal structure (person) is classified for each similar posture.
　本実施の形態では、人物の骨格構造の特徴量に基づいて分類することにより、多様な分類方法を用いることができる。なお、分類方法は、予め設定されていてもよいし、ユーザが任意に設定できるようにしてもよい。また、後述する検索方法と同じ方法により分類を行ってもよい。つまり、検索条件と同様の分類条件により分類してもよい。例えば、分類部１０４は、次の分類方法により分類を行う。いずれかの分類方法を用いてもよいし、任意に選択された分類方法を組み合わせてもよい。 In this embodiment, various classification methods can be used by classifying based on the feature amount of the skeletal structure of a person. The classification method may be set in advance or may be arbitrarily set by the user. Further, the classification may be performed by the same method as the search method described later. That is, it may be classified according to the same classification conditions as the search conditions. For example, the classification unit 104 classifies by the following classification method. Any classification method may be used, or an arbitrarily selected classification method may be combined.
＜分類方法１＞
　複数の階層による分類全身の骨格構造による分類や、上半身や下半身の骨格構造による分類、腕や脚の骨格構造による分類等を階層的に組み合わせて分類する。すなわち、骨格構造の第１の部分や第２の部分の特徴量に基づいて分類し、さらに、第１の部分や第２の部分の特徴量に重みづけを行って分類してもよい。 <Classification method 1>
Classification by multiple layers Classification by skeletal structure of the whole body, classification by skeletal structure of upper body and lower body, classification by skeletal structure of arms and legs, etc. are classified in a hierarchical combination. That is, the classification may be performed based on the feature amounts of the first portion and the second portion of the skeletal structure, and further, the feature amounts of the first portion and the second portion may be weighted and classified.
＜分類方法２＞
　時系列に沿った複数枚の画像による分類時系列に連続する複数の画像における骨格構造の特徴量に基づいて分類する。例えば、時系列方向に特徴量を積み重ねて、累積値に基づいて分類してもよい。さらに、連続する複数の画像における骨格構造の特徴量の変化（変化量）に基づいて分類してもよい。 <Classification method 2>
Classification by a plurality of images along the time series Classification is performed based on the feature amount of the skeletal structure in a plurality of images continuous in the time series. For example, the feature quantities may be stacked in the time series direction and classified based on the cumulative value. Further, it may be classified based on the change (change amount) of the feature amount of the skeletal structure in a plurality of continuous images.
＜分類方法３＞
　骨格構造の左右を無視した分類人物の右側と左側が反対の骨格構造を同じ骨格構造として分類する。 <Classification method 3>
Classification ignoring the left and right of the skeletal structure Classify the skeletal structures whose right and left sides are opposite to each other as the same skeletal structure.
　さらに、分類部１０４は、骨格構造の分類結果を表示する（Ｓ１１３）。分類部１０４は、データベース１１０から必要な骨格構造や人物の画像を取得し、分類結果として似ている姿勢（クラスタ）ごとに骨格構造及び人物を表示部１０７に表示する。図１２は、姿勢を３つに分類した場合の表示例を示している。例えば、図１２に示すように、表示ウィンドウＷ１に、姿勢ごとの姿勢領域ＷＡ１～ＷＡ３を表示し、姿勢領域ＷＡ１～ＷＡ３にそれぞれ該当する姿勢の骨格構造及び人物（イメージ）を表示する。姿勢領域ＷＡ１は、例えば立っている姿勢の表示領域であり、クラスタＣ１に分類された、立っている姿勢に似た骨格構造及び人物を表示する。姿勢領域ＷＡ２は、例えば座っている姿勢の表示領域であり、クラスタＣ２に分類された、座っている姿勢に似た骨格構造及び人物を表示する。姿勢領域ＷＡ３は、例えば寝ている姿勢の表示領域であり、クラスタＣ２に分類された、寝ている姿勢に似た骨格構造及び人物を表示する。 Further, the classification unit 104 displays the classification result of the skeletal structure (S113). The classification unit 104 acquires images of necessary skeleton structures and people from the database 110, and displays the skeleton structure and people on the display unit 107 for each posture (cluster) similar as a classification result. FIG. 12 shows a display example when the postures are classified into three. For example, as shown in FIG. 12, the posture regions WA1 to WA3 for each posture are displayed in the display window W1, and the skeletal structure and the person (image) of the posture corresponding to each of the posture regions WA1 to WA3 are displayed. The posture area WA1 is, for example, a display area of a standing posture, and displays a skeletal structure and a person similar to a standing posture classified into cluster C1. The posture area WA2 is, for example, a sitting posture display area, and displays a skeletal structure and a person similar to the sitting posture classified into cluster C2. The posture area WA3 is, for example, a display area of a sleeping posture, and displays a skeletal structure and a person similar to the sleeping posture classified into cluster C2.
　続いて、図３に示すように、画像処理装置１００は、検索処理を行う（Ｓ１０５）。検索処理では、図５に示すように、画像処理装置１００は、検索条件の入力を受け付ける（Ｓ１２１）。 Subsequently, as shown in FIG. 3, the image processing apparatus 100 performs a search process (S105). In the search process, as shown in FIG. 5, the image processing device 100 accepts the input of the search condition (S121).
　例えば、画像処理装置１００は、図１７に示すようなＵＩ画面を表示し、当該ＵＩ画面から第１のクエリ及び第２のクエリの入力を受付ける。図示するＵＩ画面は、姿勢指定の項目において、第１のクエリの入力を受付ける。当該項目は、複数の人物各々の姿勢を指定可能となっている。そして、各々の姿勢の指定方法として、「選択肢の中から選択」、「クエリ画像を入力」、「骨格モデルを利用」が選択可能になっている。これらの方法は、上述した第１のクエリ取得例１乃至３に対応する。第１のクエリ取得部１０９は、選択された方法で、第１のクエリを取得する。 For example, the image processing device 100 displays a UI screen as shown in FIG. 17, and accepts input of a first query and a second query from the UI screen. The illustrated UI screen accepts the input of the first query in the posture designation item. In this item, the posture of each of a plurality of persons can be specified. Then, as a method of designating each posture, "select from choices", "input query image", and "use skeleton model" can be selected. These methods correspond to the first query acquisition examples 1 to 3 described above. The first query acquisition unit 109 acquires the first query by the selected method.
　また、図示するＵＩ画面は、物理的距離及び時間的距離の項目において、第２のクエリの入力を受付ける。各々の指定方法として、「選択肢の中から選択」、「数値入力」、「クエリ画像を入力」が選択可能になっている。これらの方法は、上述した第２のクエリ取得例１乃至３に対応する。第２のクエリ取得部１１１は、選択された方法で、第２のクエリを取得する。 In addition, the illustrated UI screen accepts the input of the second query in the items of physical distance and temporal distance. As each designation method, "select from choices", "numerical input", and "input query image" can be selected. These methods correspond to the above-mentioned second query acquisition examples 1 to 3. The second query acquisition unit 111 acquires the second query by the selected method.
　その後、検索部１０５は、第１のクエリ及び第２のクエリに基づく画像検索を実行する（Ｓ１２２）。検索部１０５は、第１のクエリで指定される骨格構造を検索クエリとして、検索対象であるデータベース１１０に格納されている骨格構造の中から特徴量の類似度が高い骨格構造を検索し、その検索結果を利用して第２のクエリの条件を満たすシーンを検索する。 After that, the search unit 105 executes an image search based on the first query and the second query (S122). The search unit 105 searches for a skeleton structure having a high degree of similarity in feature quantities from the skeleton structures stored in the database 110 to be searched, using the skeleton structure specified in the first query as a search query. The search result is used to search for a scene that satisfies the condition of the second query.
　なお、検索部１０５は、第１のクエリで指定される骨格構造を検索クエリとした検索において、検索クエリの骨格構造の特徴量と検索対象の骨格構造の特徴量（解析対象画像から検出された骨格構造の特徴量）との類似度を算出し、算出した類似度が所定の閾値よりも高い骨格構造を抽出する。検索クエリの骨格構造の特徴量は、予め算出された特徴量を使用してもよいし、検索時に求めた特徴量を使用してもよい。 In the search using the skeleton structure specified in the first query as the search query, the search unit 105 has a feature amount of the skeleton structure of the search query and a feature amount of the skeleton structure of the search target (detected from the image to be analyzed). The degree of similarity with the feature amount of the skeleton structure) is calculated, and the skeleton structure whose calculated degree of similarity is higher than a predetermined threshold is extracted. As the feature amount of the skeleton structure of the search query, the feature amount calculated in advance may be used, or the feature amount obtained at the time of search may be used.
　本実施の形態では、分類方法と同様に、人物の骨格構造の特徴量に基づいて検索することにより、多様な検索方法を用いることができる。なお、検索方法は、予め設定されていてもよいし、ユーザが任意に設定できるようにしてもよい。例えば、検索部１０５は、次の検索方法により検索を行う。いずれかの検索方法を用いてもよいし、任意に選択された検索方法を組み合わせてもよい。複数の検索方法（検索条件）を論理式（例えばAND（論理積）、OR（論理和）、NOT（否定））により組み合わせて検索してもよい。例えば、検索条件を「（右手を挙げている姿勢）AND（左足を挙げている姿勢）」として検索してもよい。 In the present embodiment, as in the classification method, various search methods can be used by searching based on the feature amount of the skeletal structure of the person. The search method may be set in advance or may be arbitrarily set by the user. For example, the search unit 105 searches by the following search method. Either search method may be used, or an arbitrarily selected search method may be combined. A plurality of search methods (search conditions) may be combined and searched by a logical expression (for example, AND (logical product), OR (logical sum), NOT (negation)). For example, the search condition may be searched as "(posture in which the right hand is raised) AND (posture in which the left foot is raised)".
＜検索方法１＞
　高さ方向の特徴量のみによる検索人物の高さ方向の特徴量のみを用いて検索することで、人物の横方向の変化の影響を抑えることができ、人物の向きや人物の体型の変化に対しロバスト性が向上する。例えば、図１３の骨格構造５０１～５０３のように、人物の向きや体型が異なる場合でも、高さ方向の特徴量は大きく変化しない。このため、骨格構造５０１～５０３では、検索時（分類時）に同じ姿勢であると判断することができる。 <Search method 1>
Search by only the feature amount in the height direction By searching using only the feature amount in the height direction of the person, the influence of the lateral change of the person can be suppressed, and the change of the direction of the person and the body shape of the person can be suppressed. On the other hand, the robustness is improved. For example, as in the skeletal structures 501 to 503 of FIG. 13, even if the orientation and body shape of the person are different, the feature amount in the height direction does not change significantly. Therefore, in the skeletal structures 501 to 503, it can be determined that the postures are the same at the time of searching (at the time of classification).
＜検索方法２＞
　部分検索画像において人物の体の一部が隠れている場合、認識可能な部分の情報のみを用いて検索する。例えば、図１４の骨格構造５１１及び５１２のように、左足が隠れていることにより、左足のキーポイントが検出できない場合でも、検出されている他のキーポイントの特徴量を使用して検索できる。このため、骨格構造５１１及び５１２では、検索時（分類時）に同じ姿勢であると判断することができる。つまり、全てのキーポイントではなく、一部のキーポイントの特徴量を用いて、分類や検索を行うことができる。図１５の骨格構造５２１及び５２２の例では、両足の向きが異なっているものの、上半身のキーポイント（Ａ１、Ａ２、Ａ３１、Ａ３２、Ａ４１、Ａ４２、Ａ５１、Ａ５２）の特徴量を検索クエリとすることで、同じ姿勢であると判断することができる。また、検索したい部分（特徴点）に対して、重みを付けて検索してもよいし、類似度判定の閾値を変化させてもよい。体の一部が隠れている場合、隠れた部分を無視して検索してもよいし、隠れた部分を加味して検索してもよい。隠れた部分も含めて検索することで、同じ部位が隠れているような姿勢を検索することができる。 <Search method 2>
When a part of the body of a person is hidden in the partial search image, the search is performed using only the information of the recognizable part. For example, as in the skeletal structures 511 and 512 of FIG. 14, even if the key point of the left foot cannot be detected due to the hiding of the left foot, the feature amount of other detected key points can be used for the search. Therefore, in the skeletal structures 511 and 512, it can be determined that the postures are the same at the time of searching (at the time of classification). That is, it is possible to perform classification and search using the features of some key points instead of all the key points. In the examples of the skeletal structures 521 and 522 of FIG. 15, although the directions of both feet are different, the feature quantities of the key points of the upper body (A1, A2, A31, A32, A41, A42, A51, A52) are used as the search query. Therefore, it can be determined that the posture is the same. Further, the portion (feature point) to be searched may be weighted and searched, or the threshold value for determining the similarity may be changed. When a part of the body is hidden, the hidden part may be ignored and the search may be performed, or the hidden part may be added to the search. By searching including hidden parts, it is possible to search for postures in which the same part is hidden.
＜検索方法３＞
　骨格構造の左右を無視した検索人物の右側と左側が反対の骨格構造を同じ骨格構造として検索する。例えば、図１６の骨格構造５３１及び５３２のように、右手を挙げている姿勢と、左手を挙げている姿勢を同じ姿勢として検索（分類）できる。図１６の例では、骨格構造５３１と骨格構造５３２は、右手のキーポイントＡ５１、右肘のキーポイントＡ４１、左手のキーポイントＡ５２、左肘のキーポイントＡ４２の位置が異なるものの、その他のキーポイントの位置は同じである。骨格構造５３１の右手のキーポイントＡ５１及び右肘のキーポイントＡ４１と骨格構造５３２の左手のキーポイントＡ５２及び左肘のキーポイントＡ４２のうち、一方の骨格構造のキーポイントを左右反転させると、他方の骨格構造のキーポイントと同じ位置となり、また、骨格構造５３１の左手のキーポイントＡ５２及び左肘のキーポイントＡ４２と骨格構造５３２の右手のキーポイントＡ５１及び右肘のキーポイントＡ４１のうち、一方の骨格構造のキーポイントを左右反転させると、他方の骨格構造のキーポイントと同じ位置となるため、同じ姿勢と判断する。 <Search method 3>
Search ignoring the left and right of the skeletal structure Search for the skeletal structure with the opposite right and left sides of the person as the same skeletal structure. For example, as in the skeletal structures 531 and 532 of FIG. 16, the posture in which the right hand is raised and the posture in which the left hand is raised can be searched (classified) as the same posture. In the example of FIG. 16, the skeletal structure 531 and the skeletal structure 532 have different positions of the right hand key point A51, the right elbow key point A41, the left hand key point A52, and the left elbow key point A42, but other key points. The position of is the same. Of the key point A51 of the right hand and the key point A41 of the right elbow of the skeletal structure 531 and the key point A52 of the left hand and the key point A42 of the left elbow of the skeletal structure 532, when the key point of one of the skeletal structures is flipped left and right, the other It becomes the same position as the key point of the skeletal structure of, and one of the key point A52 of the left hand and the key point A42 of the left elbow of the skeletal structure 531 and the key point A51 of the right hand and the key point A41 of the right elbow of the skeletal structure 532. When the key point of the skeletal structure of is inverted left and right, it becomes the same position as the key point of the other skeletal structure, so it is judged that the posture is the same.
＜検索方法４＞
　縦方向と横方向の特徴量による検索人物の縦方向（Ｙ軸方向）の特徴量のみで検索を行った後、得られた結果をさらに人物の横方向（Ｘ軸方向）の特徴量を用いて検索する。 <Search method 4>
Search by features in the vertical and horizontal directions After searching only with the features in the vertical direction (Y-axis direction) of the person, the obtained results are further obtained using the features in the horizontal direction (X-axis direction) of the person. Search.
＜検索方法５＞
　時系列に沿った複数枚の画像による検索時系列に連続する複数の画像における骨格構造の特徴量に基づいて検索する。例えば、時系列方向に特徴量を積み重ねて、累積値に基づいて検索してもよい。さらに、連続する複数の画像における骨格構造の特徴量の変化（変化量）に基づいて検索してもよい。 <Search method 5>
Search by multiple images in chronological order Search is performed based on the feature quantity of the skeletal structure in a plurality of images consecutive in chronological order. For example, the feature quantities may be stacked in the time series direction and searched based on the cumulative value. Further, the search may be performed based on the change (change amount) of the feature amount of the skeletal structure in a plurality of consecutive images.
　さらに、検索部１０５は、骨格構造の検索結果を表示する（Ｓ１２３）。検索部１０５は、検索結果に含まれる静止画像や、シーンのサムネイル画像等を表示部１０７に表示する。 Further, the search unit 105 displays the search result of the skeletal structure (S123). The search unit 105 displays still images included in the search results, thumbnail images of scenes, and the like on the display unit 107.
　以上のように、本実施の形態では、２次元画像から人物の骨格構造を検出し、検出した骨格構造の特徴量に基づいて分類や検索を行うことを可能とした。これにより、類似度が高い似た姿勢ごとに分類することができ、また、検索クエリ（検索キー）と類似度が高い似た姿勢を検索することができる。画像から似ている姿勢を分類し表示することで、ユーザが姿勢等を指定することなく、画像中の人物の姿勢を把握することができる。分類結果の中からユーザが検索クエリの姿勢を指定できるため、予めユーザが検索したい姿勢を詳細に把握していない場合でも、所望の姿勢を検索することができる。例えば、人物の骨格構造の全体や一部等を条件として分類や検索を行うことができるため、柔軟な分類や検索が可能となる。 As described above, in the present embodiment, it is possible to detect the skeletal structure of a person from a two-dimensional image and perform classification and search based on the feature amount of the detected skeletal structure. As a result, it is possible to classify by similar postures having a high degree of similarity, and it is possible to search for similar postures having a high degree of similarity with a search query (search key). By classifying and displaying similar postures from the image, the posture of the person in the image can be grasped without the user specifying the posture or the like. Since the user can specify the posture of the search query from the classification results, the desired posture can be searched even if the user does not know the posture to be searched in detail in advance. For example, since it is possible to perform classification and search on the condition of the whole or part of the skeleton structure of a person, flexible classification and search is possible.
　また、本実施形態の場合、複数の人物各々の姿勢を指定するとともに、それらの姿勢をとる人物間の物理的距離や、複数の人物各々が指定された姿勢をとるタイミングの時間的距離等を指定して、画像検索を行うことができる。このため、所望のシーンを効率的に検索することが可能となる。 Further, in the case of the present embodiment, the postures of each of the plurality of persons are specified, the physical distance between the persons who take those postures, the time distance of the timing when each of the plurality of persons takes the designated postures, and the like. You can specify and perform an image search. Therefore, it is possible to efficiently search for a desired scene.
　例えば、ある人物が物を投げ、他の人物がその物を受け取る等のように、所定の姿勢が時間差で現れるようなシーンを効率的に検索できる。当該例の場合、「物を投げる姿勢」、「物を受け取る姿勢」、「それら２つの姿勢が現れる時間差」、「それら２つの姿勢をとる２人の物理的距離」等を適切に設定することで、効率的に上記所望のシーンを検索することができる。 For example, it is possible to efficiently search for a scene in which a predetermined posture appears with a time lag, such as one person throwing an object and another person receiving the object. In the case of this example, "throwing posture", "receiving thing", "time difference between these two postures", "physical distance between two people taking those two postures", etc. should be set appropriately. Therefore, the desired scene can be efficiently searched.
　その他、例えば、ある人物が横たわり、他の人物がその人物の上に乗って覆いかぶさる等のように、所定の姿勢が同時に現れるようなシーンも効率的に検索できる。当該例の場合、「横たわる姿勢」、「横たわる人の上に乗ってその人に覆いかぶさる姿勢」、「それら２つの姿勢をとる２人の物理的距離」等を適切に設定することで、効率的に上記所望のシーンを検索することができる。 In addition, it is possible to efficiently search for scenes in which a predetermined posture appears at the same time, for example, when a certain person lies down and another person rides on the person and covers it. In the case of this example, efficiency is achieved by appropriately setting "lying posture", "posture of riding on a lying person and covering the person", "physical distance between two people taking those two postures", and the like. The desired scene can be searched for.
（実施の形態２）
　以下、図面を参照して実施の形態２について説明する。本実施の形態では、実施の形態１における特徴量算出の具体例について説明する。本実施の形態では、人物の身長を用いて正規化することで特徴量を求める。その他については、実施の形態１と同様である。 (Embodiment 2)
Hereinafter, the second embodiment will be described with reference to the drawings. In this embodiment, a specific example of feature amount calculation in the first embodiment will be described. In the present embodiment, the feature amount is obtained by normalizing using the height of the person. Others are the same as those in the first embodiment.
　図１８は、本実施の形態に係る画像処理装置１００の構成を示している。図１８に示すように、画像処理装置１００は、実施の形態１の構成に加えて、さらに身長算出部１０８を備える。なお、特徴量算出部１０３と身長算出部１０８を一つの処理部としてもよい。 FIG. 18 shows the configuration of the image processing apparatus 100 according to the present embodiment. As shown in FIG. 18, the image processing apparatus 100 further includes a height calculation unit 108 in addition to the configuration of the first embodiment. The feature amount calculation unit 103 and the height calculation unit 108 may be combined into one processing unit.
　身長算出部（身長推定部）１０８は、骨格構造検出部１０２により検出された２次元の骨格構造に基づき、２次元の画像内の人物の起立時の高さ（身長画素数という）を算出（推定）する。身長画素数は、２次元の画像における人物の身長（２次元画像空間上の人物の全身の長さ）であるとも言える。身長算出部１０８は、検出された骨格構造の各ボーンの長さ（２次元画像空間上の長さ）から身長画素数（ピクセル数）を求める。 The height calculation unit (height estimation unit) 108 calculates the height (called the number of height pixels) of a person in a two-dimensional image when standing up based on the two-dimensional skeleton structure detected by the skeleton structure detection unit 102 (referred to as the number of height pixels). presume. It can also be said that the number of height pixels is the height of the person in the two-dimensional image (the length of the whole body of the person in the two-dimensional image space). The height calculation unit 108 obtains the number of height pixels (number of pixels) from the length (length on the two-dimensional image space) of each bone of the detected skeleton structure.
　以下の例では、身長画素数を求める方法として具体例１～３を用いる。なお、具体例１～３のいずれかの方法を用いてもよいし、任意に選択される複数の方法を組み合わせて用いてもよい。具体例１では、骨格構造の各ボーンのうち、頭部から足部までのボーンの長さを合計することで、身長画素数を求める。骨格構造検出部１０２（骨格推定技術）が頭頂と足元を出力しない場合は、必要に応じて定数を乗じて補正することもできる。具体例２では、各ボーンの長さと全身の長さ（２次元画像空間上の身長）との関係を示す人体モデルを用いて、身長画素数を算出する。具体例３では、３次元人体モデルを２次元骨格構造にフィッティング（あてはめる）することで、身長画素数を算出する。 In the following examples, specific examples 1 to 3 are used as a method for obtaining the number of height pixels. In addition, any of the methods of Specific Examples 1 to 3 may be used, or a plurality of arbitrarily selected methods may be used in combination. In Specific Example 1, the number of height pixels is obtained by totaling the lengths of the bones from the head to the foot among the bones of the skeletal structure. If the skeleton structure detection unit 102 (skeleton estimation technique) does not output the crown and feet, it can be corrected by multiplying by a constant if necessary. In Specific Example 2, the number of height pixels is calculated using a human body model showing the relationship between the length of each bone and the length of the whole body (height in the two-dimensional image space). In Specific Example 3, the number of height pixels is calculated by fitting (fitting) a three-dimensional human body model to a two-dimensional skeleton structure.
　本実施の形態の特徴量算出部１０３は、算出された人物の身長画素数に基づいて、人物の骨格構造（骨格情報）を正規化する正規化部である。特徴量算出部１０３は、正規化した骨格構造の特徴量（正規化値）をデータベース１１０に格納する。特徴量算出部１０３は、骨格構造に含まれる各キーポイント（特徴点）の画像上での高さを、身長画素数で正規化する。本実施の形態では、例えば、高さ方向は、画像の２次元座標（Ｘ－Ｙ座標）空間における上下の方向（Ｙ軸方向）である。この場合、キーポイントの高さは、キーポイントのＹ座標の値（画素数）から求めることができる。あるいは、高さ方向は、実世界の３次元座標空間における地面（基準面）に対し垂直な鉛直軸の方向を、２次元座標空間に投影した鉛直投影軸の方向（鉛直投影方向）でもよい。この場合、キーポイントの高さは、実世界における地面に対し垂直な軸を、カメラパラメータに基づいて２次元座標空間に投影した鉛直投影軸を求め、この鉛直投影軸に沿った値（画素数）から求めることができる。なお、カメラパラメータは、画像の撮像パラメータであり、例えば、カメラパラメータは、カメラ２００の姿勢、位置、撮像角度、焦点距離等である。カメラ２００により、予め長さや位置が分かっている物体を撮像し、その画像からカメラパラメータを求めることができる。撮像された画像の両端ではひずみが発生し、実世界の鉛直方向と画像の上下方向が合わない場合がある。これに対し、画像を撮影したカメラのパラメータを使用することで、実世界の鉛直方向が画像中でどの程度傾いているのかが分かる。このため、カメラパラメータに基づいて画像中に投影した鉛直投影軸に沿ったキーポイントの値を身長で正規化することで、実世界と画像のずれを考慮してキーポイントを特徴量化することができる。なお、左右方向（横方向）は、画像の２次元座標（Ｘ－Ｙ座標）空間における左右の方向（Ｘ軸方向）であり、または、実世界の３次元座標空間における地面に対し平行な方向を、２次元座標空間に投影した方向である。 The feature amount calculation unit 103 of the present embodiment is a normalization unit that normalizes the skeletal structure (skeleton information) of a person based on the calculated number of height pixels of the person. The feature amount calculation unit 103 stores the feature amount (normalized value) of the normalized skeletal structure in the database 110. The feature amount calculation unit 103 normalizes the height of each key point (feature point) included in the skeleton structure on the image by the number of height pixels. In the present embodiment, for example, the height direction is the vertical direction (Y-axis direction) in the two-dimensional coordinate (XY coordinate) space of the image. In this case, the height of the key point can be obtained from the value (number of pixels) of the Y coordinate of the key point. Alternatively, the height direction may be the direction of the vertical axis perpendicular to the ground (reference plane) in the three-dimensional coordinate space in the real world, and the direction of the vertical projection axis projected onto the two-dimensional coordinate space (vertical projection direction). In this case, the height of the key point is the vertical projection axis obtained by projecting the axis perpendicular to the ground in the real world onto the two-dimensional coordinate space based on the camera parameters, and the value along this vertical projection axis (number of pixels). ) Can be obtained. The camera parameters are image imaging parameters, and for example, the camera parameters are the posture, position, imaging angle, focal length, and the like of the camera 200. The camera 200 can take an image of an object whose length and position are known in advance, and obtain camera parameters from the image. Distortion occurs at both ends of the captured image, and the vertical direction of the real world may not match the vertical direction of the image. On the other hand, by using the parameters of the camera that took the image, you can see how much the vertical direction in the real world is tilted in the image. Therefore, by normalizing the value of the key points along the vertical projection axis projected in the image based on the camera parameters by height, it is possible to feature the key points in consideration of the deviation between the real world and the image. can. The left-right direction (horizontal direction) is the left-right direction (X-axis direction) in the two-dimensional coordinate (XY coordinates) space of the image, or the direction parallel to the ground in the three-dimensional coordinate space in the real world. Is the direction projected onto the two-dimensional coordinate space.
　図１９～図２３は、本実施の形態に係る画像処理装置１００の処理の流れを示している。図１９は、画像処理装置１００における画像取得から検索処理までの流れを示し、図２０～図２２は、図１９の身長画素数算出処理（Ｓ２０１）の具体例１～３の流れを示し、図２３は、図１９の正規化処理（Ｓ２０２）の流れを示している。 19 to 23 show the processing flow of the image processing apparatus 100 according to the present embodiment. 19 shows a flow from image acquisition to search processing in the image processing apparatus 100, and FIGS. 20 to 22 show the flow of specific examples 1 to 3 of the height pixel number calculation process (S201) of FIG. 23 shows the flow of the normalization process (S202) of FIG.
　図１９に示すように、本実施の形態では、実施の形態１における特徴量算出処理（Ｓ１０３）として、身長画素数算出処理（Ｓ２０１）及び正規化処理（Ｓ２０２）を行う。その他については実施の形態１と同様である。 As shown in FIG. 19, in the present embodiment, the height pixel number calculation process (S201) and the normalization process (S202) are performed as the feature amount calculation process (S103) in the first embodiment. Others are the same as those in the first embodiment.
　画像処理装置１００は、画像取得（Ｓ１０１）及び骨格構造検出（Ｓ１０２）に続いて、検出された骨格構造に基づいて身長画素数算出処理を行う（Ｓ２０１）。この例では、図２４に示すように、画像における直立時の人物の骨格構造の高さを身長画素数（ｈ）とし、画像の人物の状態における骨格構造の各キーポイントの高さをキーポイント高さ（ｙｉ）とする。以下、身長画素数算出処理の具体例１～３について説明する。 The image processing apparatus 100 performs height pixel number calculation processing based on the detected skeleton structure (S201) following image acquisition (S101) and skeleton structure detection (S102). In this example, as shown in FIG. 24, the height of the skeleton structure of the person standing upright in the image is the height pixel number (h), and the height of each key point of the skeleton structure in the state of the person in the image is the key point. Let it be the height (yi). Hereinafter, specific examples 1 to 3 of the height pixel number calculation process will be described.
　＜具体例１＞
　具体例１では、頭部から足部までのボーンの長さを用いて身長画素数を求める。具体例１では、図２０に示すように、身長算出部１０８は、各ボーンの長さを取得し（Ｓ２１１）、取得した各ボーンの長さを合計する（Ｓ２１２）。 <Specific example 1>
In Specific Example 1, the number of height pixels is obtained using the length of the bone from the head to the foot. In Specific Example 1, as shown in FIG. 20, the height calculation unit 108 acquires the length of each bone (S211) and totals the lengths of the acquired bones (S212).
　身長算出部１０８は、人物の頭部から足部の２次元の画像上のボーンの長さを取得し、身長画素数を求める。すなわち、骨格構造を検出した画像から、図２４のボーンのうち、ボーンＢ１（長さＬ１）、ボーンＢ５１（長さＬ２１）、ボーンＢ６１（長さＬ３１）及びボーンＢ７１（長さＬ４１）、もしくは、ボーンＢ１（長さＬ１）、ボーンＢ５２（長さＬ２２）、ボーンＢ６２（長さＬ３２）及びボーンＢ７２（長さＬ４２）の各長さ（画素数）を取得する。各ボーンの長さは、２次元の画像における各キーポイントの座標から求めることができる。これらを合計した、Ｌ１＋Ｌ２１＋Ｌ３１＋Ｌ４１、もしくは、Ｌ１＋Ｌ２２＋Ｌ３２＋Ｌ４２に補正定数を乗じた値を身長画素数（ｈ）として算出する。両方の値を算出できる場合、例えば、長い方の値を身長画素数とする。すなわち、各ボーンは正面から撮像された場合が画像中での長さが最も長くなり、カメラに対して奥行き方向に傾くと短く表示される。従って、長いボーンの方が正面から撮像されている可能性が高く、真実の値に近いと考えられる。このため、長い方の値を選択することが好ましい。 The height calculation unit 108 acquires the length of the bone on the two-dimensional image of the foot from the head of the person, and obtains the number of height pixels. That is, from the image in which the skeletal structure is detected, among the bones of FIG. 24, bone B1 (length L1), bone B51 (length L21), bone B61 (length L31) and bone B71 (length L41), or , Bone B1 (length L1), bone B52 (length L22), bone B62 (length L32), and bone B72 (length L42) are acquired. The length of each bone can be obtained from the coordinates of each key point in the two-dimensional image. The sum of these is calculated as the height pixel number (h) by multiplying L1 + L21 + L31 + L41 or L1 + L22 + L32 + L42 by a correction constant. When both values can be calculated, for example, the longer value is taken as the number of height pixels. That is, each bone has the longest length in the image when it is imaged from the front, and it is displayed short when it is tilted in the depth direction with respect to the camera. Therefore, it is more likely that the longer bone is imaged from the front, which is considered to be closer to the true value. Therefore, it is preferable to select the longer value.
　図２５の例では、ボーンＢ１、ボーンＢ５１及びボーンＢ５２、ボーンＢ６１及びボーンＢ６２、ボーンＢ７１及びボーンＢ７２がそれぞれ重ならずに検出されている。これらのボーンの合計である、Ｌ１＋Ｌ２１＋Ｌ３１＋Ｌ４１、及び、Ｌ１＋Ｌ２２＋Ｌ３２＋Ｌ４２を求め、例えば、検出されたボーンの長さが長い左足側のＬ１＋Ｌ２２＋Ｌ３２＋Ｌ４２に補正定数を乗じた値を身長画素数とする。 In the example of FIG. 25, bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 are detected without overlapping. The total of these bones, L1 + L21 + L31 + L41 and L1 + L22 + L32 + L42, is obtained, and for example, the value obtained by multiplying L1 + L22 + L32 + L42 on the left foot side where the detected bone length is long by a correction constant is taken as the height pixel number.
　図２６の例では、ボーンＢ１、ボーンＢ５１及びボーンＢ５２、ボーンＢ６１及びボーンＢ６２、ボーンＢ７１及びボーンＢ７２がそれぞれ検出され、右足のボーンＢ６１及びボーンＢ７１と左足のボーンＢ６２及びボーンＢ７２が重なっている。これらのボーンの合計である、Ｌ１＋Ｌ２１＋Ｌ３１＋Ｌ４１、及び、Ｌ１＋Ｌ２２＋Ｌ３２＋Ｌ４２を求め、例えば、検出されたボーンの長さが長い右足側のＬ１＋Ｌ２１＋Ｌ３１＋Ｌ４１に補正定数を乗じた値を身長画素数とする。 In the example of FIG. 26, bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 are detected, respectively, and the right foot bone B61 and bone B71 and the left foot bone B62 and bone B72 overlap each other. .. The total of these bones, L1 + L21 + L31 + L41 and L1 + L22 + L32 + L42, is obtained, and for example, the value obtained by multiplying L1 + L21 + L31 + L41 on the right foot side where the detected bone length is long by a correction constant is taken as the height pixel number.
　図２７の例では、ボーンＢ１、ボーンＢ５１及びボーンＢ５２、ボーンＢ６１及びボーンＢ６２、ボーンＢ７１及びボーンＢ７２がそれぞれ検出され、右足のボーンＢ６１及びボーンＢ７１と左足のボーンＢ６２及びボーンＢ７２が重なっている。これらのボーンの合計である、Ｌ１＋Ｌ２１＋Ｌ３１＋Ｌ４１、及び、Ｌ１＋Ｌ２２＋Ｌ３２＋Ｌ４２を求め、例えば、検出されたボーンの長さが長い左足側のＬ１＋Ｌ２２＋Ｌ３２＋Ｌ４２に補正定数を乗じた値を身長画素数とする。 In the example of FIG. 27, bone B1, bone B51 and bone B52, bone B61 and bone B62, bone B71 and bone B72 are detected, respectively, and the right foot bone B61 and bone B71 and the left foot bone B62 and bone B72 overlap each other. .. The total of these bones, L1 + L21 + L31 + L41 and L1 + L22 + L32 + L42, is obtained, and for example, the value obtained by multiplying L1 + L22 + L32 + L42 on the left foot side where the detected bone length is long by a correction constant is taken as the height pixel number.
　具体例１では、頭から足までのボーンの長さを合計することで身長を求めることができるため、簡易な方法で身長画素数を求めることができる。また、機械学習を用いた骨格推定技術により、少なくとも頭から足までの骨格を検出できればよいため、しゃがみ込んでいる状態など、必ずしも人物の全体が画像に写っていない場合でも精度よく身長画素数を推定することができる。 In Specific Example 1, the height can be calculated by summing the lengths of the bones from the head to the feet, so the number of height pixels can be calculated by a simple method. In addition, since it is only necessary to detect the skeleton from the head to the feet by skeleton estimation technology using machine learning, the number of height pixels can be accurately calculated even when the entire person is not always shown in the image, such as when crouching down. Can be estimated.
　＜具体例２＞
　具体例２では、２次元骨格構造に含まれる骨の長さと２次元画像空間上の人物の全身の長さとの関係を示す２次元骨格モデルを用いて身長画素数を求める。 <Specific example 2>
In Specific Example 2, the number of height pixels is obtained using a two-dimensional skeleton model showing the relationship between the length of the bone included in the two-dimensional skeleton structure and the length of the whole body of the person in the two-dimensional image space.
　図２８は、具体例２で用いる、２次元画像空間上の各ボーンの長さと２次元画像空間上の全身の長さとの関係を示す人体モデル（２次元骨格モデル）３０１である。図２８に示すように、平均的な人物の各ボーンの長さと全身の長さとの関係（全身の長さに対する各ボーンの長さの割合）を、人体モデル３０１の各ボーンに対応付ける。例えば、頭のボーンＢ１の長さは全身の長さ×０．２（２０％）であり、右手のボーンＢ４１の長さは全身の長さ×０．１５（１５％）であり、右足のボーンＢ７１の長さは全身の長さ×０．２５（２５％）である。このような人体モデル３０１の情報をデータベース１１０に記憶しておくことで、各ボーンの長さから平均的な全身の長さを求めることができる。平均的な人物の人体モデルの他に、年代、性別、国籍等の人物の属性ごとに人体モデルを用意してもよい。これにより、人物の属性に応じて適切に全身の長さ（身長）を求めることができる。 FIG. 28 is a human body model (two-dimensional skeleton model) 301 showing the relationship between the length of each bone in the two-dimensional image space and the length of the whole body in the two-dimensional image space used in the second embodiment. As shown in FIG. 28, the relationship between the length of each bone of an average person and the length of the whole body (the ratio of the length of each bone to the length of the whole body) is associated with each bone of the human body model 301. For example, the length of the bone B1 of the head is the length of the whole body × 0.2 (20%), the length of the bone B41 of the right hand is the length of the whole body × 0.15 (15%), and the length of the right foot. The length of the bone B71 is the length of the whole body × 0.25 (25%). By storing the information of the human body model 301 in the database 110, the average whole body length can be obtained from the length of each bone. In addition to the human body model of an average person, a human body model may be prepared for each attribute of the person such as age, gender, and nationality. As a result, the length (height) of the whole body can be appropriately obtained according to the attributes of the person.
　具体例２では、図２１に示すように、身長算出部１０８は、各ボーンの長さを取得する（Ｓ２２１）。身長算出部１０８は、検出された骨格構造において、全てのボーンの長さ（２次元画像空間上の長さ）を取得する。図２９は、しゃがみ込んでいる状態の人物を右斜め後ろから撮像し、骨格構造を検出した例である。この例では、人物の顔や左側面が写っていないことから、頭のボーンと左腕及び左手のボーンが検出できていない。このため、検出されているボーンＢ２１、Ｂ２２、Ｂ３１、Ｂ４１、Ｂ５１、Ｂ５２、Ｂ６１、Ｂ６２、Ｂ７１、Ｂ７２の各長さを取得する。 In Specific Example 2, as shown in FIG. 21, the height calculation unit 108 acquires the length of each bone (S221). The height calculation unit 108 acquires the lengths (lengths in the two-dimensional image space) of all the bones in the detected skeletal structure. FIG. 29 is an example in which a person in a crouched state is imaged from diagonally right behind and the skeletal structure is detected. In this example, since the face and left side of the person are not shown, the bones of the head, the left arm, and the bones of the left hand cannot be detected. Therefore, the lengths of the detected bones B21, B22, B31, B41, B51, B52, B61, B62, B71, and B72 are acquired.
　続いて、身長算出部１０８は、図２１に示すように、人体モデルに基づき、各ボーンの長さから身長画素数を算出する（Ｓ２２２）。身長算出部１０８は、図２８のような、各ボーンと全身の長さとの関係を示す人体モデル３０１を参照し、各ボーンの長さから身長画素数を求める。例えば、右手のボーンＢ４１の長さが全身の長さ×０．１５であるため、ボーンＢ４１の長さ／０．１５によりボーンＢ４１に基づいた身長画素数を求める。また、右足のボーンＢ７１の長さが全身の長さ×０．２５であるため、ボーンＢ７１の長さ／０．２５によりボーンＢ７１に基づいた身長画素数を求める。 Subsequently, as shown in FIG. 21, the height calculation unit 108 calculates the number of height pixels from the length of each bone based on the human body model (S222). The height calculation unit 108 refers to the human body model 301 showing the relationship between each bone and the length of the whole body as shown in FIG. 28, and obtains the number of height pixels from the length of each bone. For example, since the length of the bone B41 on the right hand is the length of the whole body × 0.15, the number of height pixels based on the bone B41 is obtained from the length of the bone B41 / 0.15. Further, since the length of the bone B71 of the right foot is the length of the whole body × 0.25, the number of height pixels based on the bone B71 is obtained from the length of the bone B71 / 0.25.
　このとき参照する人体モデルは、例えば、平均的な人物の人体モデルであるが、年代、性別、国籍等の人物の属性に応じて人体モデルを選択してもよい。例えば、撮像した画像に人物の顔が写っている場合、顔に基づいて人物の属性を識別し、識別した属性に対応する人体モデルを参照する。属性ごとの顔を機械学習した情報を参照し、画像の顔の特徴から人物の属性を認識することができる。また、画像から人物の属性が識別できない場合に、平均的な人物の人体モデルを用いてもよい。 The human body model referred to at this time is, for example, a human body model of an average person, but a human body model may be selected according to the attributes of the person such as age, gender, and nationality. For example, when a person's face is shown in the captured image, the attribute of the person is identified based on the face, and the human body model corresponding to the identified attribute is referred to. It is possible to recognize a person's attributes from the facial features of the image by referring to the information obtained by machine learning the face for each attribute. Further, when the attribute of the person cannot be identified from the image, the human body model of the average person may be used.
　また、ボーンの長さから算出した身長画素数をカメラパラメータにより補正してもよい。例えばカメラを高い位置において、人物を見下ろすように撮影した場合、二次元骨格構造において肩幅のボーン等の横の長さはカメラの俯角の影響を受けないが、首－腰のボーン等の縦の長さは、カメラの俯角が大きくなる程小さくなる。そうすると、肩幅のボーン等の横の長さから算出した身長画素数が実際より大きくなる傾向がある。そこで、カメラパラメータを活用すると、人物がどの程度の角度でカメラに見下ろされているかがわかるため、この俯角の情報を使って正面から撮影したような二次元骨格構造に補正することができる。これによって、より正確に身長画素数を算出できる。 Further, the number of height pixels calculated from the length of the bone may be corrected by the camera parameter. For example, when the camera is taken at a high position and looking down at a person, the horizontal length of the shoulder-width bones, etc. is not affected by the depression angle of the camera in the two-dimensional skeletal structure, but the vertical length of the neck-waist bones, etc. The length decreases as the depression angle of the camera increases. Then, the number of height pixels calculated from the horizontal length of the shoulder-width bones and the like tends to be larger than the actual number. Therefore, by utilizing the camera parameters, it is possible to know the angle at which the person is looking down at the camera, and this information on the depression angle can be used to correct the two-dimensional skeleton structure as if it were taken from the front. This makes it possible to calculate the number of height pixels more accurately.
　続いて、身長算出部１０８は、図２１に示すように、身長画素数の最適値を算出する（Ｓ２２３）。身長算出部１０８は、ボーンごとに求めた身長画素数から身長画素数の最適値を算出する。例えば、図３０に示すような、ボーンごとに求めた身長画素数のヒストグラムを生成し、その中で大きい身長画素数を選択する。つまり、複数のボーンに基づいて求められた複数の身長画素数の中で他よりも長い身長画素数を選択する。例えば、上位３０％を有効な値とし、図３０ではボーンＢ７１、Ｂ６１、Ｂ５１による身長画素数を選択する。選択した身長画素数の平均を最適値として求めてもよいし、最も大きい身長画素数を最適値としてもよい。２次元画像のボーンの長さから身長を求めるため、ボーンを正面からできていない場合、すなわち、ボーンがカメラから見て奥行き方向に傾いて撮像された場合、ボーンの長さが正面から撮像した場合よりも短くなる。そうすると、身長画素数が大きい値は、身長画素数が小さい値よりも、正面から撮像された可能性が高く、より尤もらしい値となることから、より大きい値を最適値とする。 Subsequently, the height calculation unit 108 calculates the optimum value of the number of height pixels as shown in FIG. 21 (S223). The height calculation unit 108 calculates the optimum value of the number of height pixels from the number of height pixels obtained for each bone. For example, as shown in FIG. 30, a histogram of the number of height pixels obtained for each bone is generated, and a large number of height pixels is selected from the histogram. That is, the number of height pixels longer than the others is selected from the plurality of height pixels obtained based on the plurality of bones. For example, the upper 30% is set as a valid value, and in FIG. 30, the number of height pixels by bones B71, B61, and B51 is selected. The average number of selected height pixels may be obtained as the optimum value, or the largest number of height pixels may be used as the optimum value. Since the height is calculated from the length of the bone in the 2D image, the length of the bone is taken from the front when the bone is not made from the front, that is, when the bone is tilted in the depth direction when viewed from the camera. It will be shorter than the case. Then, a value having a large number of height pixels is more likely to be captured from the front than a value having a small number of height pixels, and is a more plausible value. Therefore, a larger value is set as the optimum value.
　具体例２では、２次元画像空間上のボーンと全身の長さとの関係を示す人体モデルを用いて、検出した骨格構造のボーンに基づき身長画素数を求めるため、頭から足までの全ての骨格が得られない場合でも、一部のボーンから身長画素数を求めることができる。特に、複数のボーンから求められた値のうち、より大きい値を採用することで、精度よく身長画素数を推定することができる。 In Specific Example 2, the number of height pixels is calculated based on the detected bones of the skeleton structure using a human body model showing the relationship between the bones in the two-dimensional image space and the length of the whole body, so that all the skeletons from the head to the feet are obtained. Even if is not obtained, the number of height pixels can be obtained from some bones. In particular, the number of height pixels can be estimated accurately by adopting a larger value among the values obtained from a plurality of bones.
　＜具体例３＞
　具体例３では、２次元骨格構造を３次元人体モデル（３次元骨格モデル）にフィッティングさせて、フィッティングした３次元人体モデルの身長画素数を用いて全身の骨格ベクトルを求める。 <Specific example 3>
In Specific Example 3, the two-dimensional skeleton structure is fitted to the three-dimensional human body model (three-dimensional skeleton model), and the skeleton vector of the whole body is obtained using the number of height pixels of the fitted three-dimensional human body model.
　具体例３では、図２２に示すように、身長算出部１０８は、まず、カメラ２００の撮像した画像に基づき、カメラパラメータを算出する（Ｓ２３１）。身長算出部１０８は、カメラ２００が撮像した複数の画像の中から、予め長さが分かっている物体を抽出し、抽出した物体の大きさ（画素数）からカメラパラメータを求める。なお、カメラパラメータを予め求めておき、求めておいたカメラパラメータを必要に応じて取得してもよい。 In Specific Example 3, as shown in FIG. 22, the height calculation unit 108 first calculates the camera parameters based on the image captured by the camera 200 (S231). The height calculation unit 108 extracts an object whose length is known in advance from a plurality of images captured by the camera 200, and obtains a camera parameter from the size (number of pixels) of the extracted object. The camera parameters may be obtained in advance, and the obtained camera parameters may be acquired as needed.
　続いて、身長算出部１０８は、３次元人体モデルの配置及び高さを調整する（Ｓ２３２）。身長算出部１０８は、検出された２次元骨格構造に対し、身長画素数算出用の３次元人体モデルを用意し、カメラパラメータに基づいて、同じ２次元画像内に配置する。具体的には、カメラパラメータと、２次元骨格構造から、「実世界におけるカメラと人物の相対的な位置関係」を特定する。例えば、仮にカメラの位置を座標（０，０，０）としたときに、人物が立っている（または座っている）位置の座標（ｘ，ｙ，ｚ）を特定する。そして、特定した人物と同じ位置（ｘ，ｙ，ｚ）に３次元人体モデルを配置して撮像した場合の画像を想定することで、２次元骨格構造と３次元人体モデルを重ね合わせる。 Subsequently, the height calculation unit 108 adjusts the arrangement and height of the three-dimensional human body model (S232). The height calculation unit 108 prepares a three-dimensional human body model for calculating the number of height pixels for the detected two-dimensional skeleton structure, and arranges the three-dimensional human body model in the same two-dimensional image based on the camera parameters. Specifically, the "relative positional relationship between the camera and the person in the real world" is specified from the camera parameters and the two-dimensional skeleton structure. For example, assuming that the position of the camera is the coordinates (0, 0, 0), the coordinates (x, y, z) of the position where the person is standing (or sitting) are specified. Then, by assuming an image in which a three-dimensional human body model is placed at the same position (x, y, z) as the specified person and captured, the two-dimensional skeleton structure and the three-dimensional human body model are superimposed.
　図３１は、しゃがみ込んでいる人物を左斜め前から撮像し、２次元骨格構造４０１を検出した例である。２次元骨格構造４０１は、２次元の座標情報を有する。なお、全てのボーンを検出していることが好ましいが、一部のボーンが検出されていなくてもよい。この２次元骨格構造４０１に対し、図３２のような、３次元人体モデル４０２を用意する。３次元人体モデル（３次元骨格モデル）４０２は、３次元の座標情報を有し、２次元骨格構造４０１と同じ形状の骨格のモデルである。そして、図３３のように、検出した２次元骨格構造４０１に対し、用意した３次元人体モデル４０２を配置し重ね合わせる。また、重ね合わせるとともに、３次元人体モデル４０２の高さを２次元骨格構造４０１に合うように調整する。 FIG. 31 is an example in which a crouching person is imaged diagonally from the front left and the two-dimensional skeleton structure 401 is detected. The two-dimensional skeleton structure 401 has two-dimensional coordinate information. It is preferable that all bones are detected, but some bones may not be detected. For this two-dimensional skeleton structure 401, a three-dimensional human body model 402 as shown in FIG. 32 is prepared. The three-dimensional human body model (three-dimensional skeleton model) 402 has three-dimensional coordinate information and is a model of a skeleton having the same shape as the two-dimensional skeleton structure 401. Then, as shown in FIG. 33, the prepared three-dimensional human body model 402 is arranged and superimposed on the detected two-dimensional skeleton structure 401. In addition, the height of the three-dimensional human body model 402 is adjusted so as to match the two-dimensional skeleton structure 401.
　なお、このとき用意する３次元人体モデル４０２は、図３３のように、２次元骨格構造４０１の姿勢に近い状態のモデルでもよいし、直立した状態のモデルでもよい。例えば、機械学習を用いて２次元画像から３次元空間の姿勢を推定する技術を用いて、推定した姿勢の３次元人体モデル４０２を生成してもよい。２次元画像の関節と３次元空間の関節の情報を学習することで、２次元画像から３次元の姿勢を推定することができる。 The three-dimensional human body model 402 prepared at this time may be a model in a state close to the posture of the two-dimensional skeleton structure 401 as shown in FIG. 33, or may be a model in an upright state. For example, a 3D human body model 402 of the estimated posture may be generated by using a technique of estimating the posture of the 3D space from the 2D image using machine learning. By learning the information of the joints in the 2D image and the joints in the 3D space, the 3D posture can be estimated from the 2D image.
　続いて、身長算出部１０８は、図２２に示すように、３次元人体モデルを２次元骨格構造にフィッティングする（Ｓ２３３）。身長算出部１０８は、図３４のように、３次元人体モデル４０２を２次元骨格構造４０１に重ね合わせた状態で、３次元人体モデル４０２と２次元骨格構造４０１の姿勢が一致するように、３次元人体モデル４０２を変形させる。すなわち、３次元人体モデル４０２の身長、体の向き、関節の角度を調整し、２次元骨格構造４０１との差異がなくなるように最適化する。例えば、３次元人体モデル４０２の関節を、人の可動範囲で回転させていき、また、３次元人体モデル４０２の全体を回転させたり、全体のサイズを調整する。なお、３次元人体モデルと２次元骨格構造のフィッティング（あてはめ）は、２次元空間（２次元座標）上で行う。すなわち、２次元空間に３次元人体モデルを写像し、変形させた３次元人体モデルが２次元空間（画像）でどのように変化するかを考慮して、３次元人体モデルを２次元骨格構造に最適化する。 Subsequently, the height calculation unit 108 fits the three-dimensional human body model into the two-dimensional skeletal structure as shown in FIG. 22 (S233). As shown in FIG. 34, the height calculation unit 108 superimposes the three-dimensional human body model 402 on the two-dimensional skeletal structure 401 so that the postures of the three-dimensional human body model 402 and the two-dimensional skeletal structure 401 match. Transform the dimensional human body model 402. That is, the height, body orientation, and joint angle of the three-dimensional human body model 402 are adjusted and optimized so that there is no difference from the two-dimensional skeletal structure 401. For example, the joints of the three-dimensional human body model 402 are rotated within the range of movement of the person, the entire three-dimensional human body model 402 is rotated, and the overall size is adjusted. The fitting of the three-dimensional human body model and the two-dimensional skeletal structure is performed in the two-dimensional space (two-dimensional coordinates). That is, the 3D human body model is mapped to the 2D space, and the 3D human body model is converted into a 2D skeletal structure in consideration of how the deformed 3D human body model changes in the 2D space (image). Optimize.
　続いて、身長算出部１０８は、図２２に示すように、フィッティングさせた３次元人体モデルの身長画素数を算出する（Ｓ２３４）。身長算出部１０８は、図３５のように、３次元人体モデル４０２と２次元骨格構造４０１の差異がなくなり、姿勢が一致すると、その状態の３次元人体モデル４０２の身長画素数を求める。最適化された３次元人体モデル４０２を直立させた状態として、カメラパラメータに基づき、２次元空間上の全身の長さを求める。例えば、３次元人体モデル４０２を直立させた場合の頭から足までのボーンの長さ（画素数）により身長画素数を算出する。具体例１と同様に、３次元人体モデル４０２の頭部から足部までのボーンの長さを合計してもよい。 Subsequently, the height calculation unit 108 calculates the number of height pixels of the fitted three-dimensional human body model as shown in FIG. 22 (S234). As shown in FIG. 35, the height calculation unit 108 obtains the number of height pixels of the three-dimensional human body model 402 in that state when the difference between the three-dimensional human body model 402 and the two-dimensional skeleton structure 401 disappears and the postures match. With the optimized 3D human body model 402 upright, the length of the whole body in 2D space is obtained based on the camera parameters. For example, the height pixel number is calculated from the bone length (number of pixels) from the head to the foot when the three-dimensional human body model 402 is upright. Similar to the first embodiment, the lengths of the bones from the head to the foot of the three-dimensional human body model 402 may be totaled.
　具体例３では、カメラパラメータに基づいて３次元人体モデルを２次元骨格構造にフィッティングさせて、その３次元人体モデルに基づいて身長画素数を求めることで、全てのボーンが正面に写っていない場合、すなわち、全てのボーンが斜めに映っているため誤差が大きい場合でも、精度よく身長画素数を推定することができる。 In Specific Example 3, when a 3D human body model is fitted to a 2D skeletal structure based on camera parameters and the number of height pixels is obtained based on the 3D human body model, all bones are not shown in the front. That is, since all the bones are reflected diagonally, the number of height pixels can be estimated accurately even when the error is large.
　＜正規化処理＞
　図１９に示すように、画像処理装置１００は、身長画素数算出処理に続いて、正規化処理（Ｓ２０２）を行う。正規化処理では、図２３に示すように、特徴量算出部１０３は、キーポイント高さを算出する（Ｓ２４１）。特徴量算出部１０３は、検出された骨格構造に含まれる全てのキーポイントのキーポイント高さ（画素数）を算出する。キーポイント高さは、骨格構造の最下端（例えばいずれかの足のキーポイント）からそのキーポイントまでの高さ方向の長さ（画素数）である。ここでは、一例として、キーポイント高さを、画像におけるキーポイントのＹ座標から求める。なお、上記のように、キーポイント高さは、カメラパラメータに基づいた鉛直投影軸に沿った方向の長さから求めてもよい。例えば、図２４の例で、首のキーポイントＡ２の高さ（ｙｉ）は、キーポイントＡ２のＹ座標から右足のキーポイントＡ８１または左足のキーポイントＡ８２のＹ座標を引いた値である。 <Normalization process>
As shown in FIG. 19, the image processing apparatus 100 performs a normalization process (S202) following the height pixel number calculation process. In the normalization process, as shown in FIG. 23, the feature amount calculation unit 103 calculates the key point height (S241). The feature amount calculation unit 103 calculates the key point height (number of pixels) of all the key points included in the detected skeleton structure. The key point height is the length (number of pixels) in the height direction from the lowest end of the skeleton structure (for example, the key point of any foot) to the key point. Here, as an example, the height of the key point is obtained from the Y coordinate of the key point in the image. As described above, the key point height may be obtained from the length in the direction along the vertical projection axis based on the camera parameters. For example, in the example of FIG. 24, the height (y) of the key point A2 of the neck is a value obtained by subtracting the Y coordinate of the key point A81 of the right foot or the Y coordinate of the key point A82 of the left foot from the Y coordinate of the key point A2.
　続いて、特徴量算出部１０３は、正規化のための基準点を特定する（Ｓ２４２）。基準点は、キーポイントの相対的な高さを表すための基準となる点である。基準点は、予め設定されていてもよいし、ユーザが選択できるようにしてもよい。基準点は、骨格構造の中心もしくは中心よりも高い（画像の上下方向における上である）ことが好ましく、例えば、首のキーポイントの座標を基準点とする。なお、首に限らず頭やその他のキーポイントの座標を基準点としてもよい。キーポイントに限らず、任意の座標（例えば骨格構造の中心座標等）を基準点としてもよい。 Subsequently, the feature amount calculation unit 103 specifies a reference point for normalization (S242). The reference point is a reference point for expressing the relative height of the key point. The reference point may be preset or may be selectable by the user. The reference point is preferably at the center of the skeletal structure or higher than the center (upper in the vertical direction of the image), and for example, the coordinates of the key point of the neck are used as the reference point. The coordinates of the head and other key points, not limited to the neck, may be used as the reference point. Not limited to the key point, any coordinate (for example, the center coordinate of the skeleton structure) may be used as a reference point.
　続いて、特徴量算出部１０３は、キーポイント高さ（ｙｉ）を身長画素数で正規化する（Ｓ２４３）。特徴量算出部１０３は、各キーポイントのキーポイント高さ、基準点、身長画素数を用いて、各キーポイントを正規化する。具体的には、特徴量算出部１０３は、基準点に対するキーポイントの相対的な高さを身長画素数により正規化する。ここでは、高さ方向のみに着目する例として、Ｙ座標のみを抽出し、また、基準点を首のキーポイントとして正規化を行う。具体的には、基準点（首のキーポイント）のＹ座標を（ｙｃ）として、次の式（１）を用いて、特徴量（正規化値）を求める。なお、カメラパラメータに基づいた鉛直投影軸を用いる場合は、（ｙｉ）及び（ｙｃ）を鉛直投影軸に沿った方向の値に変換する。
Figure JPOXMLDOC01-appb-M000001
 Subsequently, the feature amount calculation unit 103 normalizes the key point height (yi) by the number of height pixels (S243). The feature amount calculation unit 103 normalizes each key point by using the key point height, the reference point, and the number of height pixels of each key point. Specifically, the feature amount calculation unit 103 normalizes the relative height of the key point with respect to the reference point by the number of height pixels. Here, as an example focusing only on the height direction, only the Y coordinate is extracted, and normalization is performed with the reference point as the key point of the neck. Specifically, the feature amount (normalized value) is obtained by using the following equation (1) with the Y coordinate of the reference point (key point of the neck) as (yc). When using the vertical projection axis based on the camera parameters, (yi) and (yc) are converted into values in the direction along the vertical projection axis.
Figure JPOXMLDOC01-appb-M000001
　例えば、キーポイントが１８個の場合、各キーポイントの１８点の座標（ｘ０、ｙ０）、（ｘ１、ｙ１）、・・・（ｘ１７、ｙ１７）を、上記式（１）を用いて、次のように１８次元の特徴量に変換する。
Figure JPOXMLDOC01-appb-M000002
 For example, when there are 18 key points, the coordinates (x0, y0), (x1, y1), ... (X17, y17) of the 18 points of each key point are set to the following using the above equation (1). It is converted into an 18-dimensional feature amount as in.
Figure JPOXMLDOC01-appb-M000002
　図３６は、特徴量算出部１０３が求めた各キーポイントの特徴量の例を示している。この例では、首のキーポイントＡ２を基準点とするため、キーポイントＡ２の特徴量は０．０となり、首と同じ高さの右肩のキーポイントＡ３１及び左肩のキーポイントＡ３２の特徴量も０．０である。首よりも高い頭のキーポイントＡ１の特徴量は－０．２である。首よりも低い右手のキーポイントＡ５１及び左手のキーポイントＡ５２の特徴量は０．４であり、右足のキーポイントＡ８１及び左足のキーポイントＡ８２の特徴量は０．９である。この状態から人物が左手を挙げると、図３７のように左手が基準点よりも高くなるため、左手のキーポイントＡ５２の特徴量は－０．４となる。一方で、Ｙ軸の座標のみを用いて正規化を行っているため、図３８のように、図３６に比べて骨格構造の幅が変わっても特徴量は変わらない。すなわち、本実施の形態の特徴量（正規化値）は、骨格構造（キーポイント）の高さ方向（Ｙ方向）の特徴を示しており、骨格構造の横方向（Ｘ方向）の変化に影響を受けない。 FIG. 36 shows an example of the feature amount of each key point obtained by the feature amount calculation unit 103. In this example, since the key point A2 of the neck is used as the reference point, the feature amount of the key point A2 is 0.0, and the feature amount of the key point A31 on the right shoulder and the key point A32 on the left shoulder at the same height as the neck are also. It is 0.0. The feature amount of the key point A1 of the head higher than the neck is -0.2. The feature amount of the right hand key point A51 and the left hand key point A52 lower than the neck is 0.4, and the feature amount of the right foot key point A81 and the left foot key point A82 is 0.9. When the person raises his / her left hand from this state, the left hand becomes higher than the reference point as shown in FIG. 37, so that the feature amount of the key point A52 of the left hand is −0.4. On the other hand, since the normalization is performed using only the coordinates of the Y axis, the feature amount does not change even if the width of the skeleton structure changes as compared with FIG. 36, as shown in FIG. 38. That is, the feature amount (normalized value) of the present embodiment shows the characteristics of the skeleton structure (key point) in the height direction (Y direction), and affects the change of the skeleton structure in the lateral direction (X direction). Do not receive.
　以上のように、本実施の形態では、２次元画像から人物の骨格構造を検出し、検出した骨格構造から求めた身長画素数（２次元画像空間上の直立時の高さ）を用いて、骨格構造の各キーポイントを正規化する。この正規化された特徴量を用いることで、分類や検索等を行った場合のロバスト性を向上することができる。すなわち、本実施の形態の特徴量は、上記のように人物の横方向の変化に影響を受けないため、人物の向きや人物の体型の変化に対しロバスト性が高い。 As described above, in the present embodiment, the skeleton structure of a person is detected from the two-dimensional image, and the number of height pixels (height when standing upright on the two-dimensional image space) obtained from the detected skeleton structure is used. Normalize each key point in the skeletal structure. By using this normalized feature amount, it is possible to improve the robustness when classification, search, or the like is performed. That is, since the feature amount of the present embodiment is not affected by the lateral change of the person as described above, it is highly robust to the change of the direction of the person and the body shape of the person.
　さらに、本実施の形態では、ＯｐｅｎＰｏｓｅ等の骨格推定技術を用いて人物の骨格構造を検出することで実現できるため、人物の姿勢等を学習する学習データを用意する必要がない。また、骨格構造のキーポイントを正規化し、データベースに格納しておくことで、人物の姿勢等の分類や検索が可能となるため、未知な姿勢に対しても分類や検索を行うことができる。また、骨格構造のキーポイントを正規化することで、明確でわかりやすい特徴量を得ることができるため、機械学習のようにブラックボックス型のアルゴリズムと異なり、処理結果に対するユーザの納得性が高い。 Further, in this embodiment, since it can be realized by detecting the skeleton structure of a person using a skeleton estimation technique such as OpenPose, it is not necessary to prepare learning data for learning the posture of the person. Further, by normalizing the key points of the skeletal structure and storing them in the database, it is possible to classify and search the posture of a person, so that it is possible to classify and search even an unknown posture. In addition, by normalizing the key points of the skeleton structure, clear and easy-to-understand features can be obtained, so that the user is highly convinced of the processing result, unlike the black box type algorithm such as machine learning.
　以上、図面を参照して本発明の実施形態について述べたが、これらは本発明の例示であり、上記以外の様々な構成を採用することもできる。 Although the embodiments of the present invention have been described above with reference to the drawings, these are examples of the present invention, and various configurations other than the above can be adopted.
　また、上述の説明で用いた複数のフローチャートでは、複数の工程（処理）が順番に記載されているが、各実施形態で実行される工程の実行順序は、その記載の順番に制限されない。各実施形態では、図示される工程の順番を内容的に支障のない範囲で変更することができる。また、上述の各実施形態は、内容が相反しない範囲で組み合わせることができる。 Further, in the plurality of flowcharts used in the above description, a plurality of steps (processes) are described in order, but the execution order of the steps executed in each embodiment is not limited to the order of description. In each embodiment, the order of the illustrated steps can be changed within a range that does not hinder the contents. In addition, the above-mentioned embodiments can be combined as long as the contents do not conflict with each other.
　上記の実施形態の一部又は全部は、以下の付記のようにも記載されうるが、以下には限定されない。
１．　複数の人物各々の姿勢を指定した第１のクエリを取得する第１のクエリ取得手段と、
　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する第２のクエリ取得手段と、
　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する検索手段と、
を有する画像処理装置。
２．　クエリ画像に含まれる人物の２次元骨格構造を検出する骨格構造検出手段と、
　前記検出された２次元骨格構造の特徴量を算出する特徴量算出手段と、
をさらに有し、
　前記第１のクエリ取得手段は、前記クエリ画像から算出された２次元骨格構造の特徴量を、前記第１のクエリとして取得する１に記載の画像処理装置。
３．　前記第１のクエリ取得手段は、予め用意された複数の選択肢の中からユーザが選択した人物の姿勢を、前記第１のクエリとして取得する１に記載の画像処理装置。
４．　前記第２のクエリ取得手段は、クエリ画像に含まれる複数の人物間の物理的距離及び時間的距離に基づき、前記第２のクエリを決定する１から３のいずれかに記載の画像処理装置。
５．　前記第２のクエリ取得手段は、予め用意された複数の選択肢の中からユーザが選択した値、又は、ユーザが入力した数値を、前記第２のクエリとして取得する１から３のいずれかに記載の画像処理装置。
６．　前記検索手段は、
　　動画像の中から、
　　前記第１のクエリで指定された第１の姿勢をとる第１の人物と、前記第１のクエリで指定された第２の姿勢をとる第２の人物が現れ、
　　前記第１の人物が前記第１の姿勢をとる第１のタイミングと、前記第２の人物が前記第２の姿勢をとる第２のタイミングの時間差は、前記第２のクエリで指定された時間的距離の条件を満たし、かつ、
　前記第１のタイミングと前記第２のタイミングとの間の時間帯の少なくとも一部において、前記第１の人物と前記第２の人物との間の物理的距離が、前記第２のクエリで指定された物理的距離の条件を満たすシーンを検索する１から５のいずれかに記載の画像処理装置。
７．　前記検索手段は、
　　複数の静止画像の中から、
　　前記第１のクエリで指定された第１の姿勢をとる第１の人物と、前記第１のクエリで指定された第２の姿勢をとる第２の人物が現れ、
　　前記第１の人物と前記第２の人物との間の物理的距離が、前記第２のクエリで指定された物理的距離の条件を満たす静止画像を検索する１から５のいずれかに記載の画像処理装置。
８．　前記第１のクエリで指定された複数の人物各々の姿勢は、同じ姿勢、又は、異なる姿勢である１から７のいずれかに記載の画像処理装置。
９．　コンピュータが、
　　複数の人物各々の姿勢を指定した第１のクエリを取得し、
　　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得し、
　　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する画像処理方法。
１０．　コンピュータを、
　　複数の人物各々の姿勢を指定した第１のクエリを取得する第１のクエリ取得手段、
　　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する第２のクエリ取得手段、
　　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する検索手段、
として機能させるプログラム。 Some or all of the above embodiments may also be described, but not limited to:
1. 1. A first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people,
A second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
A search means for executing an image search based on the first query and the second query, and
Image processing device with.
2. 2. A skeletal structure detecting means for detecting the two-dimensional skeletal structure of a person included in a query image,
A feature amount calculating means for calculating the feature amount of the detected two-dimensional skeletal structure, and
Have more
The image processing apparatus according to 1, wherein the first query acquisition means acquires a feature amount of a two-dimensional skeleton structure calculated from the query image as the first query.
3. 3. The image processing device according to 1, wherein the first query acquisition means acquires the posture of a person selected by a user from a plurality of options prepared in advance as the first query.
4. The image processing apparatus according to any one of 1 to 3, wherein the second query acquisition means determines the second query based on the physical distance and the temporal distance between a plurality of persons included in the query image.
5. The second query acquisition means describes in any one of 1 to 3 for acquiring a value selected by the user from a plurality of options prepared in advance or a numerical value input by the user as the second query. Image processing equipment.
6. The search means is
From the moving image
A first person in the first posture specified in the first query and a second person in the second posture specified in the first query appear.
The time difference between the first timing in which the first person takes the first posture and the second timing in which the second person takes the second posture is the time specified in the second query. Satisfy the conditions of target distance and
The physical distance between the first person and the second person is specified in the second query in at least a portion of the time zone between the first timing and the second timing. The image processing apparatus according to any one of 1 to 5, which searches for a scene satisfying the condition of the physical distance.
7. The search means is
From multiple still images
A first person in the first posture specified in the first query and a second person in the second posture specified in the first query appear.
The description in any one of 1 to 5 for searching for a still image in which the physical distance between the first person and the second person satisfies the condition of the physical distance specified by the second query. Image processing device.
8. The image processing apparatus according to any one of 1 to 7, wherein the postures of the plurality of persons specified in the first query are the same posture or different postures.
9. The computer
Get the first query that specifies the posture of each of the multiple people
Obtain a second query that specifies at least one of the physical distance and the temporal distance between the plurality of people who take the specified posture.
An image processing method for executing an image search based on the first query and the second query.
10. Computer,
A first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people,
A second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
A search means for performing an image search based on the first query and the second query.
A program that functions as.
１　　　　画像処理システム
１０　　　　画像処理装置
１１　　　　骨格検出部
１２　　　　特徴量算出部
１３　　　　認識部
１００　　　　画像処理装置
１０１　　　　画像取得部
１０２　　　　骨格構造検出部
１０３　　　　特徴量算出部
１０４　　　　分類部
１０５　　　　検索部
１０６　　　　入力部
１０７　　　　表示部
１０８　　　　身長算出部
１０９　　　　第１のクエリ取得部
１１０　　　　データベース
１１１　　　　第２のクエリ取得部
２００　　　　カメラ
３００、３０１　　　　人体モデル
４０１　　　　２次元骨格構造
４０２　　　　３次元人体モデル 1 Image processing system 10 Image processing device 11 Skeleton detection unit 12 Feature amount calculation unit 13 Recognition unit 100 Image processing device 101 Image acquisition unit 102 Skeleton structure detection unit 103 Feature amount calculation unit 104 Classification unit 105 Search unit 106 Input unit 107 Display unit 108 Height calculation unit 109 First query acquisition unit 110 Database 111 Second query acquisition unit 200 Camera 300, 301 Human body model 401 Two-dimensional skeleton structure 402 Three-dimensional human body model

Claims (10)

  1. 　複数の人物各々の姿勢を指定した第１のクエリを取得する第１のクエリ取得手段と、
    　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する第２のクエリ取得手段と、
    　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する検索手段と、
    を有する画像処理装置。     A first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people,
    A second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
    A search means for executing an image search based on the first query and the second query, and
    Image processing device with.
  2. 　クエリ画像に含まれる人物の２次元骨格構造を検出する骨格構造検出手段と、
    　前記検出された２次元骨格構造の特徴量を算出する特徴量算出手段と、
    をさらに有し、
    　前記第１のクエリ取得手段は、前記クエリ画像から算出された２次元骨格構造の特徴量を、前記第１のクエリとして取得する請求項１に記載の画像処理装置。     A skeletal structure detecting means for detecting the two-dimensional skeletal structure of a person included in a query image,
    A feature amount calculating means for calculating the feature amount of the detected two-dimensional skeletal structure, and
    Have more
    The image processing apparatus according to claim 1, wherein the first query acquisition means acquires a feature amount of a two-dimensional skeleton structure calculated from the query image as the first query.
  3. 　前記第１のクエリ取得手段は、予め用意された複数の選択肢の中からユーザが選択した人物の姿勢を、前記第１のクエリとして取得する請求項１に記載の画像処理装置。 The image processing device according to claim 1, wherein the first query acquisition means acquires the posture of a person selected by the user from a plurality of options prepared in advance as the first query.
  4. 　前記第２のクエリ取得手段は、クエリ画像に含まれる複数の人物間の物理的距離及び時間的距離に基づき、前記第２のクエリを決定する請求項１から３のいずれか１項に記載の画像処理装置。 The second query acquisition means according to any one of claims 1 to 3, wherein the second query acquisition means determines the second query based on the physical distance and the temporal distance between a plurality of persons included in the query image. Image processing device.
  5. 　前記第２のクエリ取得手段は、予め用意された複数の選択肢の中からユーザが選択した値、又は、ユーザが入力した数値を、前記第２のクエリとして取得する請求項１から３のいずれか１項に記載の画像処理装置。 The second query acquisition means is any one of claims 1 to 3 for acquiring a value selected by the user from a plurality of options prepared in advance or a numerical value input by the user as the second query. The image processing apparatus according to item 1.
  6. 　前記検索手段は、
    　　動画像の中から、
    　　前記第１のクエリで指定された第１の姿勢をとる第１の人物と、前記第１のクエリで指定された第２の姿勢をとる第２の人物が現れ、
    　　前記第１の人物が前記第１の姿勢をとる第１のタイミングと、前記第２の人物が前記第２の姿勢をとる第２のタイミングの時間差は、前記第２のクエリで指定された時間的距離の条件を満たし、かつ、
    　前記第１のタイミングと前記第２のタイミングとの間の時間帯の少なくとも一部において、前記第１の人物と前記第２の人物との間の物理的距離が、前記第２のクエリで指定された物理的距離の条件を満たすシーンを検索する請求項１から５のいずれか１項に記載の画像処理装置。     The search means is
    From the moving image
    A first person in the first posture specified in the first query and a second person in the second posture specified in the first query appear.
    The time difference between the first timing in which the first person takes the first posture and the second timing in which the second person takes the second posture is the time specified in the second query. Satisfy the conditions of target distance and
    The physical distance between the first person and the second person is specified in the second query in at least a portion of the time zone between the first timing and the second timing. The image processing apparatus according to any one of claims 1 to 5, which searches for a scene satisfying the condition of the physical distance.
  7. 　前記検索手段は、
    　　複数の静止画像の中から、
    　　前記第１のクエリで指定された第１の姿勢をとる第１の人物と、前記第１のクエリで指定された第２の姿勢をとる第２の人物が現れ、
    　　前記第１の人物と前記第２の人物との間の物理的距離が、前記第２のクエリで指定された物理的距離の条件を満たす静止画像を検索する請求項１から５のいずれか１項に記載の画像処理装置。     The search means is
    From multiple still images
    A first person in the first posture specified in the first query and a second person in the second posture specified in the first query appear.
    Any one of claims 1 to 5 for searching for a still image in which the physical distance between the first person and the second person satisfies the condition of the physical distance specified by the second query. The image processing apparatus described in the section.
  8. 　前記第１のクエリで指定された複数の人物各々の姿勢は、同じ姿勢、又は、異なる姿勢である請求項１から７のいずれか１項に記載の画像処理装置。 The image processing apparatus according to any one of claims 1 to 7, wherein the postures of the plurality of persons specified in the first query are the same posture or different postures.
  9. 　コンピュータが、
    　　複数の人物各々の姿勢を指定した第１のクエリを取得し、
    　　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得し、
    　　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する画像処理方法。     The computer
    Get the first query that specifies the posture of each of the multiple people
    Obtain a second query that specifies at least one of the physical distance and the temporal distance between the plurality of people who take the specified posture.
    An image processing method for executing an image search based on the first query and the second query.
  10. 　コンピュータを、
    　　複数の人物各々の姿勢を指定した第１のクエリを取得する第１のクエリ取得手段、
    　　前記指定した姿勢をとる複数の人物間の物理的距離、及び、時間的距離の少なくとも一方を指定した第２のクエリを取得する第２のクエリ取得手段、
    　　前記第１のクエリ及び前記第２のクエリに基づく画像検索を実行する検索手段、
    として機能させるプログラム。     Computer,
    A first query acquisition means for acquiring a first query that specifies the posture of each of a plurality of people,
    A second query acquisition means for acquiring a second query that specifies at least one of a physical distance and a temporal distance between a plurality of persons taking the specified postures.
    A search means for performing an image search based on the first query and the second query.
    A program that functions as.
PCT/JP2020/026534 2020-07-07 2020-07-07 Image processing device, image processing method, and program WO2022009301A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/JP2020/026534 WO2022009301A1 (en) 2020-07-07 2020-07-07 Image processing device, image processing method, and program
JP2022534527A JP7416252B2 (en) 2020-07-07 2020-07-07 Image processing device, image processing method, and program

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2020/026534 WO2022009301A1 (en) 2020-07-07 2020-07-07 Image processing device, image processing method, and program

Publications (1)

Publication Number Publication Date
WO2022009301A1 true WO2022009301A1 (en) 2022-01-13

Family

ID=79552312

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2020/026534 WO2022009301A1 (en) 2020-07-07 2020-07-07 Image processing device, image processing method, and program

Country Status (2)

Country Link
JP (1) JP7416252B2 (en)
WO (1) WO2022009301A1 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023152973A1 (en) * 2022-02-14 2023-08-17 日本電気株式会社 Image processing device, image processing method, and program
WO2023152841A1 (en) * 2022-02-09 2023-08-17 日本電気株式会社 Image processing system, image processing method, and non-transitory computer-readable medium
WO2023175945A1 (en) * 2022-03-18 2023-09-21 日本電気株式会社 Action evaluation device, action evaluation method, and non-transitory computer-readable medium

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2019091138A (en) * 2017-11-13 2019-06-13 株式会社日立製作所 Image retrieving apparatus, image retrieving method, and setting screen used therefor

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2019091138A (en) * 2017-11-13 2019-06-13 株式会社日立製作所 Image retrieving apparatus, image retrieving method, and setting screen used therefor

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023152841A1 (en) * 2022-02-09 2023-08-17 日本電気株式会社 Image processing system, image processing method, and non-transitory computer-readable medium
WO2023152973A1 (en) * 2022-02-14 2023-08-17 日本電気株式会社 Image processing device, image processing method, and program
WO2023175945A1 (en) * 2022-03-18 2023-09-21 日本電気株式会社 Action evaluation device, action evaluation method, and non-transitory computer-readable medium

Also Published As

Publication number Publication date
JP7416252B2 (en) 2024-01-17
JPWO2022009301A1 (en) 2022-01-13

Similar Documents

Publication Publication Date Title
Khraief et al. Elderly fall detection based on multi-stream deep convolutional networks
WO2022009301A1 (en) Image processing device, image processing method, and program
US20050094879A1 (en) Method for visual-based recognition of an object
CN114616588A (en) Image processing apparatus, image processing method, and non-transitory computer-readable medium storing image processing program
JP2024103572A (en) Information processing device, information processing method, and program
WO2021250808A1 (en) Image processing device, image processing method, and program
WO2021255846A1 (en) Image processing device, image processing method, and program
WO2022079794A1 (en) Image selection device, image selection method, and program
JP7435781B2 (en) Image selection device, image selection method, and program
WO2022009279A1 (en) Image selection device, image selection method, and program
US11983242B2 (en) Learning data generation device, learning data generation method, and learning data generation program
WO2022003854A1 (en) Image processing device, image processing method, and program
WO2021250809A1 (en) Image processing apparatus, image processing method, and program
JP7396364B2 (en) Image processing device, image processing method, and image processing program
WO2021229750A1 (en) Image selection device, image selection method, and program
WO2021234935A1 (en) Image selection device, image selection method, and program
WO2022079795A1 (en) Image selection device, image selection method, and program
WO2021220365A1 (en) Image classification apparatus, image classification method, and program
WO2022249278A1 (en) Image processing device, image processing method, and program
WO2023084778A1 (en) Image processing device, image processing method, and program
JP7435754B2 (en) Image selection device, image selection method, and program
WO2023084780A1 (en) Image processing device, image processing method, and program
WO2022249331A1 (en) Image processing device, image processing method, and program
WO2023152974A1 (en) Image processing device, image processing method, and program
WO2023152971A1 (en) Image processing device, image processing method, and program

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20944691

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022534527

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20944691

Country of ref document: EP

Kind code of ref document: A1