JP2011059897A - Image analysis apparatus and method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image analysis apparatus capable of efficiently performing the detection processing of an object and the tracking processing of a specific object. <P>SOLUTION: The image analysis apparatus includes: a feature value memory for storing a plurality of reference feature values in association with the direction of an object; a feature value extraction unit for extracting a plurality of feature values from each of a plurality of moving image configuring images included in a moving image; an object extraction unit for extracting the object of a specific object from the moving image configuring image based on the adaptation of the plurality of extracted feature values corresponding to the plurality of reference feature values stored so as to be associated with the direction of the object; a direction specification part for specifying the direction of the specific object; a reference feature value calculation part for calculating the reference feature value adapted to the feature value of the extracted object based on the reference feature value stored in the feature value memory; and a feature value updating part for updating the reference feature value stored so as to be associated with the specified direction by the feature value memory with the reference feature value calculated by the reference feature value calculation part. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an image analysis device, an image analysis method, and a program.

An apparatus is known that can determine the type of feature quantity and identification conditions used for identification processing for identifying whether target data such as image data is data of a specific content (see, for example, Patent Document 1). . In addition, a technique for recognizing a subject in a moving image with high accuracy is known (for example, see Patent Document 2).
JP 2005-115525 A JP 2001-34756 A Jacky S. C. Yuk, Kwan-Yee K.K. Wong, Ronald H. Y. Chung, F.A. Y. L. Chin and K. P. Chow, "Real-time multiple head shape detection and tracking system with decently tracked". Proceedings of the ensembles in the world. 02, pp. 384-389, 2006 Shai Aviaden, Ensemble Tracking, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 29, issue. 2, pp. 261-271, Feb. , 2007 Helmut Grabner, Michael Grabner, “Real-time tracking via on-line boosting”, vol. 1, pp. 47-57, BMVC2006

  If reference data that defines the type of feature quantity and the identification condition is used, it is not easy to identify a subject whose feature quantity changes. In addition, when the feature amount changes greatly, it may be detected as a different subject and the specific object may not be tracked. Thus, there has been a problem that the object detection process and the specific object tracking process cannot be made efficient.

  In order to solve the above-described problem, in the first aspect of the present invention, an image analysis apparatus that extracts an object of a specific subject from a moving image, and stores a plurality of reference feature amounts in association with the orientation of the subject. A quantity storage unit, a feature quantity extraction unit that extracts a plurality of feature quantities from each of a plurality of movie constituent images included in the movie, and a plurality of reference feature quantities that are stored in association with the orientation of the subject. An object extraction unit that extracts an object of a specific subject from a moving image composition image based on the degree of matching of a plurality of feature amounts, a direction specification unit that specifies the orientation of the specific subject, and a reference feature stored in the feature amount storage unit A reference feature amount calculation unit that calculates a reference feature amount adapted to the extracted feature amount of the object, and a reference feature stored in the feature amount storage unit in association with the specified direction. Comprising an amount, the feature amount updating unit which reference feature amount calculating unit updates the reference feature quantity calculated.

  The reference feature amount calculation unit may calculate a plurality of reference feature amounts that are adapted to the feature amount of the extracted object with higher strength as the degree of matching is larger.

  The reference feature quantity calculation unit may calculate a plurality of reference feature quantities adapted to the extracted feature quantity of the object at a frequency corresponding to the type of feature quantity.

  The feature amount extraction unit extracts a plurality of feature amounts from images with different resolutions, and the feature amount storage unit stores a plurality of reference feature amounts corresponding to the plurality of feature amounts extracted from images with different resolutions. The reference feature amount calculation unit may calculate a plurality of reference feature amounts adapted to the extracted feature amount of the object with an intensity corresponding to the resolution.

  The reference feature quantity calculation unit may calculate a plurality of reference feature quantities adapted to the extracted feature quantity of the object with greater intensity as the resolution is higher.

  The reference feature amount calculation unit may calculate a plurality of reference feature amounts corresponding to a resolution larger than a predetermined value, adapted to the extracted feature amount of the object.

  The reference feature amount calculation unit is not updated with one or more types of reference feature amounts that meet a predetermined condition used when the object extraction unit extracts an object from the moving image composition image as an adapted reference feature amount. You may have a feature-value selection part.

  The non-updated feature quantity selection unit is a standard obtained by adapting one or more types of reference feature quantities having a matching degree equal to or greater than a predetermined value used when the object extraction unit extracts an object from the moving image composition image. It may be a feature amount.

  The reference feature value calculation unit may calculate a reference feature value obtained by learning an image of a region including the extracted object as a correct image as an adapted reference feature value.

  The reference feature quantity calculation unit may calculate a reference feature quantity obtained by learning an area image other than the extracted object as an incorrect image, as an adapted reference feature quantity.

  The feature amount storage unit stores a plurality of reference feature amounts corresponding to a plurality of feature amounts extracted from images of different resolutions in association with the direction of the subject, and the direction specifying unit has a predetermined resolution. By comparing the feature quantity extracted by the feature quantity extraction unit from an image with a small resolution with the reference feature quantity with the same resolution, the fitness is calculated for each orientation of the subject, and the criterion with a higher fitness is calculated. The direction of the subject associated with the feature amount may be specified as the direction of the specific subject.

  According to a second aspect of the present invention, there is provided an image analysis method for extracting an object of a specific subject from a moving image, a feature amount storing step for storing a plurality of reference feature amounts in association with the direction of the subject, A feature amount extraction stage for extracting a plurality of feature amounts from each of the plurality of moving image constituent images, and a degree of fitness of the extracted feature amounts with respect to a plurality of reference feature amounts stored in association with the orientation of the subject. Based on the object extraction stage for extracting the object of the specific subject from the moving image composition image, the orientation specifying stage for specifying the orientation of the specific subject, and the reference feature quantity stored in the feature quantity storage stage, the feature of the extracted object A reference feature amount calculation step for calculating a reference feature amount adapted to the amount, and a reference feature amount stored in association with the orientation specified in the feature amount storage step , And a characteristic value update step of updating the reference feature amount calculated in the reference feature value calculation step.

  According to a third aspect of the present invention, there is provided a program for an image analysis apparatus for extracting an object of a specific subject from a moving image, wherein the computer stores a plurality of reference feature amounts in association with the direction of the subject. A feature amount extraction unit that extracts a plurality of feature amounts from each of a plurality of movie constituent images included in the movie, and a plurality of feature amounts extracted with respect to a plurality of reference feature amounts stored in association with the orientation of the subject Extracted based on the reference feature value stored in the object extraction unit that extracts the object of the specific subject from the moving image composition image, the direction specification unit that specifies the direction of the specific subject, and the feature storage unit A reference feature quantity calculation unit that calculates a reference feature quantity adapted to the feature quantity of the object, and a reference feature quantity stored in association with the specified direction by the feature quantity storage unit Reference feature value calculating part to function as a characteristic value update section for updating the reference feature quantity calculated.

  The above summary of the invention does not enumerate all necessary features of the present invention. In addition, a sub-combination of these feature groups can also be an invention.

  Hereinafter, the present invention will be described through embodiments of the invention, but the following embodiments do not limit the invention according to the claims. In addition, not all the combinations of features described in the embodiments are essential for the solving means of the invention.

  FIG. 1 shows an example of an image processing system 10 according to an embodiment. As described below, the image processing system 10 can function as an image processing system for a store in one embodiment.

  The image processing system 10 includes a plurality of imaging devices 100a-b that images the store space 160, a store server 170, a communication network 150, an image analysis device 110, an analysis result database 112, and a terminal device 180. In the following description, the imaging device 100a and the imaging device 100b may be collectively referred to as the imaging device 100. Similarly, in the following description, a character indicated by a numerical code may be generically named by omitting a character following the numerical code, such as an English letter at the end.

  The imaging device 100a images the store space 160. The imaging device 100a images the person 164a and the person 164b in the store space 160 together with the product shelf 162a, the product shelf 162b, and the like in the store space 160, and generates a moving image. The imaging device 100 a supplies the moving image obtained by imaging the store space 160 to the store server 170. Note that the imaging device 100b is provided at a different position from the imaging device 100a. In other respects, the function and operation of the imaging device 100b are substantially the same as the function and operation of the imaging device 100a, and thus the description of the imaging device 100 is omitted.

  The store server 170 sends the moving images supplied from the imaging device 100 a and the imaging device 100 b to the communication network 150 toward the image analysis device 110. The communication network 150 can be exemplified by a telecommunication line such as the Internet. The store server 170 is provided in the vicinity of the imaging device 100, for example. In another form, the store server 170 may be provided in the store space 160.

  Further, the store server 170 controls the imaging operation of the imaging device 100. For example, the store server 170 controls on / off of the imaging function of the imaging device 100, the imaging rate, and the like. When the imaging device 100 can perform zoom shooting, the store server 170 may control the zoom value of the imaging device 100. Further, when the imaging direction of the imaging device 100 is variable, the store server 170 may control the imaging direction of the imaging device 100.

  The image analysis device 110 is provided in a space different from the store space 160, and acquires the moving images captured by the imaging device 100 from the store server 170 through the communication network 150. The image analysis device 110 analyzes the moving image and sends the analysis result to the communication network 150 toward the terminal device 180.

  Note that the image analysis device 110 may output the analysis result to the analysis result database 112. The image analysis device 110 may read the analysis result from the analysis result database 112 and transmit the read analysis result to the terminal device 180. The analysis result database 112 may be included in the image analysis device 110. In addition, the analysis result database 112 may be provided in a space different from the space in which the image analysis device 110 is provided. In this case, the image analysis apparatus 110 may transmit the analysis result to the analysis result database 112 via the communication network 150.

  The terminal device 180 acquires the analysis result from the image analysis device 110 from the image analysis device 110 through the communication network 150. The terminal device 180 processes and displays the analysis result according to a user instruction. The terminal device 180 may be provided in a space different from the space in which the image analysis device 110 is provided. The terminal device 180 may be provided in the vicinity of the store space 160 or in the store space 160.

  Below, an example of operation | movement of the image analysis apparatus 110 is demonstrated roughly. The image analysis apparatus 110 detects an object in which a moving body such as a person 164 is captured from each of a plurality of frame images included in a moving image captured by the imaging apparatus 100, and tracks an object of the same subject in the image. . At this time, the image analysis apparatus 110 detects and tracks an object using a plurality of weak classifiers.

  Specifically, the image analysis apparatus 110 calculates a discrimination result from a plurality of weak classifiers that calculate a discrimination result from a low resolution image, a plurality of weak classifiers that calculate a discrimination result from a medium resolution image, and a high resolution image. It has a plurality of weak classifiers. These weak classifiers calculate a discrimination result based on the feature amount extracted from the image.

  The weak classifier for a low resolution image is a weak classifier for extracting an object of a specific type of subject such as a person, and various types of feature amounts are extracted from the low resolution image to calculate a discrimination result. Thereby, a human object can be detected with high robustness. Note that the weak classifier for low-resolution images is used without being adapted to the object to be tracked. The weak classifier for low-resolution images is constructed so that it can detect a human object with high robustness by, for example, prior learning.

  Similarly, the weak classifier for the medium resolution image extracts various types of feature amounts and calculates the identification result. Weak classifiers for medium-resolution images are constructed with multiple weak classifiers for extracting objects of a specific type of subject, such as a person, and multiple weak classifiers that are adapted to tracked objects Is done. A weak classifier for extracting an object of a specific type of subject is constructed by prior learning or the like, similarly to the weak classifier for low-resolution images, and is used without being adapted to the object to be tracked. Of the weak classifiers for medium resolution images, weak classifiers that are adapted to the person to be tracked are learned for each frame, for example, by online learning.

  The weak classifier for high-resolution images also extracts various types of feature amounts and calculates the identification result. The weak classifier for high-resolution images is adapted to the tracking target object for each frame by online learning.

  When a weak classifier is constructed by prior learning, a reinforcement learning method using a teacher image can be used. Further, the weak classifier group may be constructed by group learning using boosting such as Adaboost. As online learning, an online boosting technique can be used.

  The image analysis apparatus 110 extracts an object using a weak classifier having a different combination for each frame image, or extracts an object using a weak classifier obtained by adapting a weak classifier having a different combination for each frame image. Or extract. As described above, the image analysis apparatus 110 extracts an object using a classifier that combines a weak classifier that robustly detects a specific type of object and a tracking process that can be adapted to a subject. Therefore, the image analysis apparatus 110 can robustly detect a specific type of object and can track it with high accuracy.

  In the present embodiment, “extracting an object” refers to detecting and tracking a specific object from an image region. Further, the image analysis apparatus 110 may output position information such as coordinate information for identifying an area where the extracted object exists as an analysis result. In addition, the image analysis apparatus 110 may output an object image obtained by cutting out a region where an object exists.

  As described above, the image analysis apparatus 110 searches globally in the frame image using the class identifier in order to find a new object. Then, the image analysis apparatus 110 uses the observation information of both the previous frame image and the current frame image of the tracking target object to perform each process of object position prediction, object identification, and weak classifier update. Line up and keep track of objects.

  As a classifier for detecting a specific type of object, a class classifier can be used. In this case, the class discriminator can be said to be a discriminator for discriminating species. On the other hand, an identifier for tracking an object is required to have the ability to both identify a species and identify each individual belonging to the species according to the environment of the object to be tracked.

  For example, when there is no other person around the person to be tracked, the identification ability required as a classifier for tracking may be relatively low. For example, it is only necessary to distinguish a non-human object such as a background from a human object. That is, it is only necessary to identify the seed level. However, in order to correctly acquire a flow line at a timing when a plurality of persons cross each other and at timings before and after the timing, the discrimination capability required for the tracking discriminator is relatively high. Specifically, the ability to distinguish a plurality of persons is required.

  As described above, according to the image analysis apparatus 110, the discriminator is flexibly adapted to the tracking object, so that the discriminating ability of the discriminator is automatically changed according to surrounding environmental changes (for example, background fluctuations, congestion fluctuations, etc.) It will be adjusted to. That is, the classifier included in the image analysis device 110 has a polymorphic feature. Accordingly, the classifier included in the image analysis apparatus 110 can be a class level classifier or an individual object classifier depending on the situation.

  As described above, according to the image analysis apparatus 110, the discriminator is adjusted in accordance with the change in the environment, so that the flexibility is extremely high. Further, according to the image analysis apparatus 110, the classifier is constructed according to the tracking of the object, so that the already detected object can be extracted without applying the classifier in the initial state. As described above, according to the image analysis apparatus 110, since the detection process and the tracking process are not independent, it is possible to cope with a severe change in the environment.

  Note that the storage medium 90 stores a program for the image analysis apparatus 110. The program stored in the storage medium 90 is provided to an electronic information processing apparatus such as a computer that functions as the image analysis apparatus 110 according to the present embodiment. The CPU included in the computer operates according to the contents of the program and controls each unit of the computer. The program executed by the CPU causes the computer to function as the image analysis apparatus 110 described with reference to this figure and subsequent figures.

  Examples of the storage medium 90 include an optical recording medium such as a DVD or PD, a magneto-optical recording medium such as an MO or MD, a magnetic recording medium such as a tape medium or a hard disk device, a semiconductor memory, a magnetic memory, etc. It can be illustrated. Further, a storage device such as a hard disk or a RAM provided in a server system connected to a dedicated communication network or the Internet can function as the storage medium 90.

  The storage medium 90 may store a program for the image processing system 10. The program for the image processing system 10 includes at least one of the analysis result database 112, the store server 170, the terminal device 180, and the imaging device 100 described with reference to this figure and the following drawings, in addition to the image analysis device 110. As such, the computer can function.

  FIG. 2 shows an example of a block configuration of the image analysis apparatus 110. The image analysis apparatus 110 includes an image acquisition unit 200, a feature amount extraction unit 210, a fitness calculation unit 220, an object extraction unit 230, a reference feature amount calculation unit 240, an object region estimation unit 250, a direction identification unit 260, and a feature amount update unit. 270, a feature amount storage unit 280, and an output unit 290. The reference feature value calculation unit 240 includes an update feature value selection unit 242, an update feature value calculation unit 244, and a non-update feature value selection unit 246.

  The image acquisition unit 200 acquires the moving image captured by the imaging apparatus 100 from the store server 170 through the communication network 150. The moving image acquired by the image acquisition unit 200 is supplied to the feature amount extraction unit 210 and the object extraction unit 230.

  The feature amount extraction unit 210 extracts a feature amount from the frame image included in the moving image. Examples of the feature quantity include a brightness feature quantity such as a brightness distribution, various wavelet feature quantities, Haar-like feature quantities, Edgelet feature quantities, EOH feature quantities, HOG feature quantities, and the like.

  The fitness level calculation unit 220 outputs the fitness level of the feature value extracted by the feature value extraction unit 210 with respect to the reference feature value. Specifically, the feature quantity storage unit 280 stores the reference feature quantity, and the matching degree calculation unit 220 outputs the matching degree with respect to the reference feature quantity stored in the feature quantity storage unit 280.

  Note that the feature quantity extraction unit 210 extracts a feature quantity for each type of different feature quantities. The feature amount storage unit 280 stores reference feature amounts for different types. The fitness level calculation unit 220 calculates the fitness level for each type.

  The different types of feature quantities may be feature quantities having different types of feature quantities such as Haar-like feature quantities and Edgelet feature quantities as described above. In addition, the different types of feature amounts can include feature amounts calculated by different calculation methods. Different calculation methods include extracting from images of different resolutions. That is, even if the types of feature quantities themselves are the same, feature quantities extracted from images with different resolutions can be regarded as different types of feature quantities in this embodiment.

  The object extraction unit 230 determines whether the target image is an image of a specific object based on the fitness for each type calculated by the fitness calculation unit 220. That is, the object extraction unit 230 can perform extraction.

  Note that the feature quantity extraction unit 210, the fitness level calculation unit 220, the feature quantity storage unit 280, and the object extraction unit 230 may be implemented as a process of a classifier. For example, focusing on one type of feature quantity, the process from when the feature quantity extraction unit 210 extracts a feature quantity until the fitness level calculation unit 220 outputs the fitness level is referred to as processing by one weak classifier. Can be considered. Including the processing of the object extraction unit 230 that determines whether or not the image is an object image based on the fitness for a plurality of types, it can be regarded as a process of one classifier for extracting an object.

  If the object extraction unit 230 determines that the image is an object image, the object peripheral image, which is an image of a predetermined area including the object, is cut out from the frame image and supplied to the reference feature value calculation unit 240. The reference feature amount calculation unit 240 uses the image supplied from the object extraction unit 230 to calculate a reference feature amount for extracting an object from the next frame image. The reference feature amount calculated by the reference feature amount calculation unit 240 is stored in the feature amount storage unit 280 by the feature amount update unit 270. As a result, the reference feature value stored in the feature value storage unit 280 is updated with the reference feature value calculated by the reference feature value calculation unit 240.

  Then, an object is extracted from the next frame image by the processes in the feature amount extraction unit 210, the fitness calculation unit 220, and the object extraction unit 230 described above. By repeating this operation, an object is extracted from each of the plurality of frame images.

  Note that the reference feature amount calculation unit 240 calculates a reference feature amount adapted to the object peripheral image by learning using the object peripheral image. As described above, the reference feature value calculation unit 240 can calculate a feature value adapted to the object image of the current frame. Therefore, the reference feature value calculation unit 240 can calculate a reference feature value suitable for tracking a specific object.

  Note that the reference feature value adapted to the object peripheral image is calculated by the reference feature value calculation unit 240. Updating one reference feature value means that the processing of the weak classifier described above is adapted. To do. That is, the feature amount update processing from when the reference feature amount is calculated by the reference feature amount calculation unit 240 until it is stored in the feature amount storage unit 280 can be regarded as processing for constructing an adaptive weak classifier. When the reference feature value calculation process in the reference feature value calculation unit 240 is based on learning, the feature value update process can be regarded as the online learning process of the classifier described above.

  Note that the reference feature value calculation unit 240 can update the reference feature value according to the orientation of the subject imaged as the extracted object. For example, the reference feature amount calculation unit 240 may store a reference feature amount for the feature amount storage unit 280 to extract an object for each direction of the subject. In this case, the reference feature amount calculation unit 240 can select and update the reference feature amount corresponding to the orientation of the subject imaged as the extracted object.

  Note that the orientation specifying unit 260 specifies the orientation of a subject imaged as an object. When the feature amount storage unit 280 stores the reference feature amount for each orientation, the orientation specifying unit 260 determines the orientation stored in the feature amount storage unit 280 in association with the most suitable reference feature amount, for the subject. Can be oriented. The reference feature value calculation unit 240 can perform the above-described reference feature value update process based on the direction specified by the direction specification unit 260.

  In addition, the reference feature value calculation unit 240 may calculate the reference feature value based on the fitness level calculated by the fitness level calculation unit 220. For example, as a reference feature value for extracting an object from the next frame image, a reference feature value having a higher fitness can be incorporated by excluding a reference feature value having a lower fitness. In other words, the combination of weak classifiers can be changed. For this reason, it is possible to improve the accuracy of tracking processing for tracking an object of a specific subject.

  Further, the reference feature value calculation unit 240 can calculate the reference feature value based on the position of the object in each frame. For example, the reference feature amount calculation unit 240 learns by using an object image in the current frame as a correct image, or learns by using an image in an area other than the object area in the current frame as an incorrect image. The reference feature amount can be calculated.

  The object area estimation unit 250 estimates the position of the object on the next frame image. Specifically, the position of the object on the next frame image can be estimated based on the moving speed of the object in the moving image and the position of the object in the current frame image. The object region estimation unit 250 may cut out an image of a region including the estimated position from the current frame image and supply the image to the reference feature value calculation unit 240.

  In this case, the reference feature value calculation unit 240 can calculate the reference feature value by learning the image supplied from the object region estimation unit 250 as an incorrect image. As a result, the reference feature value set stored in the feature value storage unit 280 is strengthened so that the image around the object in the next frame image is difficult to detect, so the detection probability of the object from the next frame image Can be increased.

  Here, the operation of each unit in the reference feature quantity calculation unit 240 will be briefly described. The update feature quantity selection unit 242 specifies the type to be updated. For example, as the processing by the update feature quantity selection unit 242, a new reference feature quantity can be introduced by excluding the kind of reference feature quantity having a low fitness as described above. Then, the update feature quantity calculation unit 244 adapts the type of feature quantity selected by the update feature quantity selection unit 242 by learning or the like. Further, the non-update feature quantity selection unit 246 may select a feature quantity of a type used for object extraction even in the next frame image without adaptation. The operation of each unit in the reference feature quantity calculation unit 240 will be described in detail later.

  Further, the output unit 290 outputs information on the object extracted by the object extraction unit 230 to the outside. For example, the output unit 290 may output the position of the extracted object and the image of the extracted object itself to the analysis result database 112, the terminal device 180, and the like. As described above, according to the image analysis device 110, the object detection process and the object tracking process can be performed integrally, and the amount of calculation for tracking the object can be reduced.

  FIG. 3 schematically shows an example of processing contents in the image analysis apparatus 110. Hereinafter, functions and operations of the image analysis apparatus 110 will be described by taking processing by a classifier or weak classifier as an example.

  The classifier set 404 includes a plurality of classifiers 402a to 402c that can be used by the image analysis apparatus 110 for object extraction processing. Each of the discriminators 402a to 402c corresponds to a different direction of the subject. Specifically, the classifier 402a is formed by a plurality of weak classifiers 400a to 400c suitable for detecting the front face of a person.

  Here, the weak classifier 400a is formed by one or more weak classifiers 400 for low-resolution images. For example, the weak classifier 400a is formed by one or more weak classifiers 400 for an image having a resolution of 8 × 8 pixels. In the following drawings, one weak classifier 400 is indicated by one horizontally long rectangle. The weak classifier 400b is formed by one or more weak classifiers 400 for the medium resolution image, and the weak classifier 400c is formed by one or more weak classifiers 400 for the high resolution image. For example, the weak classifier 400b is formed by one or more weak classifiers 400 for an image having a resolution of 16 × 16 pixels, and the weak classifier 400c is one or more weak classifiers 400 for an image having a resolution of 32 × 32 pixels. It is formed by. A weak classifier having a relatively low resolution such as the weak classifier 400a functions as a classifier part for detecting a species, and a weak class having a relatively high resolution such as the weak classifier 400c (or the weak classifier 400b). The classifier can function as a classifier part that identifies individual objects. As will be described later, the weak classifier 400 having a relatively high resolution is adapted to an individual object by online learning. In this way, by using a hierarchical weak classifier configuration corresponding to the resolution, individual objects can be detected and tracked at high speed.

  The classifier 402b is formed by a plurality of weak classifiers 400 suitable for detecting a right profile of a person, and the classifier 402c is suitable for detecting a profile of a person's left profile. A plurality of weak classifiers 400 are formed. Except for this point, the discriminator 402b and the discriminator 402c have substantially the same configuration as the discriminator 402a, and thus description thereof is omitted.

  Hereinafter, a process in which the image analysis apparatus 110 generates the classifier set 464 used for object extraction from the classifier set 404 will be described with reference to FIG. Here, for convenience of explanation, it is assumed that the object extraction unit 230 has already extracted a specific object from the previous frame image in the process 480. The object extraction unit 230 supplies information indicating the object region to the reference feature amount calculation unit 240.

  At the stage where the object is extracted, the fitness level calculation unit 220 calculates the fitness level using the classifiers included in the classifier set 464. The fitness calculated by the fitness calculation unit 220 is supplied to the reference feature value calculation unit 240. At the stage where the object is extracted, the orientation specifying unit 260 specifies the orientation of the subject indicated by the object, and the specified orientation is also supplied to the reference feature value calculating unit 240. At the stage where the object is extracted, the object region estimation unit 250 estimates the region where the object exists in the next frame image, and information indicating the estimated region is supplied to the reference feature value calculation unit 240. Yes.

  The update feature amount calculation unit 244 generates an object peripheral image 412 that is an image around the object based on the information indicating the object region (processing 410). The object peripheral image here includes an object region and a region around the object. Then, the update feature quantity calculation unit 244 generates a plurality of teacher images 422 having a predetermined number of pixels using the object peripheral image 412 (processing 420).

  The teacher image 422 includes a correct image and an incorrect image. The update feature amount calculation unit 244 generates a plurality of correct images from the image of the object area in the object peripheral image 412. In addition, the update feature amount calculation unit 244 generates a plurality of incorrect images from an image of a region including a region other than the object in the object peripheral image 412.

  In process 430, the non-updated feature quantity selection unit 246 selects a classifier used by the object extraction unit 230 to extract an object in the next frame image from the weak classifiers included in the classifier 402. The discriminator selected in the process 430 is used for object extraction without being applied to the object peripheral image.

  In the following description, it is assumed that the object of the front face of the person is extracted by the orientation specifying unit 260. In this case, the weak classifier 400 selected in the process 430 and the weak classifier 400 adapted in the process 450 are both selected from the weak classifiers 400 forming the classifier 402a. Of course, when the right side profile object is extracted, it is selected from the classifier 402b, and when the left side profile object is extracted, it is needless to say that the classifier 402c is selected.

  At this time, the weak classifier 400 having a high degree of fitness for the object may be selected. A predetermined number of weak classifiers 400 may be selected from the weak classifiers 400a for low-resolution images. In addition to the weak classifiers 400a for low resolution images, a predetermined number of weak classifiers 400 may be selected from the weak classifiers 400a for medium resolution images.

  In this way, all weak classifiers that can function as weak classifiers for species detection are replicated (for example, all weak classifiers 400a are replicated) or partially replicated (for example, one of the weak classifiers 400b). A weak classifier for detecting an individual object is formed. In other words, a part of the weak classifier for detection and tracking is formed by complete inheritance and partial inheritance of the weak classifier for species detection. Thereby, it is determined that the object detected by the weak classifier that is completely inherited and partially inherited belongs to the species detected by the weak classifier. These inherited weak classifiers may be constructed in advance by offline learning.

  On the other hand, in order to identify individual objects, a weak classifier having the ability to detect the characteristics of each individual (for example, characteristics of eyes, mouth, etc.) is generated by a part of the weak classifier 400b and the weak classifier 400c. . Specifically, a part of the weak classifier 400b and the weak classifier 400c are evolved by online learning (for example, the AdaBoost method) to reflect the individual characteristics. The advanced weak classifier 400c incorporates the ability to identify each of a plurality of individual objects being tracked (for example, individual objects belonging to the same species) through online learning.

  In this way, the weak classifier is inherited and evolved according to the hierarchical structure corresponding to the resolution. The process of classifier evolution includes the process of selecting, excluding, and updating weak classifiers. In particular, in the update process, new features of individual objects are incorporated into the classifier by online machine learning techniques. Accordingly, since the weak classifier for detecting the species and the weak classifier for detecting the individual object are linked, the individual object can be detected and tracked at high speed and with high accuracy.

  This evolution process will be described using the functional blocks of the image analysis apparatus 110. The weak classifier 400 to be adapted to the object peripheral image is selected by the process of the update feature amount selection unit 242. Specifically, in process 440, each weak classifier 400 included in the classifier set 404 performs determination processing on the object peripheral image, and the identification result 442 is calculated. In a process 450, a classifier set 464 is generated based on the identification result 442.

  For example, the weak classifier 400 to be adapted is selected by the processing by the update feature value selection unit 242. The selected classifier is applied to the object peripheral image and used for object extraction. At this time, the weak classifier 400 having a lower fitness to the object may be preferentially selected as the weak classifier to be adapted. A predetermined number of weak classifiers 400 may be selected from the weak classifiers 400b for medium resolution images. In addition to the weak classifiers 400 from the weak classifiers 400b, a predetermined number of weak classifiers 400 may be selected from the weak classifiers 400c for high resolution images.

  Here, the weak classifier 400 selected by the update feature quantity selection unit 242 is applied to the object peripheral image 412 by the processing by the update feature quantity calculation unit 244. Here, the weak classifier 400 may be adapted by an online boosting method using the teacher image 422. Then, an adapted weak classifier 460 is generated by the processing by the feature amount updating unit 270.

  A classifier set 464 is constructed by the weak classifier 460 adapted in the process 450 and the weak classifier 400 selected in the process 430. Although the weak classifier 400 selected in the process 430 is not adapted, it is a weak classifier used for extracting an object from the next frame image. Yes.

  As described above, each of the weak classifiers 460a is a weak classifier that performs the same classification process as the weak classifier 400a. A predetermined number of weak classifiers 460 among the weak classifiers 460a are used for object extraction processing from the next frame image.

  Also, some of the weak classifiers 460 of the weak classifiers 460b may be weak classifiers that perform the same classification process as the corresponding weak classifiers 400 of the weak classifiers 400b. Some of these weak classifiers 460 are used for object extraction processing from the next frame image. The other weak classifiers 460 may be weak classifiers to which the corresponding weak classifiers 400 of the weak classifiers 400 b are adapted in the process 450. The adapted weak classifier is used for object extraction processing from the next frame image.

  The weak classifier 460 selected as the weak classifier to be updated among the weak classifiers 460c is a weak classifier to which the corresponding weak classifier 400 of the weak classifiers 400c is adapted in the process 450, These weak classifiers are used for object extraction processing.

  In the process 480, the object extraction unit 230 extracts an object from the next frame image using the weak classifier 460 to be used for the object extraction process in the classifier set 464. Processing 470 indicates processing in which the object region estimation unit 250 estimates a region where an object exists on the next frame image. In the process 480, the object extraction unit 230 uses the weak classifier included in the classifier set 464 for the estimated area image 472 that is the image of the area estimated in the process 470 to perform object search / recognition processing. do.

  As described above, according to the image analysis apparatus 110, when constructing a classifier that detects and tracks an individual object, a part is inherited from the parent weak classifier. And the ability to detect the feature of an individual object is built into some other weak classifiers by an online learning method. By performing the inheritance and evolution processing in this way, it is possible to provide a higher discrimination ability than when performing online learning from the weak classifier in the initial state while maintaining the species detection ability.

  As described above, the classifier included in the image analysis apparatus 110 includes a weak classifier corresponding to a feature amount from a coarse image and a weak classifier corresponding to a feature amount from a fine image formed in a hierarchy. Further, the discriminator is formed in a form including a weak discriminator that can sufficiently represent the attributes of the species corresponding to the feature amount from the rough image. Therefore, the classifier included in the image analysis apparatus 110 can function as a classifier that detects the seed. Furthermore, according to the image analysis apparatus 110, the weak classifier that identifies the individual object to be tracked is evolved by online learning, so that the classifier can have the ability to detect the unique feature of the individual object. . Accordingly, it is possible to construct a discriminator that can detect and identify individual objects in order to track a plurality of objects in a video scene.

  According to the image analysis apparatus 110, detection / tracking processing is performed on an already detected object using the inherited and evolved weak classifier without using the weak classifier in the initial state. Therefore, the close dependency between the tracking process and the detection process can be used without separating the tracking process and the detection process, and the detection and tracking can be performed efficiently. For this reason, according to the image analysis apparatus 110, the adaptability of the classifier to the environment can be enhanced.

  FIG. 4 schematically shows an example of processing contents related to the processing 450. Assume that the classifier 502a is a classifier used for detection processing of a specific object. With reference to this figure, the process in the case of constructing a new discriminator used for the object extraction process for the next frame image will be described. Here, in particular, a description will be given centering on processing for constructing a weak classifier adapted to an object peripheral image. In addition, although the process with respect to the discrimination device 462a with respect to a front face is demonstrated below, it cannot be overemphasized that the process with respect to the discrimination device 462b and the discrimination device 462c is also the same.

  The classifier 502a indicates a classifier actually used for object extraction among the classifiers 462a. The classifier 502a is divided into a weak classifier 500a and a weak classifier 500b. The weak classifier 500a includes a weak classifier 460 used for object extraction in the weak classifier 460a and a weak classifier 460 used for object extraction in the weak classifier 460b and cannot be applied to the object peripheral image. Show. The weak classifier 500b is a weak classifier that can be applied to an object peripheral image.

  The weak classifier 500a includes a part of the high resolution weak classifier 460a and a part of the medium resolution weak classifier 460b. When a specific object is extracted, the object extraction unit 230 uses the weak classifier 500a to extract the specific object from one or more subsequent frame images. In addition, when the orientation of the face of the imaged person changes significantly, another discriminator 462 corresponding to the face orientation, such as the discriminator 462b, is used. Also in this case, after an object having a specific face orientation is extracted, a discriminator constructed for the specific face orientation from one or more subsequent frame images may be used.

  In this manner, a combination of low resolution and some medium resolution weak classifiers 500a selected from the weak classifiers 400a for human detection, which are suitable for extracting a specific object, include one or more subsequent classifiers. It is used for object extraction processing from the frame image. For this reason, it is possible to reduce the amount of calculation while reducing the probability of erroneous detection of other objects.

  Next, processing for the weak classifier 500b will be described. The weak classifier 500b is formed by P weak classifiers 500. The P weak classifiers 500 give the classifiers the ability to identify individual objects. In process 510, N weak discriminators to be excluded are selected by the process by the update feature value selection unit 242. Specifically, the N weak classifiers may be selected as the weak classifiers that are excluded in order from the one having the lowest degree of matching with the image of the object. As shown in the weak classifier 500b, the weak classifier 500b is divided into (PN) weak classifiers 520A and N weak classifiers 520B. When the weak classifier 520B is excluded in the exclusion process in the process 530, (PN) weak classifiers 520A remain.

  As described above, the classification performance of the weak classifiers corresponding to the P feature amounts is evaluated by the individual objects. Based on the evaluation results, the discrimination performance of the P weak discriminators is ranked, and N weak discriminators having lower discrimination performance are excluded.

  Next, a process for generating the weak classifier 562 used for the object extraction process from the next frame image from the classifier 402a will be described. In process 550, the candidate of M weak classifiers to be added is selected from the weak classifiers 400 included in the classifier 402a by the process by the update feature quantity calculation unit 244. In this case, M weak classifiers 400 may be selected in descending order of the degree of fitness of the object to the image. The selected weak classifier is shown as weak classifier 552.

  In this way, using the individual object in the current frame image as the evaluation data, the weak classifiers corresponding to the M feature quantities are selected from the weak classifiers 400 included in the classifier 402a. It should be noted that M feature values having a high evaluation value of discrimination performance are selected from the inherited weak classifier 400a or weak classifier 400b, and M weak classifications corresponding to the selected M feature quantities are selected. The classifier may be selected from the weak classifier 400c. Here, evaluation is performed including weak classifiers used to detect other kinds of objects, and M weak classifiers including at least weak classifiers used to detect other kinds of objects are selected. May be.

  In process 560, N weak classifiers are used as weak classifiers used for object extraction processing from the next frame image by the online boosting technique using the selected weak classifier 552 and the teacher image 422. A device 562 is determined.

  Specifically, in a specific weak classifier 552, feature amounts are extracted from each of the teacher images 422. By comparing the extracted feature value and the reference feature value, a determination value for determining whether or not the teacher image 422 is a correct image by the specific weak classifier 552 is determined for each teacher image 422. Further, an error value between the teacher signal indicating whether or not the teacher image 422 is a correct image and the determination value is determined for each teacher image 422. Then, the reference feature amount calculation unit 240 calculates a reference feature amount that minimizes a value obtained by weighting the error value determined for each of the teacher images 422. By using the calculated reference feature amount as a reference feature amount used for determination by the specific weak classifier 552, the specific weak classifier 552 can be adapted to the teacher image.

  Similarly, by calculating a reference feature amount that minimizes the weighted error value for each of a plurality of weak classifiers 552, each of the plurality of weak classifiers 552 can be adapted to the teacher image 422. The update feature quantity selection unit 242 selects the N weak classifiers 552 with smaller weighted errors as the weak classifiers 562 used for the object extraction process from the next frame image with higher priority. As described above, N effective weak classifiers 562 are selected from the M weak classifiers 552 by performing online learning using the peripheral image of the object to be extracted as a teacher image.

  By the processing by the update feature quantity calculation unit 244, (P−N) weak classifiers 520A and N weak classifiers 562 are combined to form P weak classifiers. A classifier that is actually used for object extraction processing from the next frame image is constructed by the formed weak classifier and the weak classifier 500a. Thereby, an effective classifier for detecting / tracking the individual object is formed. The object extraction unit 230 extracts an object from the next frame image using the constructed discriminator.

  Then, the classifier for extracting an object from a further next frame image is constructed by regarding the classifier constructed in the current frame image as the classifier 502 in this figure and performing the processing described in this figure. Is done. By repeating such processing, a specific object can be detected and tracked from a plurality of frame images included in the moving image.

  FIG. 5 shows an example of image areas of correct and incorrect images. An area 610 on the frame image 600 indicates an area where an object extracted from the frame image 600 exists. A region 620 indicates a region around the extracted object. The images of the areas 610 and 620 correspond to the object peripheral images described with reference to FIGS. 3 and 4.

  When generating the teacher image 422, the updated feature amount calculation unit 244 generates a plurality of correct images from the image corresponding to the region 610. In addition, the update feature quantity calculation unit 244 generates a plurality of incorrect images from the image corresponding to the region 620. The classifier can be adapted to the frame peripheral image by learning using a plurality of correct images and a plurality of incorrect images.

  FIG. 6 shows another example of the image areas of the correct image and the incorrect image. Assume that the frame image 700-1, the frame image 700-2, and the frame image 700-3 are continuous frame images captured in this order. The update feature amount calculation unit 244 generates a teacher image 422 for extracting an object from the frame image 700-3 on the frame image 700-1 and the frame image 700-2.

  An area 710b on the frame image 700-2 corresponds to the area 610 described in FIG. 5, and a plurality of correct images are generated from the area 710b. The area 710b may include a peripheral area of the extracted object (for example, an area that is two to three pixels wider than the extracted object area). The image including the peripheral region is also set as the image region of the correct image, so that it is possible to cope with the variation of the tracked object.

  An area 720a on the frame image 700-2 corresponds to the area 620 described with reference to FIG. 5, and a plurality of incorrect images are generated from the area 720a. The area 720a is an image area serving as a background around the object. The image area of the incorrect image may be set to an area where the tracked object is detected in the frame image 700-2.

  An area 710 a on the frame image 700-1 is an object area extracted from the frame image 700-1, and a plurality of correct images are generated from the area 710 a in the processing in the update feature amount calculation unit 244. Note that the object extracted from the area 710a on the frame image 700-1 and the object extracted from the area 710b on the frame image 700-2 are both objects of the same subject.

  An area 720b on the frame image 700-2 is an area from which another object is extracted from the frame image 700-2, and a plurality of incorrect images are generated from the area 720b in the process in the update feature amount calculation unit 244. Is done. An incorrect image may be generated from the area 720b of another object near the object being tracked. Thus, by generating an incorrect image from the area of another object, the probability of detecting the other object as a specific object can be reduced.

  An area 720c indicates an area where an object existing in the area 720a is predicted to exist in the next frame image 700-3. The region 720c is estimated by the object region estimation unit 250 based on the motion of the object. A plurality of incorrect images are generated from the region 720c on the frame image 700-2 by the processing in the update feature amount calculation unit 244.

  In the region 720c on the frame image 700-2, there is a high probability that a subject existing around a specific object in the frame image 700-3 is captured. Therefore, by setting the area 720c on the frame image 700-2 as an incorrect image, another object in the vicinity of the tracking target object is detected as the tracking target object from the frame image 700-3. The probability that it will end up can be reduced.

  Note that the object extraction unit 230 may extract an object from the region 720c of the frame image 700-3. That is, the extraction target region from which the object extraction unit 230 extracts an object and the target region for generating an incorrect image may be the same region on the image region. Note that the target area for generating the incorrect image may be smaller than the extraction target area.

  As described with reference to FIGS. 5 and 6, not only the object image in the current frame but also the image of the object in the previous frame is used as the correct image. In addition, the incorrect image is not only the background image of the peripheral area of the object, but also the background image near the position of the object at the timing of the next frame, and the image of another object positioned near the object to be tracked Is also used. Further, according to the image analysis apparatus 110, it is possible to incorporate not only learning in frame units but also online learning including object movement and object pattern change. For this reason, the capability of the discriminator to cope with environmental changes can be enhanced.

  The online learning process for generating a weak classifier for identifying an individual object does not have to be performed every frame. Tracking in moving images requires high speed and real-time performance. For the purpose of ensuring real-time performance, the frequency of online learning may be appropriately changed.

  For example, when the object is stationary or when the amount of movement (change amount) of the object is smaller than a predetermined value, online learning is not necessary. If the surrounding environment has not changed substantially or the object has not changed substantially, the object can be detected and tracked without online learning.

  In addition, when the similarity of the object obtained by identifying with the weak classifier for identifying the individual object is higher than a predetermined first value, online learning is not necessary. If the current classifier determines a sufficiently high degree of similarity, it means that the characteristics of the individual object are already built into the weak classifier, so that the individual object is sufficient even without online learning. Can be detected and tracked. In addition, online learning may not be performed when the similarity is lower than a predetermined second value (a value smaller than the first value). A remarkably low degree of similarity means that the object is likely to be occluded or misdetected. If online learning is performed in this case, it may become impossible to detect and track an object that is no longer occluded. For this reason, when the degree of similarity is extremely low, it can be said that it is desirable not to perform learning in view of the stability of the instance discriminator.

  As described above, according to the image analysis apparatus 110, it is possible to construct a discriminator that can perform the object detection process and the object recognition process in a lump. For this reason, the amount of calculation required for these processes can be remarkably reduced.

  In FIG. 4 to FIG. 6 and the like, the function and operation of the image analysis apparatus 110 have been described by taking processing by the classifier or weak classifier as an example. As described above, the functions and operations of the feature amount extraction unit 210, the fitness calculation unit 220, the feature amount storage unit 280, and the object extraction unit 230 can be regarded as functions and operations of a weak classifier. The adaptation and update of the reference feature amount can be regarded as the selection and adaptation of the discriminator. Therefore, each component of the image analysis device 110 operates as follows with respect to the reference feature amount, thereby realizing a function equivalent to the function specifically described in connection with the classifier in FIGS. 4 to 6 and the like. can do.

  That is, the image analysis apparatus 110 can extract a specific object from a moving image by the operation described below. Object extraction includes tracking by specifying a position for each object. The feature amount storage unit 280 stores a plurality of reference feature amounts of different types. The feature amount extraction unit 210 extracts a plurality of feature amounts of different types from each of a plurality of frame images included in the moving image. The object extraction unit 230 extracts an object from the frame image based on the degree of matching of the plurality of extracted feature amounts with respect to the plurality of reference feature amounts stored in the feature amount storage unit 280.

  Based on the reference feature quantity stored in the feature quantity storage unit 280, the reference feature quantity calculation unit 240 applies a plurality of reference feature quantities adapted to the extracted feature quantity of the object with a predetermined strength according to the type. calculate. Then, the feature amount update unit 270 updates the plurality of reference feature amounts stored in the feature amount storage unit 280 with the reference feature amounts calculated by the reference feature amount calculation unit 240.

  Note that the reference feature amount calculation unit 240 may calculate a plurality of reference feature amounts that are adapted to the feature amount of the extracted object with higher strength as the degree of matching is higher. Further, the reference feature value calculation unit 240 may calculate a plurality of reference feature values adapted to the extracted feature value of the object at a frequency according to the type.

  The feature amount storage unit 280 stores a plurality of reference feature amounts corresponding to a plurality of feature amounts extracted from images with different resolutions. Then, the feature amount extraction unit 210 extracts a plurality of feature amounts from images with different resolutions. The reference feature value calculation unit 240 calculates a plurality of reference feature values that are adapted to the extracted feature value of the object with an intensity corresponding to the resolution. The reference feature amount calculation unit 240 may calculate a plurality of reference feature amounts that are adapted to the extracted object feature amount with greater intensity as the resolution is higher. Note that the reference feature value calculation unit 240 may calculate a plurality of reference feature values corresponding to a resolution larger than a predetermined value, adapted to the extracted feature value of the object.

  Regarding the operation of the reference feature quantity calculation unit 240, the update feature quantity selection unit 242 gives priority to the type having a higher degree of matching among the types other than the type of the reference feature quantity selected by the non-update feature quantity selection unit 246. select. The update feature value calculation unit 244 calculates one or more reference feature values adapted to the feature value of the extracted object for the type selected by the update feature value selection unit 242. The update feature quantity calculation unit 244 selects, with higher priority, a reference feature quantity having a higher degree of adaptability to the extracted feature quantity of the extracted object from among the plurality of adapted reference feature quantities. To do. As the fitness, a learning error in the learning process can be exemplified.

  Note that the object extraction unit 230 may extract an object using at least P predetermined reference feature amounts that can be updated. Then, the update feature quantity calculation unit 244 may select the N types of reference feature quantities having higher fitness levels with higher priority as the adapted reference feature quantities. Then, the non-updated feature quantity selection unit 246 selects (PN) types of reference feature quantities whose matching level is equal to or greater than a predetermined value used when the object extraction unit 230 extracts an object from the frame image. The adapted reference feature value is used. In this case, N types of feature values selected in the order of fitness may be used as reference feature values.

  Then, the update feature quantity selection unit 242 selects M types of feature quantities having a higher degree of matching and greater than N. At this time, M types of feature quantities may be selected in descending order of fitness. The reference feature value calculation unit 240 calculates a reference feature value adapted to the feature value of the extracted object for the M types selected by the update feature value selection unit 242. Then, the reference feature quantity calculation unit 240 may preferentially select N types of reference feature quantities having a higher fitness among the calculated M types of reference feature quantities as the adapted reference feature quantities.

  The non-updated feature quantity selection unit 246 is a reference feature quantity obtained by adapting one or more types of reference feature quantities that meet a predetermined condition used when the object extraction unit 230 extracts an object from a frame image. As good as That is, one or more types of reference feature values that meet a predetermined condition are not updated and are used by the object extraction unit 230 for object extraction processing in the next frame image.

  When selecting a feature quantity that is not to be updated, the non-update feature quantity selection unit 246 has one or more matching degrees that are used when the object extraction unit 230 extracts an object from the frame image and that has a predetermined value or more. These types of reference feature values may be used as the adapted reference feature values. Further, the non-updated feature quantity selection unit 246 may select the adapted reference feature quantity based on the teacher image 422. Specifically, the non-updated feature quantity selection unit 246 selects a reference feature quantity with a combination that has a probability that the correct answer can be determined with respect to the correct image by learning based on the teacher image 422 higher than a predetermined value. You can do it. Further, the non-updated feature quantity selection unit 246 further selects a reference feature quantity having a combination in which the probability that an incorrect answer can be determined for an incorrect answer image is higher than a predetermined value by learning based on the teacher image 422. You may choose.

  In the process of adapting the feature amount, the reference feature amount calculation unit 240 calculates the reference feature amount obtained by learning the image of the region including the extracted object as the correct image as the adapted reference feature amount. It's okay. In addition, the reference feature value calculation unit 240 may calculate a reference feature value obtained by learning an image of a region other than the extracted object as an incorrect image as an adapted reference feature value.

  Note that the reference feature value calculation unit 240 may calculate the reference feature value by using an image of a region including a specific object extracted from the frame image as a correct image. The reference feature amount calculation unit 240 calculates the reference feature amount by setting an image of a region including a specific object extracted from the frame image before the frame image as a correct image. Here, the object extraction unit 230 may extract a specific object from the next frame image based on the degree of matching of the feature amount extracted from the next frame image by the feature amount extraction unit 210 with respect to the reference feature amount.

  The reference feature quantity calculation unit 240 may calculate the reference feature quantity by setting an image of a region that does not include the extracted specific object in the frame image as an incorrect image. Further, the reference feature value calculation unit 240 may calculate the reference feature value by setting an image of a region that is a predetermined distance away from a specific object extracted from the frame image as an incorrect answer image. Further, the reference feature value calculation unit 240 may calculate the reference feature value by using an image of a region including another object extracted from the frame image as an incorrect image.

  The object region estimation unit 250 estimates a region where a specific object should exist in another frame image based on the movement amount of the specific object extracted from the moving image. The reference feature amount calculation unit 240 may calculate the reference feature amount by using an image including the region estimated by the object region estimation unit 250 in another frame image as an incorrect image.

  The degree-of-fit calculation unit 220 calculates the degree of match of the feature amount extracted by the feature amount extraction unit 210 with respect to the reference feature amount for each type. Then, the object extraction unit 230 extracts an object from each of the plurality of frame images based on the degree of matching calculated for each type. In this case, the object extraction unit 230 may extract a specific object when the number of feature amounts for which a fitness level greater than a predetermined fitness level is calculated is greater than a predetermined value. In addition, the object extraction unit 230 may extract a specific object when the total value obtained by weighting the fitness with a predetermined weight is larger than a predetermined value.

  Note that the feature amount storage unit 280 stores a plurality of reference feature amounts in association with the direction of the subject. The feature amount extraction unit 210 extracts a plurality of feature amounts from each of a plurality of frame images included in the moving image. The object extraction unit 230 extracts the object of the specific subject from the frame image based on the degree of matching of the plurality of feature amounts extracted with respect to the plurality of reference feature amounts stored in association with the direction of the subject.

  The direction specifying unit 260 specifies the direction of the specific subject. At this time, the orientation may be specified based on the degree of matching with the reference feature amount. For example, the orientation specifying unit 260 associates the feature amount storage unit 280 with a reference feature amount in which a degree of fitness greater than a predetermined value is calculated among a plurality of feature amounts stored in the feature amount storage unit 280. May be specified as the orientation of the specific subject.

  As described above, the orientation specifying unit 260 compares the feature amount extracted by the feature amount extraction unit 210 from an image having a resolution smaller than a predetermined resolution with the reference feature amount having the same resolution, thereby determining the orientation of the subject. A goodness of fit may be calculated for each. Then, the orientation specifying unit 260 may specify the orientation of the subject associated with the reference feature amount for which a greater matching degree is calculated as the orientation of the specific subject.

  Then, the reference feature value calculation unit 240 calculates a reference feature value adapted to the extracted object from the reference feature values stored in the feature value storage unit 280. The feature amount update unit 270 may update the reference feature amount stored in association with the specified direction in the feature amount storage unit 280 with the reference feature amount calculated by the reference feature amount calculation unit 240. Thereby, it is possible to strengthen the reference feature amount of an appropriate combination according to the direction of the subject by learning.

  As described above, in the image analysis apparatus 110, a detector that detects a species by introducing the concept of inheritance and evolution into a class discriminator using the close relationship between detection processing and tracking processing. A class classifier for object tracking is constructed from the class classifier. Through online learning that realizes evolution, the ability to respond to environmental changes can be dynamically incorporated into the classifier. For this reason, according to the discriminator realized by the image analysis apparatus 110, it is possible to discriminate between persons and cope with environmental fluctuations.

  For example, according to the image analysis apparatus 110, a change in the shape of the object due to a change in the posture of the subject, a change in the environment in which the subject exists, including a lighting change and a background change, a change in imaging performance (for example, a change in contrast, It is possible to appropriately detect and track the fluctuation in resolution performance such as blurring. In addition, even in an environment where occlusion may occur, such as objects overlapping each other or being shielded by surrounding objects, detection and tracking can be performed appropriately. Further, even when the imaging apparatus 100 itself moves, it can be appropriately detected and tracked. Also, when tracking a plurality of objects, each object can be appropriately distinguished. In addition, since the individual object can be easily detected, the cooperation and cooperation between the plurality of imaging devices 100 is facilitated. Further, even when the regularity of the movement of the object is relatively low or when the information for estimating the movement of the object is insufficient, the detection and tracking can be appropriately performed. Even when video noise occurs, it can be detected and tracked appropriately. Thus, according to the image analysis apparatus 110, it is possible to extract individual objects stably and with high accuracy against such environmental fluctuations.

  As described above with reference to FIGS. 1 to 6, according to the image processing system 10, a specific customer can be accurately tracked, so that customer information can be accurately analyzed or the customer is interested. You can accurately analyze the products. Thus, the image processing system 10 can function as an image processing system for a store.

  In addition, the image processing system 10 can be applied to various uses. For example, according to the image processing system 10, advertisement content can be provided to the customer based on information on products that the customer is interested in. In this case, the image processing system 10 can function as a signage system. The image processing system 10 can also function as a monitoring system or a crime prevention system that identifies what the suspicious person has noticed. The image processing system 10 can function as an authentication system that authenticates a specific subject.

  Further, the image analysis device 110 may acquire and analyze moving images of different store spaces 160. Thereby, a customer's trend etc. can be compared between stores. In the above description, the functions and operations of the image processing system 10 have been described as a store system. However, the image processing system 10 can be used for various purposes other than for a store.

  As mentioned above, although this invention was demonstrated using embodiment, the technical scope of this invention is not limited to the range as described in the said embodiment. It will be apparent to those skilled in the art that various modifications or improvements can be added to the above-described embodiment. It is apparent from the scope of the claims that the embodiments added with such changes or improvements can be included in the technical scope of the present invention.

  The order of execution of each process such as operations, procedures, steps, and stages in the apparatus, system, program, and method shown in the claims, the description, and the drawings is particularly “before” or “prior to”. It should be noted that the output can be realized in any order unless the output of the previous process is used in the subsequent process. Regarding the operation flow in the claims, the description, and the drawings, even if it is described using “first”, “next”, etc. for convenience, it means that it is essential to carry out in this order. It is not a thing.

1 is a diagram illustrating an example of an image processing system 10 according to an embodiment. 2 is a diagram illustrating an example of a block configuration of an image analysis device 110. FIG. It is a figure which shows typically an example of the processing content in the image analysis apparatus. It is a figure which shows an example of the processing content in the process 450 typically. It is a figure which shows an example of the image area | region of a correct image and an incorrect image. It is a figure which shows another example of the image area | region of a correct image and an incorrect image.

10 image processing system 90 storage medium 100 imaging device 110 image analysis device 112 analysis result database 150 communication network 160 store space 162 product shelf 164 person 170 store server 180 terminal device 200 image acquisition unit 210 feature amount extraction unit 220 fitness calculation unit 230 Object extraction unit 240 Reference feature amount calculation unit 242 Update feature amount selection unit 244 Update feature amount calculation unit 246 Non-update feature amount selection unit 250 Object region estimation unit 260 Orientation specification unit 270 Feature amount update unit 280 Feature amount storage unit 290 Output unit 400 weak classifier 402 classifier 404 classifier set 410, 420, 430, 440, 450, 470, 480, 510, 530, 550, 560 processing 412 object peripheral image 422 teacher image 442 identification result 460 weak classifier 462 Another 464 classifier set 472 estimated area image 500,520,552,562 weak classifier 502 classifier 600, 700 frame images 610,620,710,720 area

Claims (13)

An image analysis device that extracts an object of a specific subject from a video,
A feature amount storage unit that stores a plurality of reference feature amounts in association with the direction of the subject;
A feature amount extraction unit that extracts a plurality of feature amounts from each of a plurality of movie constituent images included in the movie;
An object extraction unit that extracts an object of the specific subject from the moving image composition image based on the degree of matching of the plurality of extracted feature amounts with respect to a plurality of reference feature amounts stored in association with the direction of the subject;
An orientation identifying unit that identifies the orientation of the specific subject;
A reference feature value calculation unit that calculates a reference feature value adapted to the feature value of the extracted object from the reference feature value stored in the feature value storage unit;
An image analysis apparatus comprising: a feature amount update unit that updates a reference feature amount stored in the feature amount storage unit in association with the specified orientation with a reference feature amount calculated by the reference feature amount calculation unit. The image analysis device according to claim 1, wherein the reference feature amount calculation unit calculates a plurality of reference feature amounts adapted to the extracted feature amount of the object with higher strength as the degree of matching is higher. 3. The image according to claim 1, wherein the reference feature amount calculating unit calculates a plurality of reference feature amounts adapted to the extracted feature amount of the object at a frequency according to a feature amount type. Analysis device. The feature amount extraction unit extracts a plurality of feature amounts from images of different resolutions,
The feature amount storage unit stores a plurality of reference feature amounts corresponding to a plurality of feature amounts extracted from images of different resolutions,
The image analysis apparatus according to claim 1, wherein the reference feature amount calculation unit calculates a plurality of reference feature amounts adapted to the extracted feature amount of the object with an intensity corresponding to the resolution. . The image analysis apparatus according to claim 4, wherein the reference feature amount calculation unit calculates a plurality of reference feature amounts adapted to the extracted feature amount of the object with greater intensity as the resolution is larger. The image analysis apparatus according to claim 5, wherein the reference feature amount calculation unit calculates a plurality of reference feature amounts corresponding to a resolution larger than a predetermined value, adapted to the extracted feature amount of the object. The reference feature amount calculation unit includes:
A non-updated feature quantity selection unit that uses one or more types of reference feature values that meet a predetermined condition used when the object extraction unit extracts an object from the moving image constituent image as the adapted reference feature value The image analysis apparatus according to claim 1, comprising: The non-update feature quantity selection unit uses one or more types of reference feature quantities that are used when the object extraction unit extracts an object from the moving image constituent image and the degree of fitness is a predetermined value or more. The image analysis apparatus according to claim 7, wherein the adapted reference feature amount is used. 9. The reference feature amount calculation unit according to claim 1, wherein the reference feature amount calculation unit calculates a reference feature amount obtained by learning an image of a region including the extracted object as a correct image as the adapted reference feature amount. An image analyzing apparatus according to claim 1. The reference feature amount calculation unit calculates, as the adapted reference feature amount, a reference feature amount obtained by learning an image of the region other than the extracted object as an incorrect image. The image analysis apparatus according to any one of 9. The feature amount storage unit stores a plurality of reference feature amounts corresponding to a plurality of feature amounts extracted from images with different resolutions in association with the direction of the subject,
The orientation specifying unit compares the feature amount extracted by the feature amount extraction unit from an image having a resolution smaller than a predetermined resolution with a reference feature amount having the same resolution, thereby matching the degree of fitness for each direction of the subject. The image analysis apparatus according to any one of claims 1 to 10, wherein the orientation of the subject associated with the reference feature amount for which a greater degree of fitness is calculated is identified as the orientation of the specific subject. An image analysis method for extracting an object of a specific subject from a video,
A feature amount storing step for storing a plurality of reference feature amounts in association with the direction of the subject;
A feature amount extraction stage for extracting a plurality of feature amounts from each of a plurality of movie constituent images included in the movie;
An object extraction step of extracting the object of the specific subject from the moving image composition image based on the degree of matching of the plurality of extracted feature amounts with respect to a plurality of reference feature amounts stored in association with the direction of the subject;
A direction identifying step of identifying a direction of the specific subject;
A reference feature amount calculating step of calculating a reference feature amount adapted to the extracted feature amount of the object from the reference feature amount stored in the feature amount storage step;
An image analysis method comprising: a feature amount update step of updating the reference feature amount stored in association with the specified orientation in the feature amount storage step with the reference feature amount calculated in the reference feature amount calculation step. A program for an image analysis device that extracts an object of a specific subject from a video, the computer
A feature amount storage unit that stores a plurality of reference feature amounts in association with the direction of the subject;
A feature amount extraction unit that extracts a plurality of feature amounts from each of a plurality of movie constituent images included in the movie,
An object extraction unit that extracts the object of the specific subject from the moving image composition image based on the degree of matching of the plurality of extracted feature amounts with respect to the plurality of reference feature amounts stored in association with the direction of the subject;
A direction identifying unit that identifies a direction of the specific subject;
A reference feature value calculation unit that calculates a reference feature value adapted to the feature value of the extracted object from the reference feature value stored in the feature value storage unit;
A program that causes a reference feature value stored in the feature value storage unit in association with the specified direction to function as a feature value update unit that updates the reference feature value calculated by the reference feature value calculation unit.

Images