JP6296769B2 - Object detection device - Google Patents

Description

  The present invention relates to an object detection apparatus that detects an object from an image.

  2. Description of the Related Art Conventionally, an object detection device that detects a target object such as a human body or a vehicle from an image has been studied for various purposes such as a photo or video image search or an image sensor for crime prevention. In recent years, many researches have been made on object detection devices using machine-learned classifiers. For example, the learning classifier uses feature quantities extracted from each of a large number of learning images in which a human body is photographed and a large number of learning images in which a human body is not photographed. It is generated by machine learning of an identification boundary that divides the space. When a feature value extracted from an image for which the presence / absence of a human body is to be determined is input to this classifier, whether or not the human body is reflected in the image depends on which side of the feature boundary the feature value is located in the feature space. Determine whether.

  For example, Patent Document 1 proposes an object detection device that detects an object appearing in an input image. In this object detection device, a learning image (specimen image) is divided into a plurality of cells. Each cell includes a partial discriminator that identifies the presence / absence of a portion of the object, and an overall discriminator that can identify the presence / absence of the object from cell identification index values output by all the partial discriminators.

JP 2011-215883 A

  Since the object detection device described in Patent Document 1 uses the identification result for each cell to detect the object using the overall classifier, a part of the object is hidden by another object However, the object can be detected with high accuracy. However, depending on the position of the cell, the characteristic part of the target object appearing in the input image (for example, the head when the target object is a person) does not fit in one cell, and spans multiple cells. In some cases, the identification accuracy of the partial classifier that learned the feature of the part is lowered, and the object cannot be detected. In addition, if the cell is enlarged so that the part fits in one cell, the proportion of the part features in the whole cell decreases, and the classification accuracy of the partial classifier decreases depending on the number of learning of the partial classifier. As a result, the performance of the overall discriminator for identifying the presence / absence of the target object also deteriorates.

  An object of the present invention is to provide an object detection device capable of accurately detecting a target object using the characteristics of the part even when the position where the characteristic part of the target object appears in the image fluctuates. It is in.

In order to solve such a problem, the present invention provides an object detection device for identifying whether or not a target object image exists in a detection target image using a part feature that is an image feature of one or a plurality of parts constituting the target object. A part suitability calculator that sets a part window of a predetermined size at a plurality of positions in the detection target image and calculates a part suitability that is suitable for the part feature within the part window; Provided is an object detection device including at least a distribution generation unit that generates a distribution of part suitability and a distribution identification unit that identifies the presence or absence of the target object using the distribution of part suitability as a feature amount.
Thus, the present invention identifies the presence / absence of the target object using the situation in which the part features constituting the target object are distributed in the detection target image as a feature quantity, so that one part suitability, for example, the peak value of the part suitability The identification accuracy of the target object is improved as compared with the identification using.

  The part suitability calculation unit preferably determines a search range in which a part window is set for each part. Thereby, when the position where each part of the target object in the detection target image is likely to be located is known in advance, the range where the part may be located can be set as the search range. The accuracy of distribution can be further improved.

  It is preferable that the part suitability calculation unit changes the size of the part window within a predetermined change amount range. As a result, even if there is a change in the size of the target object image in the detection target image, the size of the part window can be changed to cope with the change, so that the distribution accuracy of the part suitability can be further improved. it can.

  The distribution generation unit uses the voting vector from the reference position where each part is located in the detection target image set for each part window to the same voting position, to vote from the part window position where the part suitability is calculated. It is preferable to generate the site suitability distribution. As a result, even if the site suitability at some sites cannot be calculated accurately due to occlusion, etc., the site suitability by other sites has been voted. The influence is reduced, and the identification accuracy of the target object is improved.

  It is also preferable that the distribution generation unit divides the detection target image into a plurality of regions and generates a fitness distribution for each of the divided regions. Thereby, when the approximate position for each part in the detection target image is known, it is possible to use a fitness distribution that places importance on the arrangement of the part.

  The object detection device according to the present invention has an effect that the target object can be accurately detected by using the feature of the part even when the position where the characteristic part of the target object appears in the image fluctuates. .

It is a schematic structure figure of a human body detection system by this embodiment. It is a schematic diagram for demonstrating site | part window setting information. (A)-(c) is a schematic diagram for demonstrating a mode that a site | part window is set to a detection target image. (A)-(d) is a schematic diagram for demonstrating site | part conformity distribution. It is a schematic diagram for demonstrating a human body distribution discriminator. It is a flowchart which shows the operation | movement of the human body detection process of the human body detection apparatus by this embodiment. (A)-(d) is a schematic diagram for demonstrating the process which produces | generates site | part fitness distribution for every group.

Hereinafter, an embodiment in which the object detection device of the present invention is applied to a human body detection device will be described with reference to the drawings.
The target object of the object detection device is not limited to a human body, and may be any object that appears in an image, such as a vehicle, a bag, or a door. Further, in the present embodiment, an example will be described in which the human body detection device detects a human body using distributions of all part suitability such as a head, a right shoulder, a left shoulder, a trunk, a right leg, and a left leg. In the present embodiment, six site suitability levels are used. However, the present invention is not limited to this, and a human body may be detected using one site suitability distribution. Or you may divide a site | part into details and may use many parts.

  FIG. 1 is a diagram showing a schematic configuration of a human body detection system 100 according to the present embodiment. The human body detection system 100 includes a human body detection device 1 and an imaging device 2.

  The imaging device 2 is a camera that captures a predetermined imaging region, for example, a photoelectric conversion element (for example, a CCD sensor, a C-MOS, etc.) that is arranged two-dimensionally and outputs an electrical signal corresponding to the amount of received light. And an imaging optical system for forming an image of the monitoring region on the photoelectric conversion element. The imaging device 2 is connected to the human body detection device 1 and outputs the captured image to the human body detection device 1.

  The human body detection device 1 includes a central processing unit (CPU), a micro processing unit (MPU), peripheral circuits, terminals, various memories, and the like, and detects a human body that is a detection target from a captured image captured by the imaging device 2. The human body is extracted from the photographed image. The human body detection device 1 includes an image acquisition unit 10, an output unit 20, a storage unit 30, and an image processing unit 40. Hereinafter, each part of the human body detection apparatus 1 will be described in detail.

  The image acquisition unit 10 is an interface that is connected to the imaging device 2, acquires a captured image from the imaging device 2, and outputs the captured image to the image processing unit 40 and its control circuit. In this embodiment, the image acquisition unit 10 acquires a captured image from the imaging device 2 and outputs the acquired image to the image processing unit 40. However, the image acquisition unit 10 acquires an image from a medium such as a hard disk and outputs the image to the image processing unit 40. Also good. Hereinafter, an image acquired by the image acquisition unit 10 and output to the image processing unit 40 is referred to as an input image.

The output unit 20 is an interface connected to an external device (not shown) and its control circuit. When the output unit 20 receives a result signal including a determination result as to whether or not a human body is included in the input image from the image processing unit 40, the output unit 20 converts the result signal into a signal in a format that can be received by the connected external device and outputs the signal. To do. When the determination result indicates that the input image includes a human body, the result signal further includes an image including the human body extracted from the input image.
Note that the result signal may include, instead of the image including the human body extracted from the input image, the input image and information indicating the coordinate information and the size of the region in which the human body is shown in the input image. Accordingly, for example, by using an external device connected to the output unit 20 as an object tracking device, it is possible to monitor a human body existing in the imaging region while tracking it.
The output unit 20 may output each information to an external device such as a monitoring center device via a communication line such as a general public line or a mobile phone line.

The storage unit 30 includes a semiconductor memory such as a ROM (Read Only Memory) and a RAM (Random Access Memory), or a magnetic recording medium and its access device or an optical recording medium and its access device. The storage unit 30 stores a computer program and various data for controlling the human body detection device 1, and inputs / outputs such information to / from the image processing unit 40. The various data includes at least part window setting information 31, part learning data 32, and distribution learning data 33.
The part learning data 32 includes information for calculating a part feature quantity that is one or more feature quantities useful for part identification, and a part calculated for an image in which a part is shown in the feature quantity space for the part feature quantity. Includes information such as the identification boundary that separates the space in which the feature quantity is distributed from the other spaces, information on which side the part feature quantity is located on the identification boundary, and the distance between the part feature quantity and the identification boundary . The part learning data 32 is determined by prior learning from a plurality of learning part images in which parts are shown and a plurality of learning non-part images in which parts are not shown. The part learning data 32 is used when the part amount of fitness is calculated by the feature amount calculator 43 and the part fitness calculator 44 described later.
The distribution learning data 33 is an identification boundary that separates a space in which a part suitability distribution calculated for an image showing a human body is distributed from other spaces in a feature amount space for a part suitability distribution, which will be described later. Information for determining which side the degree distribution is located on the identification boundary and the distance between the site fitness distribution and the identification boundary. The distribution learning data 33 is determined by pre-learning from a part fitness distribution generated from a plurality of learning human body images and a part fitness distribution generated from a plurality of learning non-human images. The distribution learning data 33 is used when determining whether or not a human body is included in the input image in the human body distribution discriminator 46 and the human body detection means 47 described later.
The part window setting information 31 will be described later.

The image processing unit 40 includes a so-called computer such as a memory, a peripheral circuit thereof, and a microprocessor thereof, detects a human body while referring to the storage unit 30 with respect to an input image acquired from the image acquisition unit 10, and outputs a result signal. Output to the output unit 20. The image processing unit 40 includes a detection window setting unit 41, a part window setting unit 42, a feature amount calculation unit 43, a part adaptation degree calculation unit 44, a part adaptation degree distribution generation unit 45, a human body distribution classifier 46, and a human body detection unit 47. Have. Each means of the image processing unit 40 is a functional module realized by software operating on a microprocessor. Each unit of the image processing unit 40 may be configured by an independent integrated circuit, firmware, a microprocessor, and the like.
Hereinafter, each unit of the image processing unit 40 will be described in detail.

  The detection window setting means 41 is a detection window having a size in consideration of the size of the human body to be detected in the input image, which is assumed from the resolution and installation conditions of the imaging device 2 and the standard height of the human body (for example, 170 cm). The entire input image is set to be scanned sequentially. Hereinafter, the image in the detection window set by the detection window setting means 41 at each scanning position in the input image is referred to as a detection target image.

  The detection window setting unit 41 normalizes the size of the detection target image and generates a detection target image with the scale adjusted. In this embodiment, normalization is performed to 60 pixels horizontally and 120 pixels vertically. Then, as will be described later, the part window setting means 42 has a predetermined position and a predetermined size for each part of the human head, right shoulder, left shoulder, trunk, right leg, and left leg based on the part window setting information 31. The part window is sequentially set as the detection target image, and the part window image is output.

  Here, the part window setting information 31 will be described with reference to FIG. The part window setting information 31 includes a reference position 211, information for detecting each part of the human head 201, right shoulder 202, left shoulder 203, torso 204, right leg 205, and left leg 206 in the detection target image. A reference size 212, a search range 213, a size change amount 214, and a voting vector 215 are included. 2 exemplifies the part window setting information 31 when the size of the detection target image is normalized to 60 pixels horizontally and 120 pixels vertically.

  The reference position 211 is a coordinate value representing the centroid position of each part determined empirically as the centroid position of the part window and the upper left of the detection target image as the origin (0, 0). As shown in FIG. 2, the head (30, 30), right shoulder (10, 45), left shoulder (50, 45), trunk (30, 60), right leg (20, 85), left leg (40, 85). ) Is stored. In this embodiment, the position of the center of gravity of the part is determined empirically. However, the average position of the center of gravity of the circumscribed rectangle of each part image in the plurality of learning person images used for the prior learning of the part learning data 32 is determined. It may be used. It should be noted that the human image used for the prior learning is normalized and scaled in the same manner as the detection target image.

  The reference size 212 is represented by the number of pixels in the vertical and horizontal directions using a circumscribed rectangle of each part determined empirically as a part window. In the present embodiment, as shown in FIG. 2, a head 30 × 30, a right shoulder 20 × 20, a left shoulder 20 × 20, a torso 30 × 50, a right leg 45 × 50, and a left leg 45 × 50 are stored. . In the present embodiment, the reference size 212 is determined empirically, but using the average of the length and width of the circumscribed rectangle of each part in the plurality of learning part images used for the prior learning of the part learning data 32, and the like. Also good.

  The search range 213 is a range in which the position of each part may fluctuate and deviate from the reference position, and is set for performing a scanning search in the part window. Specifically, when a part size window of the reference size 212 is set at the reference position 211 of each part, the part window is expressed in a ratio that the part size window of the same reference size 212 overlaps the part window. The larger the overlapping ratio, the more the range is set closer to the reference position 211. The smaller the overlapping ratio is, the farther from the reference position 211 is set. A small value is set for a part that can vary greatly from the reference position 211, and a large value is set for a part that can vary only slightly from the reference position 211. In the present embodiment, as shown in FIG. 2, the head 70%, right shoulder 80%, left shoulder 80%, trunk 90%, right leg 70%, and left leg 70% are stored. In other words, the body is set to have the smallest fluctuation, and the head and the left and right legs have the largest fluctuation. These values are determined experimentally and empirically. In the present embodiment, the search range 213 is determined empirically, but the average amount of deviation from the reference position of the circumscribed rectangle of each part in the plurality of part images for learning used for the prior learning of the part learning data 32 Etc. may be used.

  The size change amount 214 is an amount that changes the size of the part window not only for the part size window of the reference size 211 but also for the body shape that differs depending on the person and the environment changes that differ depending on the time of imaging. This is expressed as a ratio of enlargement or reduction from 212. In this embodiment, as shown in FIG. 2, the head 90-110%, right shoulder 90-110%, left shoulder 90-110%, trunk 80-120%, right leg 90-110%, left leg 90-110. % Is remembered. The step size for changing the part window is set every 1%. These values are determined experimentally and empirically. In the present embodiment, the size change amount 214 is determined empirically, but is calculated from the change range of the size of the circumscribed rectangle of each part in the plurality of learning part images used for the prior learning of the part learning data 32. May be. Further, the step size is determined by design according to the calculation load.

  The voting vector 215 is a vector having the reference position 211 of each part window as a starting point and the centroid position of the torso located near the center of the human body as an end point, and is used when generating a part fitness distribution described later.

  Here, with reference to FIG. 3, a state in which the part window setting means 42 sets a part window in the detection target image will be described. FIG. 3 is a diagram schematically showing the positional relationship between the detection target image and the part window when the human body is photographed in an ideal situation for the detection target image.

  FIG. 3A shows a case where the reference size part windows 301 to 306 are arranged at the reference position. In FIG. 3A, the reference position of each part window 301 to 306 is indicated by a black circle (●), and the voting vector is indicated by an arrow. When a human body that matches the learned result is photographed, the voting vectors of the respective part windows 301 to 306 are concentrated on the position of the center of gravity of the trunk as shown in FIG.

  FIG. 3B shows a part window search range and a voting vector using the head as an example. In FIG. 3B, the outer edge of the search range of the part window setting information is indicated by a dotted line 311, the barycentric position of each part window is indicated by a black circle (●), and the voting positions voted according to the respective voting vectors are shown. . As shown in the figure, the voting positions are dispersed according to the search range of the part window setting information shown in FIG. Therefore, even if the head position shifts due to a change in the posture of the human body or a difference in the photographing environment, if the head exists within the search range, the characteristics indicated by the head can be captured. In the present embodiment, for each part of the right shoulder, the left shoulder, the trunk, the right leg, and the left leg, a part window is sequentially set in the search range, and the part suitability distribution is created by voting the part suitability. Such part suitability distribution is an effective feature of the whole human body having each part even if the part suitability is low by matching the end points of the voting vectors of each part. Incidentally, if the purpose is to detect the head, only the head fitness distribution may be used.

  FIG. 3C shows a state in which the size of the part window is changed while the position is fixed at the reference position in accordance with the size change amount of the head in the part window setting information shown in FIG. . Since the position is fixed, the voting position remains the same as the position of the center of gravity of the trunk, but the degree of matching of the part described later changes due to the change in the size of the part window. This corresponds to fluctuations due to the size of the head that is photographed and the photographing angle.

  As shown in FIGS. 3A, 3B, and 3C, the part window setting unit 42 sequentially sets part windows in the detection target image while changing the position and size of each part window.

  The feature amount calculating unit 43 calculates a region feature amount from each region window image set by the region window setting unit 42 in the detection target image. In the present embodiment, the feature quantity calculation means 43 uses HOG (Histograms of Oriented Gradients) feature quantities as the part feature quantities. The HOG feature amount is calculated by setting a plurality of blocks in each part window image and normalizing a histogram with the sum of gradient intensities in each gradient direction as a frequency for each of a plurality of cells into which each block is divided.

  When the part feature amount is inputted, the part degree-of-compatibility calculation means outputs a part suitability degree indicating the possibility that each part is included in the part window in which the part feature amount is calculated. That is, the part suitability calculation means 44 includes a head suitability calculation means 441 that calculates the fit degree of the head, a right shoulder suitability calculation means 442 that calculates the fit degree of the right shoulder, and a left shoulder that calculates the fit degree of the left shoulder. The fitness calculation means 443, the body fitness calculation means 444 for calculating the body fitness, the right leg fitness calculation means 445 for calculating the right leg fitness, and the left leg fitness calculation means for calculating the left leg fitness. 446.

In the present embodiment, the feature amount calculation means 43 and the part suitability calculation means 44 are realized by a real Adaboost discriminator using the part learning data 32. Details of the real Adaboost classifier are disclosed, for example, in RESchapire and Y. Singer, Improved boosting algorithms using confidence-rated predictions, Machine Learning, 37, pp.297-336, 1999.
Here, the “part conformity degree calculation unit” used in the claims is “part part window setting means 42”, “feature amount calculation means 43” and “part conformity degree calculation means 44” in the present embodiment. It is added that it is realized in.

  The part suitability distribution generation means 45 votes the part suitability calculated for each part window set by the part window setting means 42 according to the vote vector of each part, and the voting value (the sum of the voted part suitability). ) Is created as a site fitness distribution.

Here, voting will be described with reference to FIG. First, as shown in FIG. 3 (b), the site suitability distribution generation means 45 sets the site suitability calculated from each site window set in the search range for the head to a voting vector 215 shown in FIG. Therefore, vote. Similarly, the part suitability distribution generation unit 45 votes the part suitability calculated from each part window for each part of the right shoulder, left shoulder, torso, right leg, and left leg according to the voting vector 215 shown in FIG. To do.
Further, as shown in FIG. 3C, the site suitability distribution generation unit 45 calculates the site fit calculated from each site window in which the reference size 212 in FIG. 2 is changed according to the head size change amount 214. Vote the degree according to the voting vector 215. Similarly, the part suitability distribution generation unit 45 also applies the reference size 212 of FIG. 2 according to the size change amount 214 for each part of the right shoulder, left shoulder, trunk, right leg, and left leg from each part window. Vote the calculated part conformity according to a voting vector.

  Next, the site fitness distribution, which is the result of voting the site suitability, will be described with reference to FIG. FIG. 4 is a diagram schematically showing a site fitness distribution generated for a detection target image showing a human body or a detection target image showing an object other than a human body. FIG. 4 shows a site fitness distribution in which the horizontal direction and the vertical direction of the detection target image are the x-axis and the y-axis, respectively, and the voting value at the position corresponding to each pixel of the detection target image is the z-axis.

  FIG. 4A shows a site fitness distribution 401 generated from a detection target image 400 in which a human body is shown in an ideal situation. As shown in the figure, in the part fitness distribution 401, the vote value is highest at the center of gravity position 402 of the trunk, and the vote value decreases as the distance from the center of gravity position 402 of the trunk is increased.

  FIG. 4B shows a site fitness distribution 411 generated from the detection target image 410 in which a human body with a tilted head is shown. As shown in the figure, in the part fitness distribution 411, the vote value at the center of gravity position 412 of the trunk is lower than the vote value at the center of gravity position 402 of the trunk of FIG. The voting value at positions 413 and 414 deviated from the position is high.

  FIG. 4C shows a site suitability distribution 421 generated from the detection target image 420 in which the human body is not shown and the outerwear hung on the hanger is shown instead. As shown in the figure, the detection target image 420 shows something similar to the right shoulder, left shoulder and torso of the human body, but does not show anything similar to the head, right leg and left leg. The peak voting value of the fitness distribution 421 is extremely low.

  FIG. 4D shows a site fitness distribution 431 generated from a detection target image 430 in which a human body is not shown and a tree is shown instead. As shown in the figure, the branches of the tree shown in the detection target image 430 are similar to the head, right shoulder, left shoulder, and trunk of the human body. However, the site fitness distribution 431 is extremely low throughout.

  In this way, when the human body is shown in the detection target image, even if the posture of the human body is slightly deviated, the part is shown at a position deviated from the reference position, so the vote value increases somewhere in the distribution. . On the other hand, when the human body is not shown in the detection target image, objects similar to all parts of the human body are not detected at the assumed positions, and the distribution of vote values becomes low throughout.

  The part fitness distribution generation unit 45 arranges the vote values at the positions corresponding to the pixels of the detection target image (60 pixels × 120 pixels) in a line in the order of the positions in order to be processed by the human body distribution discriminator 46 described later. The arranged 7200-dimensional vectors are generated as the site fitness distribution.

  The human body distribution discriminator 46 discriminates whether or not a human body image is present in the detection target image using the distribution learning data 33 stored in the storage unit 30 when the part suitability distribution is input as the feature amount. Output the result. The human body distribution classifier 46 is realized by using a support vector machine that uses the distribution learning data 33.

  Here, the principle for identifying the presence or absence of a human body image in the human body distribution discriminator 46 will be described with reference to FIG. FIG. 5 is a diagram schematically showing a two-dimensional feature amount space for simplification of the 7200-dimensional vector part fitness distribution. Actually, it is a 7200-dimensional feature space. In FIG. 5, each part fitness distribution learned by the support vector machine is indicated by x. A group 501 is a group of site fitness distributions generated from a learning image in which a human body is photographed, and a group 502 is a group of site fitness distributions generated from a learning image in which an object other than a human body is captured. . In the group 501, a site fitness distribution 503 generated from the detection target image 400 in which the human body is shown in the ideal state of FIG. 4A and a human body whose head is tilted in FIG. 4B are shown. The part suitability distribution 504 generated from the detection target image 410 is plotted in the feature amount space. In the group 502, a part matching distribution 505 generated from the detection target image 420 in which the outerwear hung on the hanger in FIG. 4C is shown, and a detection target in which the tree in FIG. A site fitness distribution 506 generated from the image 430 is plotted in the feature amount space. The identification boundary 507 is determined by learning so as to separate the group 501 and the group 502 and is stored as distribution learning data 33. The human body distribution discriminator 46 discriminates whether or not the human body image is included depending on which side of the identification boundary 507 the part fitness distribution of the input detection target image is located.

Further, the human body distribution discriminator 46 calculates the distribution fitness indicating the possibility that the site fitness distribution is generated from the detection target image including the human body based on the distance between the input site fitness distribution and the identification boundary. Output. As will be described later, the distribution suitability is used when the human body detection unit 47 determines the human body image to be output when there are a plurality of detection target images determined to include the human body.
As described above, the site suitability distribution is a result of voting the site suitability for a plurality of sites to a 7200-dimensional feature amount space, so that human body identification can be made taking into account the 7200 features.

  When it is determined by the human body distribution classifier 46 that the detection target image includes a human body, the human body detection unit 47 associates the position and size of the detection target image in the input image with the distribution suitability in the storage unit 30. Remember.

  Note that the human body detection means 47 only detects a human body on the detection target image when a predetermined number (for example, 2) or more of the detection target images determined to include a human body overlap each other by a predetermined size (for example, 50%). May be determined to be included. When a human body is included in the input image, there is a high possibility that it is determined that a human body is included in a plurality of detection target images. Thus, detection accuracy can be improved.

In addition, when there are a plurality of detection target images determined to include a human body, the human body detection unit 47 reads the position and size of each detection target image from the storage unit 30, and the detection target images overlap each other by a predetermined size or more. It is determined whether or not. The human body detection means 47 determines that each detection target image includes the same human body when the detection target images overlap each other by a predetermined size or more.
In that case, the human body detection unit 47 reads out the distribution fitness stored in association with each detection target image from the storage unit 30, and outputs the detection target image having the highest distribution fitness among the detection target images. To decide.
The part suitability distribution generation unit 45 stores the sum of the part suitability calculated for each detection target image in the storage unit 30 in association with the detection target image. Then, the human body detection unit 47 reads the sum of the part suitability stored in association with each detection target image from the storage unit 30, and outputs the detection target image having the largest sum of the part suitability among the detection target images. The human body image to be determined may be determined.
Or the human body detection means 47 may determine the detection target image to output using the Mean Shift method. In this case, the human body detection unit 47 sets a three-dimensional space in which the horizontal direction in the input image is the X axis, the vertical direction is the Y axis, and the size of the detection target image is the Z axis. The human body detection unit 47 plots points corresponding to the center position and size of each detection target image in the three-dimensional space. Then, the human body detection means 47 searches for a point at which the density of the plotted points becomes maximum by means of the Mean Shift method, and determines a human body image to output a detection target image corresponding to the maximum point.

  The human body detection means 47 determines whether or not a human body is included in the input image depending on whether or not there are one or more detection target images determined to include the human body, and the determination result and the human body are included. When there are one or more determined detection target images, a result signal including information on the human body image is sent to the output unit 20.

Hereinafter, the overall operation flow of the human body detection device 1 will be described with reference to FIG.
In step S <b> 101, the image acquisition unit 10 of the human body detection device 1 acquires an image captured by the imaging device 2 as an input image and outputs it to the image processing unit 40.
In step S102, the detection window setting unit 41 sets a detection window for the input image acquired from the image acquisition unit 10, acquires an image in the detection window in the input image as a detection target image, and a part window setting unit 42. Output to. The detection window is set until the entire input image is scanned sequentially in step S109 described later.

The processes in steps S103 to S105 are performed with reference to the part window setting information 31 and the part learning data 32 stored in the storage unit 30.
In step S <b> 103, the part window setting unit 42 sets a part window for the detection target image, and outputs the part window image to the feature amount calculation unit 43. Specifically, with respect to the head 201, first, a part window is set to a position where the center of gravity of the rectangular window of “30 × 30” of the reference size 212 becomes “(30, 30)” of the reference position 211 and detected. The part window image within the set part window in the target image is output to the feature amount calculating means.

  In step S104, first, the feature amount calculating unit 43 calculates a portion feature amount that is a HOG feature for the portion window image. Then, the part suitability calculation means refers to this part feature value and the part learning data stored in the storage unit 30 to calculate the part suitability and output it to the part suitability distribution generation means 45. Here, since it is the case of the head 201, the head suitability calculating means 441 calculates the degree of matching with the part feature amount using the part learning data learned for the head. If the right shoulder is not the head, the right shoulder fitness calculating means 442, the left shoulder fitness calculating means 443 if the left shoulder, the torso fitness calculating means 444 if the torso, the right leg fitness if the right leg. If the degree calculation means 445 is the left leg, the left leg fitness degree calculation means 446 is selected to calculate the site fitness.

In step S105, the site suitability distribution generation unit 45 refers to the voting vector 215 and votes for the site suitability. Here, since it is the case of the head 201, the part suitability is voted at (0, 30). The part suitability distribution generation means 45 sequentially adds the voted part suitability to the vote value at that position. After that, if all the votes are not performed in the order of the head 201, the right shoulder 202, the left shoulder 203, the trunk 204, the right leg 205, and the left leg 206 shown in FIG. 2, the process returns to step S103 to set the next part window. On the other hand, if the processing for all the part windows has been completed, the part suitability distribution is output to the human body distribution discriminator 46, and the process proceeds to step S106.
In step S106, using the distribution learning data 33 stored in the storage unit 30 as the part suitability distribution as the feature amount, the human body distribution classifier 46 calculates the distribution suitability and determines the presence or absence of the human body. And if it is determination that a human body exists (S107-Yes), the position, size, and distribution adaptability of the processing detection window which were processed will be temporarily stored. On the other hand, if it is determined that the human body does not exist (S107-No), the process directly proceeds to step S109 to determine whether a detection window has been set for the entire input image. If the setting of the detection window is not completed (S109-No), the process returns to step S102 and the same process is executed. On the other hand, if the setting of the detection window has been completed (S109-Yes), the process proceeds to step S110.
In step S110, it is determined whether or not a human body exists in the input image based on whether or not the detection window temporarily stored in step S108 by the human body detection unit 47 exists.
When it is determined that the human body is included in the input image (step S111-Yes), the human body detection unit 47 identifies the human body image to be output as the detection target image including the human body (step S112). On the other hand, if the human body detection unit 47 determines that the human body is not included in the input image (No in step S111), the process of specifying the human body image to be output is not performed, and the process proceeds to step S113. Next, when it is determined that the human body is included in the result signal and the input image, the human body detection unit 47 sends the human body image to the output unit 20 (step S113), ends a series of steps, and then displays the next input image. get.

  For example, when the imaging apparatus 2 is installed so that the human body just fits in the input image, the image processing unit 40 does not need to set a detection window in the input image. It may be used as In that case, the detection window setting means 41 is omitted from the image processing unit 40, and the processes of steps S102 and S109 are omitted in the flowchart of FIG.

  The preferred embodiments of the present invention have been described above, but the present invention is not limited to these embodiments. For example, in the present embodiment, an example has been described in which the human body detection device detects a human body using a distribution of part suitability for a plurality of parts. However, a human body is detected using a distribution of part suitability for one part. May be. When a human body is detected using the distribution of part suitability for one part, the human body cannot be detected unless the part is shown in the detection target image, but the part is shown in the detection target image. If so, the human body can be detected.

  Further, the site fitness distribution generation unit may generate a site fitness distribution separately for each site instead of generating one site fitness distribution for a plurality of sites. In that case, the image processing unit has a distribution discriminator corresponding to each part, and the human body detecting means inputs the part suitability distribution for the part for each part to the corresponding distribution discriminator. It is determined whether or not a human body is included in the detection target image based on the output identification information. For example, the human body detection means determines whether or not a human body is included in the detection target image based on whether or not the sum of the output identification information is equal to or greater than a determination threshold. In this case, the human body detection device needs to learn the distribution classifier for each part in advance, but for each part, even if no other part is shown, the part fitness distribution is not affected by the influence. On the other hand, even when the part is not shown, the human body can be detected from the identification information about the other part.

  In the present embodiment, the example in which the part window setting unit sets the part windows having different sizes has been described. However, only the part window having the reference size may be set. In that case, it may not be possible to cope with fluctuations due to the size of the part specific to the human body, the imaging angle, etc., but a standard human body in an ideal environment can be detected.

  Further, in the present embodiment, the example in which the part suitability is calculated by the classifier has been described, but the part suitability may be calculated by pattern matching. In that case, a pattern of an image in which an averaging process or the like is performed on a plurality of images showing a part is generated in advance and stored in the storage unit. The part suitability calculating means obtains the degree of similarity between each part window and the image pattern stored in the storage unit, and uses the obtained degree of similarity as the part suitability. The degree of similarity can be, for example, a normalized cross-correlation value between a part window and an image obtained by enlarging / reducing an image pattern to the same size as the part window.

FIG. 7 is a schematic diagram for explaining processing for generating a site suitability distribution for each group.
7A to 7D show the same detection target image 700. FIG. Each part window set by the part window setting means has a center position of 4 depending on which of the upper right area 702, upper left area 712, lower right area 722, and lower left area 732 of the detection target image 700 is present. It is classified into one group. The part suitability distribution generation unit generates part suitability distributions 701, 711, 721, and 731 from the part suitability calculated from the part windows included in the group for each classified group. On the other hand, the image processing unit has a distribution classifier corresponding to each classified group, and the human body detection means determines whether or not a human body is included in the detection target image based on the identification information output from the distribution classifier for each group. Determine whether. For example, the human body detection means determines whether or not a human body is included in the detection target image based on whether or not the sum of the output identification information is equal to or greater than a determination threshold. In this case, since the human body detection device can generate the site fitness distribution without being influenced by other groups, even if the site fitness distribution could not be properly generated in some groups, A human body can be detected with high accuracy.
Further, in this case, the human body can be detected with high accuracy even if the setting of the part window is not limited to the search range and is set for the entire detection target image.

  Also, the site suitability calculation means does not calculate the site suitability for each detection window, but calculates the site suitability for all the site windows set in the input image, and calculates the calculated site suitability. The degree may be stored in the storage unit. In that case, the part suitability calculation means acquires the part suitability by reading the corresponding part suitability from the storage unit in step S104 of FIG. As a result, it is not necessary to calculate the part suitability for overlapping part windows in each detection window, and the processing load can be reduced.

  As described above, those skilled in the art can make various modifications in accordance with the embodiment to be implemented within the scope of the present invention.

DESCRIPTION OF SYMBOLS 1 Human body detection apparatus 2 Imaging device 10 Image acquisition part 20 Output part 30 Storage part 40 Image processing part 41 Detection window setting means 42 Part window setting means 43 Feature-value calculation means 44 Part conformity degree calculation means 45 Part conformity degree distribution generation means 46 Human body distribution classifier 47 Human body detection means

Claims (5)

An object detection device for identifying whether or not a target object image exists in a detection target image using a part feature that is an image feature of one or more parts constituting the target object,
A storage unit that stores the positional relationship between the position of the part window set in the detection target image and the voting position for each part;
Setting the site window of a predetermined size into a plurality of positions in the detection target image for each of the sites, site fitness calculating unit for calculating a site fit is approximately conforms to the characteristic portions relative to the portion within the window When,
The part suitability for each part window set for each part is voted to a voting position corresponding to the positional relationship of the part, and a distribution of the total value of the part suitability voted for each voting position is generated. A distribution generator;
A distribution identifying unit that identifies the presence / absence of the target object using a feature amount having the total value at each voting position as an element ;
An object detection apparatus characterized by including at least.   The object detection apparatus according to claim 1, wherein the part suitability calculation unit defines a search range in which the part window is set for each part.   The object detection apparatus according to claim 1, wherein the part suitability calculation unit changes the size of the part window within a predetermined change amount range. The storage unit stores a reference position of each part window for each of the one or a plurality of parts and, as the positional relationship, a voting vector from the position of the part window set in the detection target image to the voting position. Remember,
4. The object detection device according to claim 1 , wherein the voting position is set to be located at the same position with respect to the reference position of each part window . 5. The distribution generation unit, the detection target image is divided into a plurality of regions of the object detection apparatus according to any one of claims 1 to 4 to produce a distribution of the site fit for each the divided regions.

Images