TWI698805B - System and method for detecting and tracking people - Google Patents

Abstract

This invention discloses a system and a method for detecting and tracking people for applying to an image provided by the fisheye camera, comprising a mask generating module for generating a plurality of detecting masks having different sizes in different areas of the screen, a foreground-background segmentation module splitting the image from the screen into foreground and background regions, a mask screening module applying a plurality of detection masks to the foreground region to filter out the detection mask of the candidate, a masking image capturing module applying the detection mask of the candidate to the original image to extract a mask image of the candidate, and a mask image classification module classifying the mask image of the candidate by using a classifier corresponding to the candidate mask of the candidate to obtain a mask image of the real character.

Description

Person detection and tracking system and method

This case is about a person detection and tracking technology, in detail, it is about a system and method for person detection and tracking in the picture provided by a fisheye camera.

With the rapid development of cameras, network technology, and artificial intelligence, the application of intelligent video surveillance technology has grown substantially. The main types of cameras used in video surveillance include projection cameras and fish-eye cameras, most of which are projective cameras. The imaging mechanism is similar to the human visual system, but the main disadvantage is its shooting range. limits. Usually, a field requires multiple cameras to cover its scope, resulting in increased hardware costs and data volume, as well as the complexity of mutual calibration and information integration.

Compared with projection cameras, the main advantage of top-illuminated fisheye cameras lies in their 360-degree coverage. At least a single camera can be used, which can completely cover a considerable space without the problem of camera dead ends. In terms of hardware cost and use Each has its advantages, so it is an important option for many indoor surveillance applications. In addition, a top-illuminated fisheye camera is used to move objects (including people) When used for tracking and identification purposes, another advantage is the lower interactive masking phenomenon, so it can have a higher accuracy in positioning, counting, tracking and other issues.

However, the existing technology for human detection or tracking on the image of the top-illuminated fisheye camera is far less than that of the projection camera. In addition to factors with fewer sources of information, the main challenge comes from the variation of the orientation and appearance of the characters in different positions in the image. The existing technology is mainly divided into two categories: the first category is only the separation of foreground/background, each foreground block is regarded as a person and tracked, but it is not recognized whether the foreground block is a person, so there may be other The environment of moving objects in the foreground is difficult; the second type is to perform image correction or dewarping first, and then apply traditional image human detection methods, but normalization or flattening usually only handles the area near the outer edge of the image. It will cause the image to be distorted, and it is still unable to deal with the change of the character in the center of the image (below the camera).

In view of the above situation, there are still many shortcomings in the above-mentioned conventional technologies, which need to be improved urgently.

In order to solve the above and other problems, this case discloses a system and method for human detection and tracking, which can be used in the image provided by a fisheye camera.

The human detection and tracking system in this case includes: a front background segmentation module, which divides the image from the frame into a foreground area and a background area; a mask generation module, which is used to generate different areas in different areas of the frame A plurality of detection masks of different sizes; a mask filtering module applies the plurality of detection masks to the foreground area to filter out candidate detection masks; masks The image capture module applies the detection mask of the candidate object to the image to capture the mask image of the candidate object; and the mask image classification module uses the detection mask corresponding to the candidate object The classifier classifies the mask image of the candidate object to obtain the mask image that is actually a person.

The person detection and tracking system in this case further includes a person tracking unit, which includes a mask feature comparison module, a person sequence update module, and a specific clothing recognition module, wherein the mask feature comparison module The group compares the mask images of subsequent images according to the image features calculated by the mask image classification module, so that when the comparison matches, the tracking person sequence is included.

The classifier in the human detection and tracking system in this case has a training phase and a classification phase. During the training phase, the classifier targets a plurality of radial regions from the center of the image to the center of each detection mask. Establish a classifier model corresponding to each detection mask. In the classification stage, use the candidate object's detection mask for the candidate object's mask image to select the corresponding classifier to classify the actual object Masked images of people.

The method of human detection and tracking in this case includes: capturing an image from the screen to divide the image into a foreground area and a background area; applying a plurality of detection masks to the foreground area to filter out candidate detections Mask; apply the detection mask of the candidate object to the image to extract the mask image of the candidate object; and use the classifier corresponding to the detection mask of the candidate object to mask the candidate object The images are classified to obtain masked images that are actually people.

The method of human detection and tracking in this case further includes calculating image features of the masked image of the candidate object, and comparing subsequent images based on the image features The masked image of, so that when the comparison matches, it will be included in the tracking person sequence.

In the method of human detection and tracking in this case, the method for generating the plurality of detection masks includes: placing the character model at different positions in the scene to calculate the area occupied by the character model in the image or the character model In the depth of field of the image, a plurality of detection masks of different sizes are generated in different areas in the frame.

Therefore, this case uses a plurality of detection masks for human detection. The detection mask itself contains information about the direction and size variation of people in different positions, and covers the shadow caused by fixed objects in the scene. Therefore, Real-time detection can be achieved. In addition, this case also trains individual classifier models for different radial distance regions in the image, so people with different shapes in different regions can be effectively detected. Furthermore, this case can also concatenate the detected persons into a sequence, thereby achieving the effect of real-time tracking and recording of the trajectory, which has higher accuracy and practicability than the conventional technology.

1‧‧‧Screen image

100‧‧‧Fisheye Camera

11‧‧‧Foreground area

12‧‧‧Detection mask

2‧‧‧People detection and tracking system

200‧‧‧Personal mask setting unit

210‧‧‧Scene Setting Module

220‧‧‧Mask generation module

300‧‧‧People Detection Unit

310‧‧‧Image capture module

320‧‧‧Front background segmentation module

330‧‧‧Mask filter module

340‧‧‧Mask image capture module

350‧‧‧Mask image classification module

400‧‧‧People Tracking Unit

410‧‧‧Mask feature comparison module

420‧‧‧Character Sequence Update Module

430‧‧‧Specific clothing recognition module

501~519‧‧‧Cube

A~C‧‧‧People

E, F‧‧‧Fixed objects

S201~S206‧‧‧Step

Figure 1A is a schematic diagram of the configuration of an embodiment of the system for human detection and tracking; Figure 1B is a schematic diagram of the configuration of a specific embodiment of the system for human detection and tracking; Figure 2 is a schematic diagram of the configuration of the human detection and tracking system A schematic diagram of the flow of an embodiment of the method; Figure 3 is a block diagram of a specific embodiment of the method for human detection and tracking in this case; Figure 4 is a fisheye shot of the system and method for human detection and tracking in this case Schematic diagram of the scene of the camera; Figure 5 is a schematic diagram of the fish-eye camera of the system and method for human detection and tracking; Figure 6 is the screen of the fish-eye camera of the system and method for human detection and tracking A schematic diagram of the foreground area of the captured image; Figure 7A is a schematic diagram of the detection mask of the system and method of human detection and tracking; Figure 7B is the detection mask of the system and method of human detection and tracking Figure 8 is a schematic diagram of the application of the detection mask in the foreground area of the system and method for human detection and tracking; Figure 9 is a screened view of the system and method for human detection and tracking Figure 10 is a schematic diagram of the radial area of the fisheye camera screen of the system and method for human detection and tracking; and Figures 11A to 11D are the system for human detection and tracking. The actual picture taken by the fisheye camera and the application of the detection mask of the method.

The following specific examples are used to illustrate the implementation of this case. Those familiar with the art can easily understand the other advantages and effects of this case from the content disclosed in this article. The structure, ratio, size, etc. shown in the drawings in this manual are only used to match the content disclosed in the manual for the understanding and reading of those who are familiar with the art, and are not used to limit the implementation of this case. Therefore, any modification, change or adjustment will not affect the The effect and the achievable purpose should still fall within the scope of the technical content disclosed in this case.

Referring to FIGS. 1A and 1B, the human detection and tracking system 2 of this case includes a human mask setting unit 200, a human detection unit 300, and a human tracking unit 400.

The character mask setting unit 200 includes a scene setting module 210 and a mask generation module 220. The scene setting module 210 is used for the user to set the parameters of the fisheye camera, the scene range, the position and size information of the fixed objects in the scene, and the actual size parameters of the characters. The mask generation module 220 is used to generate a plurality of detection masks with different sizes in different areas of the fisheye camera's screen. The mask generation module 220 can be based on the parameters or information input by the user, etc. , Place the character model in different positions in the scene. In other words, the character model can be generated according to the actual size parameters of the character set by the user, and then the character model can be placed in different positions in the scene to calculate the range of the image occupied by the character model Or the depth of field of the character model in the image, that is, according to the parameters of the fisheye camera that have been set, the scene range, the position and size information of the fixed objects in the scene, and the actual size of the character, the character model in the image can be calculated Depth of field, thereby generating multiple detection masks of different sizes in different areas of the screen.

In addition, the mask generation module 220 also calculates the depth (distance from the camera) of the character model and the scene object at each pixel position of the image, and by comparing their relative depths, the detection mask area is actually visible (not covered by the scene object). ) Scope. In this way, the mask generation module 220 can generate a set of detection masks. The information of each detection mask includes its range in the original image, the coordinates of each pixel in the image covered by the detection mask in the original image, The actual visible area in the image covered by the detection mask and the coordinate position of the detection mask in the real world. Therefore, in the subsequent human detection stage, the scene objects can be processed efficiently without any additional judgment The problem of partial occlusion of characters.

The person detection unit 300 includes an image capturing module 310, a front background segmentation module 320, a mask filtering module 330, a mask image capturing module 340, and a mask image classification module 350. The image capturing module 310 can capture an image from the frame captured by the fisheye camera, and the front background segmentation module 320 divides the image into a foreground area and a background area, thereby updating the background model and calculating the foreground area in the screen.

The mask filtering module 330 uses the plurality of detection masks calculated by the human mask setting unit 200 to apply them to the foreground area of the image, and then calculates the proportion of the foreground area included in each detection mask range to filter A mask whose foreground ratio reaches a given threshold or more, that is, a detection mask that may be a person (also called a candidate) is generated.

The mask image capturing module 340 applies the detection mask that may be a person (candidate) to the original image to capture the mask image that may be a person (candidate), that is, the original image is likely to be a person ( The part covered by the detection mask of the candidate object), in which the captured image can be corrected.

The mask image classification module 350 can calculate the image features of the mask image that may be a person (candidate), and then use a classifier corresponding to the detection mask that may be a person (candidate) to determine each possible person ( The mask image of the candidate) generates a score of the possibility of being a person, thereby classifying the mask image that is a person. In detail, for the training of the classifier, the training data used is the masked image that was manually marked as human or non-human by the operator in advance. In the training phase, this classifier is targeted at several radial images in the fisheye camera image. Area (segmented according to the ratio of the distance between the center of the mask and the center of the image relative to the radius of the fisheye camera image), use the center of the mask to be located in this area For training mask images in the domain, a different classifier model is established, and in the actual classification stage, the classifier model of the region where each mask is located is also used to classify it. In other words, the classifier establishes a classifier model corresponding to each detection mask for a plurality of radial regions from the center of the image to the center of each detection mask, so that in the training phase, the probability is The mask image of the person (candidate) uses the area where the detection mask of the person (candidate) is located to select a corresponding classifier to classify the mask image that is actually a person.

The person tracking unit 400 includes a mask feature comparison module 410, a person sequence update module 420, and a specific clothing recognition module 430, wherein the mask feature comparison module 410 can be calculated according to the mask image classification module 350 The image feature of the mask image is compared with the mask image of the subsequent image, so that when the comparison matches, the tracking person sequence is included. In detail, the character tracking unit 400 maintains a set of records, including all the character sequences being tracked, and each character sequence includes the movement track of a character from entering the scene to leaving the scene. For each new camera frame, the person detection unit 300 generates and sends to the person tracking unit 400, including a set of detection masks and/or mask images, and the probability scores that these detection masks contain actual people (Masks whose probability scores are lower than a preset threshold are excluded), and the mask feature comparison module 410 compares these masks with the sequence of people being tracked, according to their appearance similarity and location conformity. (The final position of the character sequence and the real world coordinates of the mask) and the possibility that the masked image is a character. Determine the mask with the best degree of conformity with each character sequence. If the degree of conformity reaches a preset threshold Above, the character sequence update module 420 will cover Add a character sequence to the mask position, and update the appearance and position information of the character sequence with the mask’s appearance and position information. For the character sequence that cannot be connected with the mask selected by the human detection unit 300, its position and appearance information will not be updated temporarily, and it will continue to be tracked in a certain number of frames. The mask link among them determines that the character sequence has left the scene and stops tracking. The person sequence update module 420 is also responsible for the initialization of the person sequence. The method is to first select the high person possibility mask selected by the person detection unit 300, if it cannot be connected to the existing sequence, and its person possibility is higher than the neighbor The other masks are used to initialize the new character sequence. The user can also select part of the scene as the scene entrance area during the scene setting stage, and initialize the character sequence in such areas. The purpose of the specific clothing recognition module 430 is to identify the person who recognizes the specific clothing (such as a shop assistant wearing a uniform). The clothing appearance model used in the recognition phase is in the initialization and setting phase of the system, and the mask with the special clothing is manually selected. The mask image is created.

Next, referring to Figure 2, the method of human detection and tracking in this case can be outlined in Figure 2. In step S201, an image is captured from the screen; in step S202, the image is divided into a foreground area and a background area; in step S203, a plurality of detection masks are applied to the foreground area to filter out possible people ( In step S204, the detection mask that may be a person (candidate) is applied to the original image to extract the mask image that may be a person (candidate); and in step S204 In S205, a classifier corresponding to the detection mask that may be a person (candidate) is used to classify the masked image that may be a person (candidate); in step S206 In, the masked image that is actually a person is obtained.

In step S203, the method for generating the plurality of detection masks includes: placing the character model in different positions in the scene to calculate the range of the character model in the image or the depth of field of the character model in the image, by This generates a plurality of detection masks with different sizes in different areas in the frame.

In step S205, during the training phase of the classifier, the classifier establishes a classifier model corresponding to each detection mask for a plurality of radial regions from the center of the image to the center of each detection mask , In the classification stage of the classifier, the mask image that may be a person (candidate) is used to select the corresponding classifier in the area of the detection mask that may be a person (candidate) to classify the real The mask image of the character. After step S206, the method further includes the following method: calculating the image feature of the mask image that may be a person (candidate), and comparing the mask image of the subsequent image according to the image feature, so as to include the tracking when the comparison matches Character sequence.

Therefore, this case first uses the detection mask to filter out the detection masks that may be people (candidates) in the foreground area of the image, so as to extract the masked images that may be people (candidates) from the original image, and then use The classifier scores the masked images to classify the masked images that are actually people. The size of the detection mask in this case can be pre-calculated based on scene information, etc., and the classifier in this case is also pre-trained, and the corresponding classifier is selected based on the distance between the detection mask and the image center, thereby generating the possibility of detection mask Is the score of the character.

Continue to refer to Figure 3, which is a block diagram of a specific embodiment of the method for human detection and tracking in this case. The following diagrams are matched with the diagrams in Figures 4 to 10 And the actual diagrams 11A~11D.

Boxes 501~502 indicate: The fisheye camera 100 shown in Figure 4 shoots the scene below from a vertical downward angle. There are two fixed objects E and F (such as a store shelf) and three people in the scene image. A, B, and C. Block 503 indicates: in the initial setting stage, the operator first uses the scene setting module to set the parameters of the camera and the actual size of the person. In addition, the operator can also use the scene according to the screen image 1, as shown in Figure 5 or 11A. The setting module sets the scene range, the position and size information of fixed objects in the scene, as shown in Figure 11B, and then the mask generation module automatically calculates the depth information of the scene, as shown in Figure 11C.

Block 504 indicates that the mask generation module automatically calculates the length and width of people in different positions of the scene in the image, and generates a detection mask used in the image. Figure 7A shows the position and range of the detection mask corresponding to the positions of the three characters A, B, and C in the original image; the actual mask shape (the dotted ellipse as shown in the figure), the number or density can be operated者 Settings. The mask generation module also calculates the depth information of the person in the image for each detection mask, and compares it with the depth information of the scene object to determine the actual visible area of each detection mask, as shown in Figure 7B Show.

Blocks 505~507 indicate that the image capturing module captures the frame captured by the fisheye camera and sends it to the front background area segmentation module to automatically construct and update the background image model. The front background area segmentation module can calculate the difference between the screen image 1 and the background image model, and generate a grayscale or binary image to mark the foreground area 11 in the screen image 1, as shown in FIG.

Box 508 means: apply all detection masks to this foreground area, Calculate the ratio of the foreground area in the detection mask, that is, as shown in Figure 8, in the screen image 1, the detection mask 12 is merged with the foreground area 11, and then the detection mask whose foreground ratio reaches a critical value is filtered out Cover, as shown in Figure 9. As shown in Figure 11D, several detection masks are applied to the image of the actual screen. Then, apply the filtered detection mask to the original image to generate a corrected mask image.

Blocks 509, 510, and 512 indicate that: the mask image classification module calculates the image features of the mask images that have passed the screening detection mask, and then uses the trained classifier to generate real human possibilities for each mask image A sexual score, which is used to distinguish whether the mask image is likely to be a person (candidate) or other foreground objects other than a person, and to exclude the mask image whose probability score is lower than a predetermined threshold. Specifically, the image of the fisheye camera screen is divided into several radial regions (according to the ratio of the distance between the center of the detection mask and the image center to the image radius, as shown by the dotted line in Figure 10, where the square It may be the background), each establishes different classifier models to correspond to the appearance differences of the characters at different radial distances, and in the actual classification stage uses the classifier model of the area where each detection mask is located to classify the mask image. In addition, block 511 indicates that the specific clothing recognition module can compare the appearance features of the character sequence with the specific clothing features set by the operator, and if the degree of conformity is higher than a threshold, it will mark the sequence as a character with specific clothing (such as Store clerk in uniform).

Block 513 indicates: the mask feature comparison module calculates the feature values for tracking of the retained mask image, including appearance (color, texture, etc.) and its position in the real world, and uses these features to compare with the tracking data. Characters Compare the listed features and calculate their similarity.

Blocks 514 to 519 indicate that: for each person sequence in the tracking, select the mask with the best degree of conformity. If the degree of conformity exceeds a preset threshold, the person sequence update module will detect this The position of the mask is added to the character sequence (extending its track record), and the appearance and position information of the character sequence are updated with the appearance and position information of the detection mask. For the human sequence that cannot be linked to the selected detection mask, its position and appearance information will not be updated temporarily, and it will be tracked continuously in a certain number of frames. If none of these screens can be matched with the detection mask. If the cover is connected, it is determined that the character sequence has left the scene and the tracking is stopped. For the detection mask that cannot be connected to the existing human sequence, if the human probability score reaches a threshold and is higher than other adjacent detection masks, then the new human sequence is initialized.

In summary, the object detection and tracking system and method in this case can be roughly divided into the pre-calculation and generation of the detection mask, the calculation of the possibility (or score) that the masked image may be a person, and the selection of adjacent screens The detection masks are connected in series for the person tracking of the person sequence.

First, the operator sets the parameters of the camera, the scope of the scene, the position and size information of the fixed objects in the scene, and the parameters of the character size, and then places the three-dimensional character model at different positions in the scene, and then uses the camera's imaging mechanism Calculate the range of the character model in the image to obtain the length and width of the character mask at different positions in the image; calculate the relative depth of the character model and the scene object at each pixel position of the image to obtain the actual visible area of the character mask ( The range not covered by scene objects). Above steps A series of people detection masks will be generated. The information of each mask includes its range in the original image, the coordinates of each pixel in the mask image in the original image, the actual visible area in the mask image, and the mask in the real world Coordinate position in. This method can effectively deal with the problem of partial occlusion of people by scene objects without any additional judgment in the subsequent human detection stage.

Secondly, for the frame captured by the fisheye camera, the front background area is segmented to update the background model and calculate the foreground area in the screen, and then apply the pre-calculated human detection mask to the foreground area in the image. Calculate the ratio of the foreground area included in each mask range, filter the mask whose foreground ratio reaches a given threshold or more, and apply it to the original image again to generate an image of the mask area that has been normalized. The selected mask image is first calculated for its image characteristics, and then the trained person classifier is used to generate a score for the likelihood of each mask image being a person. This classifier will use a classifier model based on the image features of people in this region according to the radial region where the mask is located in the original image (the distance between the center of the mask and the center of the image relative to the radius of the fisheye camera image). Therefore, it can solve the problem that the shape of the person in the fisheye camera image varies with the position.

Finally, each character sequence contains a character's activity trajectory from entering the scene to leaving the scene. For each new camera screen, the selected mask that may be a person (including the possibility score of the person actually being a person) is compared with the sequence of the person being tracked, according to their appearance similarity, position coincidence, and mask The possibility that the image is a person, determine the mask with the best degree of conformity with each person sequence, if the degree of conformity exceeds the preset threshold, the mask position is added to the person sequence, and the mask is The appearance and location information update the appearance and location information of the character sequence. For the character sequence that cannot be linked to any selected mask, its position and appearance information will not be updated temporarily, and it will be tracked continuously in a certain number of screens. If it is not connected to the mask in these screens , It is judged that the character sequence has left the scene and stop tracking. For the initialization of the character sequence, the method is to first select the mask of high character possibility. If it cannot be connected with the existing sequence and its character possibility is higher than other adjacent masks, then use it to initialize the new character sequence. It is also possible to select some areas of the scene as the scene entrance area during the scene setting stage, and to limit the initial character sequence to such areas.

The above-mentioned embodiments only illustrate the effects of this case, and are not used to limit the case. Anyone familiar with this technique can modify and change the above-mentioned implementation aspects without departing from the spirit and scope of the case. Therefore, the scope of protection of the rights in this case should be listed in the scope of patent application described later.

2‧‧‧People detection and tracking system

220‧‧‧Mask generation module

320‧‧‧Front background segmentation module

330‧‧‧Mask filter module

340‧‧‧Mask image capture module

350‧‧‧Mask image classification module

Claims (10)

A system for human detection and tracking, used in the picture provided by a fisheye camera. The system includes: a front background segmentation module that divides the image from the picture into a foreground area and a background area; To generate a plurality of detection masks of different sizes in different areas of the screen; the mask filtering module applies the plurality of detection masks to the foreground area to filter out candidate detection masks Mask; mask image capture module, apply the detection mask of the candidate object to the image to capture the mask image of the candidate object; and mask image classification module, use the detection mask of the candidate object The classifier corresponding to the mask classifies the mask image of the candidate object to obtain the mask image that is actually a person. For example, the system described in item 1 of the scope of patent application further includes a character mask setting unit. The character mask setting unit includes a scene setting module and the mask generation module, wherein the scene setting module is used for setting Camera parameters, scene range, position and size information of fixed objects in the scene, or actual size parameters of characters. The system described in item 2 of the scope of patent application, wherein the mask generation module generates a character model according to the set actual size parameters of the character, so as to place the character model in different positions in the scene to calculate the character The model occupies the scope of the image, or according to the parameters of the set camera, the scope of the scene, the position and size of fixed objects in the scene, or the characters The actual size parameter is used to calculate the depth of field of the character model in the image, thereby generating a plurality of detection masks with different sizes in different areas in the frame. For example, the system described in item 1 of the scope of patent application further includes a person detection unit, which includes an image capture module, the front background segmentation module, the mask filter module, and the mask image capture A module and the mask image classification module are obtained, wherein the mask image classification module further calculates image features of the mask image of the candidate object. For the system described in item 4 of the scope of patent application, in the training phase, the classifier creates a corresponding detection mask for a plurality of radial regions from the center of the image to the center of each detection mask The masked classifier model, in the classification stage, selects a corresponding classifier for the masked image of the candidate object using the area where the candidate object's detection mask is located to classify the masked image that is actually a person. For example, the system described in item 4 of the scope of patent application further includes a person tracking unit, which includes a mask feature comparison module, a person sequence update module, and a specific clothing recognition module, wherein the mask feature The comparison module compares the mask images of the subsequent images according to the image features calculated by the mask image classification module, so that when the comparison matches, the tracking person sequence is included. A method for human detection and tracking used in a picture provided by a fisheye camera includes: capturing an image from the picture to divide the image into a foreground area and a background area; Apply a plurality of detection masks to the foreground area to filter out the detection mask of the candidate object; apply the detection mask of the candidate object to the image to capture the mask image of the candidate object; and use The classifier corresponding to the detection mask of the candidate object classifies the mask image of the candidate object to obtain a mask image that is actually a person. Such as the method described in item 7 of the scope of patent application, wherein the method for generating the plurality of detection masks includes: generating a character model according to the set actual size parameters of the character to place the character model in different positions in the scene, To calculate the range of the character model occupying the image, or to calculate the depth of field of the character model in the image based on the set camera parameters, scene range, position and size information of fixed objects in the scene, or actual size parameters of the character, In this way, a plurality of detection masks with different sizes are generated in different areas in the frame. Such as the method described in item 7 of the scope of patent application, wherein, in the training phase, the classifier creates a corresponding detection mask for a plurality of radial regions from the center of the image to the center of each detection mask The masked classifier model, in the classification stage, selects a corresponding classifier for the masked image of the candidate object using the area where the candidate object's detection mask is located to classify the masked image that is actually a person. For example, the method described in item 7 of the scope of patent application further includes calculating image features of the mask image of the candidate object, and comparing the mask images of subsequent images according to the image features, so that when the comparison matches, the tracking person sequence is included .

Images