A target tracking method

Through YOLO neural network identification and clustering processing, feature points in the target area are extracted and the action trajectory of the target object is constructed, which solves the problem that the YOLO neural network cannot track the same target, and achieves accurate tracking of the target object.

CN118608569BActive Publication Date: 2024-11-05CHENGDU AEROSPACE KAITE ELECTROMECHANICAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411001078.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-05
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

In the prior art, YOLO neural networks can only achieve classification of goals, but cannot achieve tracking of the same goal.

Method used

By collecting images in the management area, identifying and labeling the target box using the YOLO neural network, cutting out the target area, performing clustering processing and feature point extraction, building the action trajectory of the target object, and tracking the same target. The specific steps include: clustering according to the pixel value distance, calculating cluster features, matching cluster feature vectors, filtering out multiple target areas of the same target object, and obtaining the target object's action trajectory through fitting processing.

Benefits of technology

Effective tracking of the same target is achieved, tracking accuracy is improved, and the motion trajectory of the same target object in the image can be correctly identified and tracked.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118608569B_ABST
    Figure CN118608569B_ABST
Patent Text Reader

Abstract

The invention discloses a target tracking method, and belongs to the technical field of image processing. The invention adopts a YOLO neural network to identify each frame of original image, adopts a target frame to mark each target position and each target type on each frame of original image, cuts out each target area according to the position of the target frame, extracts each target area of ​​the same type, and implements screening out target areas with consistent classification types, and then screens out multiple target areas belonging to the same target object from each target area of ​​the same type, extracts feature points from the multiple target areas belonging to the same target object, constructs a motion trajectory of the target object, and implements tracking of the same target.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a target tracking method. Background Art

[0002] In modern surveillance, target tracking is one of the core functions, which can provide continuous monitoring of moving targets in a specific area. With the development of artificial intelligence technology, target tracking methods based on deep learning have attracted widespread attention due to their high accuracy and automation. Among them, YOLO neural network, as a popular real-time target detection algorithm, is often used in target tracking systems for preliminary target recognition and positioning due to its fast and accurate characteristics.

[0003] There are multiple types of targets in one frame of image. After the image is recognized by the YOLO neural network, each type of target is marked with a corresponding target box and corresponding classification type to achieve classification of each target in the image. However, the YOLO neural network only realizes the classification of targets and cannot track the same target. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, a target tracking method provided by the present invention solves the problem that the prior art uses the YOLO neural network to only classify targets but cannot track the same target.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a target tracking method, comprising the following steps:

[0006] S1, collecting images in the management area to obtain multiple frames of original images;

[0007] S2, using the YOLO neural network to recognize each frame of the original image, and marking each target on each frame of the original image with a target box;

[0008] S3, cutting out the target area from each frame of the original image according to the position of the marked target frame;

[0009] S4, clustering each target area of ​​the same type, and screening out multiple target areas of the same target object based on cluster features;

[0010] S5. Extract feature points from multiple target areas and construct a movement trajectory of the target object.

[0011] The beneficial effects of the present invention are as follows: the present invention adopts the YOLO neural network to recognize each frame of the original image, adopts the target frame to mark the position and type of each target on each frame of the original image, cuts out each target area according to the position of the target frame, extracts each target area of ​​the same type, and screens out target areas with the same classification type, and then screens out multiple target areas belonging to the same target object from each target area of ​​the same type, extracts feature points from the multiple target areas belonging to the same target object, constructs the movement trajectory of the target object, and realizes tracking of the same target.

[0012] Furthermore, the S4 comprises the following sub-steps:

[0013] S41, performing clustering processing on each pixel point on each target area of ​​the same type according to the pixel value distance of each pixel point, to obtain multiple clusters on the target area;

[0014] S42, calculating each cluster feature according to the area ratio of each cluster in the target area and the pixel contrast value of each cluster in the target area;

[0015] S43, using each cluster feature on a target area as an element to form a cluster feature vector;

[0016] S44, calculating the matching degree between each cluster feature vector and other cluster feature vectors;

[0017] S45, extracting target regions corresponding to all cluster feature vectors whose matching degree is greater than a matching threshold, and obtaining multiple target regions of the same target object.

[0018] The beneficial effect of the above further scheme is as follows: the present invention performs clustering processing on each pixel point on each target area according to the pixel value distance of each pixel point, obtains multiple clusters on the target area, and realizes classification by pixel value. Each cluster is the area where each pixel point with similar pixel value is located. The feature of each cluster is calculated by the area ratio and pixel contrast value of each cluster on the target area. The cluster features on a target area are used as elements to form a cluster feature vector. The matching degree of each cluster feature vector and other cluster feature vectors is calculated. When the matching degree is greater than the matching threshold, it means that the target areas corresponding to the two cluster feature vectors corresponding to the matching degree belong to the same target object.

[0019] Furthermore, the formula for calculating each cluster feature in S42 is: , , , where θ is the cluster feature, e is a natural constant, and γ Q is the area ratio of each cluster in the target area, γ p Q is the pixel contrast value of each cluster in the target area, cis the number of pixels in each cluster, Q target is the number of pixels in the target area, P c is the mean pixel value of each cluster, P max is the maximum pixel value mean in all clusters, P min is the minimum mean pixel value among all clusters.

[0020] The beneficial effects of the above further scheme are as follows: the present invention reflects the area ratio by the ratio of the number of pixel points. There are multiple clusters in a target area. The area ratio of the cluster in the target area and the pixel contrast value are extracted. The exponential function is used to enhance the area ratio and the pixel contrast value. The area ratio and the pixel contrast value are combined to characterize the characteristics of each cluster, reflecting the specificity of each cluster.

[0021] Furthermore, the S44 comprises the following sub-steps:

[0022] S441, taking a cluster feature vector as a comparison vector, setting a flag bit for each of the other cluster feature vectors, where the initial value of the flag bit is 0;

[0023] S442, taking a cluster feature from the comparison vector as a comparison feature;

[0024] S443, calculating the similarity between the comparison feature and the cluster features in other cluster feature vectors;

[0025] S444, determine whether the similarity is greater than the similarity threshold. If so, add 1 to the flag bits of other cluster feature vectors whose similarity is greater than the similarity threshold, and jump to S442 to select the next cluster feature until all cluster features in the comparison vector are traversed. If not, directly jump to S442 to select the next cluster feature until all cluster features in the comparison vector are traversed.

[0026] S445. Use the flag bit as the matching degree between the cluster feature vector and another cluster feature vector.

[0027] The beneficial effect of the above further scheme is as follows: the present invention uses a cluster feature vector as a comparison vector, takes a cluster feature from the comparison vector as a comparison feature, calculates the similarity between the comparison feature of the comparison vector and the cluster features in other cluster feature vectors, and when the similarity is greater than a similarity threshold, it indicates that the corresponding other cluster feature vectors have cluster features similar to the comparison feature. Therefore, the corresponding flag bit is incremented by 1, and the cluster features in the comparison vector are traversed. When there are multiple cluster features in one other cluster feature vector whose similarity is greater than the similarity threshold, the corresponding flag bit indicates the degree of matching between the other cluster feature vector and the comparison vector.

[0028] Furthermore, the S5 comprises the following sub-steps:

[0029] S51, matching clusters on each target area to obtain multiple sets of matching cluster sets, and selecting the best matching cluster set from the multiple sets of matching cluster sets;

[0030] S52, retaining edge pixels for each matching cluster in the optimal matching cluster set to obtain a cluster outline;

[0031] S53, extracting feature points for each cluster outline;

[0032] S54, performing fitting processing on multiple feature points corresponding to the optimal matching cluster set to obtain the action trajectory of the target object.

[0033] The beneficial effect of the above further scheme is: in step S4, only multiple target areas belonging to the same target object are screened out, but there are multiple clusters on each target area, therefore, the clusters corresponding to the same structure on the same target object are matched to form a matching cluster set, from which the optimal matching cluster set is screened out, and the edge pixel points of each matching cluster in the optimal matching cluster set are retained to obtain the cluster contour, and feature points are extracted from the cluster contour, that is, the feature points of the same structure on the same target object in different images are obtained, therefore, the multiple feature points corresponding to the optimal matching cluster set are fitted to obtain the motion trajectory of the target object, thereby improving the accuracy of motion trajectory tracking.

[0034] Furthermore, the S51 includes the following sub-steps:

[0035] S511, calculating the similarity between the jth cluster feature on a target area and each cluster feature on other target areas, where the initial value of j is 1 and j is a positive integer;

[0036] S512, extracting the cluster corresponding to the maximum similarity greater than the similarity threshold from each other target area, obtaining the matching cluster on each other target area, and forming the jth matching cluster set with each matching cluster;

[0037] S513, adding the similarities corresponding to the matching clusters in the j-th matching cluster set to obtain the j-th similarity value;

[0038] S514, add 1 to j and jump to step S511 until all cluster features in the target area are traversed;

[0039] S515: Select the matching cluster set with the largest similarity value as the optimal matching cluster set.

[0040] The beneficial effect of the above further scheme is: calculating the similarity between the cluster features on a target area and the cluster features of each other target area. Since there are multiple cluster features of each other target area, there are also multiple similarities. Take the cluster corresponding to the maximum similarity and the maximum similarity is greater than the similarity threshold, and obtain the cluster that matches it, which is named the matching cluster. There are multiple target areas. Therefore, when iterating once, there is one matching cluster on each target area. The matching clusters are formed into the jth matching cluster set. The jth matching cluster set represents the set of imaging conditions of the same structure on the same target object on different images. The similarities corresponding to each matching cluster in a set of matching clusters are added to obtain a similarity value, and the matching cluster set corresponding to the maximum similarity value is taken to select the matching clusters with the best similarity.

[0041] Furthermore, the S53 comprises the following sub-steps:

[0042] S531, calculating the geometric center coordinate point of each cluster outline according to the coordinates of each pixel point on each cluster outline;

[0043] S532, taking each pixel point on the cluster outline as a comparison pixel point, and calculating the direction angle distribution value between the comparison pixel point and the geometric center coordinate point;

[0044] S533: Use the comparison pixel point corresponding to the maximum direction angle distribution value as a feature point.

[0045] The beneficial effect of the above further scheme is: the present invention takes each pixel point from the cluster contour as a comparison pixel point, calculates the direction angle distribution value of the comparison pixel point and the geometric center coordinate point, and reflects the specificity of the comparison pixel point through the direction angle distribution value.

[0046] Furthermore, the formula for calculating the direction angle distribution value of the comparison pixel point and the geometric center coordinate point in S532 is: , , , where γ is the directional angle distribution value, e is a natural constant, and θ c To compare the direction angle between the pixel point and the geometric center coordinate point, θ m To compare the direction angle between the mth pixel point in the pixel point neighborhood and the geometric center coordinate point, x c is the horizontal coordinate of the comparison pixel, y c is the ordinate of the comparison pixel, x o is the horizontal coordinate of the geometric center coordinate point, y o is the ordinate of the geometric center coordinate point, x m is the horizontal coordinate of the mth pixel in the neighborhood of the comparison pixel on the cluster contour, y mis the ordinate of the mth pixel in the neighborhood of the comparison pixel on the cluster contour, | | is the absolute value operation, m is a positive integer, M is the number of neighborhood pixels, and arctan is the inverse tangent function.

[0047] The beneficial effect of the above further scheme is: the present invention extracts the direction angles of the comparison pixel points, and the neighborhood pixel points and the geometric center coordinate points, and reflects the shape characteristics at the comparison pixel points by comparing the direction angle distributions of the comparison pixel points and the neighborhood pixel points. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flow chart of a target tracking method. DETAILED DESCRIPTION

[0049] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0050] like Figure 1 As shown, a target tracking method comprises the following steps:

[0051] S1, collecting images in the management area to obtain multiple frames of original images;

[0052] S2, using the YOLO neural network to recognize each frame of the original image, and marking each target on each frame of the original image with a target box;

[0053] S3, cutting out the target area from each frame of the original image according to the position of the marked target frame;

[0054] S4, clustering each target area of ​​the same type, and screening out multiple target areas of the same target object based on cluster features;

[0055] S5. Extract feature points from multiple target areas and construct a movement trajectory of the target object.

[0056] In this embodiment, the process of using the YOLO neural network in S2 to recognize each frame of the original image is a prior art.

[0057] In this embodiment, the same type in step S4 refers to the type of target recognized by the YOLO neural network. For example, the target types include ships and people. The same type refers to a target area of ​​the "ship" type or a target area of ​​the "person" type.

[0058] In step S4, since there may be multiple target objects in the same type of target region, multiple target regions belonging to the same target object are screened out according to the cluster features.

[0059] The S4 comprises the following sub-steps:

[0060] S41, performing clustering processing on each pixel point on each target area of ​​the same type according to the pixel value distance of each pixel point, to obtain multiple clusters on the target area;

[0061] S42, calculating each cluster feature according to the area ratio of each cluster in the target area and the pixel contrast value of each cluster in the target area;

[0062] S43, using each cluster feature on a target area as an element to form a cluster feature vector;

[0063] S44, calculating the matching degree between each cluster feature vector and other cluster feature vectors;

[0064] S45, extracting target regions corresponding to all cluster feature vectors whose matching degree is greater than a matching threshold, and obtaining multiple target regions of the same target object.

[0065] The present invention performs clustering processing on each pixel point on each target area according to the pixel value distance of each pixel point, obtains multiple clusters on the target area, and realizes classification according to pixel value. Each cluster is the area where each pixel point with similar pixel value is located. Each cluster feature is calculated by the area proportion and pixel contrast value of each cluster on the target area. Each cluster feature on a target area is used as an element to form a cluster feature vector. The matching degree of each cluster feature vector and other cluster feature vectors is calculated. When the matching degree is greater than the matching threshold, it means that the target areas corresponding to the two cluster feature vectors corresponding to the matching degree belong to the same target object.

[0066] The formula for calculating each cluster feature in S42 is: , , , where θ is the cluster feature, e is a natural constant, and γ Q is the area ratio of each cluster in the target area, γ p Q is the pixel contrast value of each cluster in the target area, c is the number of pixels in each cluster, Q target is the number of pixels in the target area, P c is the mean pixel value of each cluster, P max is the maximum pixel value mean in all clusters, P min is the minimum mean pixel value among all clusters.

[0067] The present invention reflects the area ratio by the ratio of the number of pixel points. There are multiple clusters in a target area. The area ratio of the cluster in the target area and the pixel contrast value are extracted. The exponential function is used to enhance the area ratio and the pixel contrast value. The area ratio and the pixel contrast value are combined to characterize the characteristics of each cluster, reflecting the specificity of each cluster.

[0068] In the present invention, since the cluster features of each cluster are arranged differently when forming the cluster feature vector, it is impossible to calculate the matching degree using calculation methods such as cosine similarity. It is necessary to use the method of step S44 to compare the cluster features one by one to accurately find multiple target areas of the same target object.

[0069] The S44 comprises the following sub-steps:

[0070] S441, taking a cluster feature vector as a comparison vector, setting a flag bit for each of the other cluster feature vectors, where the initial value of the flag bit is 0;

[0071] S442, taking a cluster feature from the comparison vector as a comparison feature;

[0072] S443, calculating the similarity between the comparison feature and the cluster features in other cluster feature vectors;

[0073] S444, determine whether the similarity is greater than the similarity threshold. If so, add 1 to the flag bits of other cluster feature vectors whose similarity is greater than the similarity threshold, and jump to S442 to select the next cluster feature until all cluster features in the comparison vector are traversed. If not, directly jump to S442 to select the next cluster feature until all cluster features in the comparison vector are traversed.

[0074] S445. Use the flag bit as the matching degree between the cluster feature vector and another cluster feature vector.

[0075] The numerical value of the flag bit of the present invention represents that there are several cluster features with similarities greater than a similarity threshold in a corresponding other cluster feature vector.

[0076] The present invention uses a cluster feature vector as a comparison vector, takes a cluster feature from the comparison vector as the comparison feature, calculates the similarity between the comparison feature of the comparison vector and the cluster features in other cluster feature vectors, and when the similarity is greater than a similarity threshold, it means that the corresponding other cluster feature vectors have cluster features similar to the comparison feature. Therefore, the corresponding flag bit is incremented by 1, and the cluster features in the comparison vector are traversed. When there are multiple cluster features in one other cluster feature vector whose similarity is greater than the similarity threshold, the corresponding flag bit indicates the matching degree between the other cluster feature vector and the comparison vector.

[0077] In the sub-steps of step S5, multiple target areas of the same target object are processed.

[0078] The S5 comprises the following sub-steps:

[0079] S51, matching clusters on each target area to obtain multiple sets of matching cluster sets, and selecting the best matching cluster set from the multiple sets of matching cluster sets;

[0080] S52, retaining edge pixels for each matching cluster in the optimal matching cluster set to obtain a cluster outline;

[0081] S53, extracting feature points for each cluster outline;

[0082] S54, performing fitting processing on multiple feature points corresponding to the optimal matching cluster set to obtain the action trajectory of the target object.

[0083] A feature point is extracted from each cluster outline, so the best matching cluster set corresponds to multiple feature points.

[0084] In step S4, the present invention only screens out multiple target areas belonging to the same target object, but there are multiple clusters on each target area. Therefore, the clusters corresponding to the same structure on the same target object are matched to form a matching cluster set, from which the optimal matching cluster set is screened out, and the edge pixel points of each matching cluster in the optimal matching cluster set are retained to obtain the cluster outline, and feature points are extracted from the cluster outline, that is, feature points of the same structure on the same target object on different images are obtained. Therefore, the multiple feature points corresponding to the optimal matching cluster set are fitted to obtain the motion trajectory of the target object, thereby improving the tracking accuracy of the motion trajectory.

[0085] The S51 comprises the following sub-steps:

[0086] S511, calculating the similarity between the jth cluster feature on a target area and each cluster feature on other target areas, where the initial value of j is 1 and j is a positive integer;

[0087] S512, extracting the cluster corresponding to the maximum similarity greater than the similarity threshold from each other target area, obtaining the matching cluster on each other target area, and forming the jth matching cluster set with each matching cluster;

[0088] S513, adding the similarities corresponding to the matching clusters in the j-th matching cluster set to obtain the j-th similarity value;

[0089] S514, add 1 to j and jump to step S511 until all cluster features in the target area are traversed;

[0090] S515: Select the matching cluster set with the largest similarity value as the optimal matching cluster set.

[0091] The present invention calculates the similarity between a cluster feature on a target area and the cluster feature of each other target area. Since there are multiple cluster features of each other target area, there are also multiple similarities. The cluster corresponding to the maximum similarity and the maximum similarity being greater than a similarity threshold is taken to obtain a cluster matching it, which is named a matching cluster. There are multiple target areas. Therefore, when iterating once, there is one matching cluster on each target area. The matching clusters are formed into a j-th matching cluster set. The j-th matching cluster set represents a set of imaging conditions of the same structure on the same target object on different images. The similarities corresponding to each matching cluster in a set of matching clusters are added to obtain a similarity value. The matching cluster set corresponding to the maximum similarity value is taken to select each matching cluster with the best similarity.

[0092] The S53 comprises the following sub-steps:

[0093] S531, calculating the geometric center coordinate point of each cluster outline according to the coordinates of each pixel point on each cluster outline;

[0094] S532, taking each pixel point on the cluster outline as a comparison pixel point, and calculating the direction angle distribution value between the comparison pixel point and the geometric center coordinate point;

[0095] S533: Take the comparison pixel point corresponding to the maximum direction angle distribution value as a feature point.

[0096] The present invention takes each pixel point on the cluster outline as a comparison pixel point, calculates the direction angle distribution value of the comparison pixel point and the geometric center coordinate point, and reflects the specificity of the comparison pixel point through the direction angle distribution value.

[0097] The formula for calculating the direction angle distribution value of the comparison pixel point and the geometric center coordinate point in S532 is: , , , where γ is the directional angle distribution value, e is a natural constant, and θ c To compare the direction angle between the pixel point and the geometric center coordinate point, θ m To compare the direction angle between the mth pixel point in the pixel point neighborhood and the geometric center coordinate point, x c is the horizontal coordinate of the comparison pixel, y c is the ordinate of the comparison pixel, x o is the horizontal coordinate of the geometric center coordinate point, y o is the ordinate of the geometric center coordinate point, x m is the horizontal coordinate of the mth pixel in the neighborhood of the comparison pixel on the cluster contour, y m is the ordinate of the mth pixel in the neighborhood of the comparison pixel on the cluster contour, | | is the absolute value operation, m is a positive integer, M is the number of neighborhood pixels, and arctan is the inverse tangent function.

[0098] The present invention extracts the direction angles of the contrasting pixel points, the neighboring pixel points and the geometric center coordinate point, and reflects the shape features at the contrasting pixel points by comparing the direction angle distributions of the pixel points and the neighboring pixel points.

[0099] In this embodiment, the threshold is set based on experience or experiments.

[0100] The present invention adopts the YOLO neural network to recognize each frame of the original image, uses a target frame to mark the position and type of each target on each frame of the original image, cuts out each target area according to the position of the target frame, extracts each target area of ​​the same type, and screens out target areas with consistent classification types, then screens out multiple target areas belonging to the same target object from each target area of ​​the same type, extracts feature points from the multiple target areas belonging to the same target object, constructs the movement trajectory of the target object, and realizes tracking of the same target.

[0101] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A target tracking method, It is characterized in that The following steps are involved: S1, collecting images in the management area to obtain multiple frames of original images; S2, using the YOLO neural network to recognize each frame of the original image, and marking each target on each frame of the original image with a target box; S3, cutting out the target area from each frame of the original image according to the position of the marked target frame; S4, clustering each target area of ​​the same type, and screening out multiple target areas of the same target object based on cluster features; S5, extracting feature points from multiple target areas and constructing the movement trajectory of the target object; The S4 comprises the following sub-steps: S41, performing clustering processing on each pixel point on each target area of ​​the same type according to the pixel value distance of each pixel point, to obtain multiple clusters on the target area; S42, calculating each cluster feature according to the area ratio of each cluster in the target area and the pixel contrast value of each cluster in the target area; S43, using each cluster feature on a target area as an element to form a cluster feature vector; S44, calculating the matching degree between each cluster feature vector and other cluster feature vectors; S45, extracting target regions corresponding to all cluster feature vectors whose matching degree is greater than a matching threshold, and obtaining multiple target regions of the same target object; The formula for calculating each cluster feature in S42 is: , , , where θ is the cluster feature, e is a natural constant, and γ Q is the area ratio of each cluster in the target area, γ p Q is the pixel contrast value of each cluster in the target area, c is the number of pixels in each cluster, Q target is the number of pixels in the target area, P c is the mean pixel value of each cluster, P max is the maximum pixel value mean in all clusters, P min is the minimum pixel value mean among all clusters; The S44 comprises the following sub-steps: S441, taking a cluster feature vector as a comparison vector, setting a flag bit for each of the other cluster feature vectors, where the initial value of the flag bit is 0; S442, taking a cluster feature from the comparison vector as a comparison feature; S443, calculating the similarity between the comparison feature and the cluster features in other cluster feature vectors; S444, determine whether the similarity is greater than the similarity threshold. If so, add 1 to the flag bits of other cluster feature vectors whose similarity is greater than the similarity threshold, and jump to S442 to select the next cluster feature until all cluster features in the comparison vector are traversed. If not, directly jump to S442 to select the next cluster feature until all cluster features in the comparison vector are traversed. S445, using the flag bit as the matching degree between the cluster feature vector and another cluster feature vector; The S5 comprises the following sub-steps: S51, matching clusters on each target area to obtain multiple sets of matching cluster sets, and selecting the best matching cluster set from the multiple sets of matching cluster sets; S52, retaining edge pixels for each matching cluster in the optimal matching cluster set to obtain a cluster outline; S53, extracting feature points for each cluster outline; S54, performing fitting processing on multiple feature points corresponding to the optimal matching cluster set to obtain the action trajectory of the target object; The S51 comprises the following sub-steps: S511, calculating the similarity between the jth cluster feature on a target area and each cluster feature on other target areas, where the initial value of j is 1 and j is a positive integer; S512, extracting the cluster corresponding to the maximum similarity greater than the similarity threshold from each other target area, obtaining the matching cluster on each other target area, and forming the jth matching cluster set with each matching cluster; S513, adding the similarities corresponding to the matching clusters in the j-th matching cluster set to obtain the j-th similarity value; S514, add 1 to j and jump to step S511 until all cluster features in the target area are traversed; S515, selecting the matching cluster set with the largest similarity value as the optimal matching cluster set; The S53 comprises the following sub-steps: S531, calculating the geometric center coordinate point of each cluster outline according to the coordinates of each pixel point on each cluster outline; S532, taking each pixel point on the cluster outline as a comparison pixel point, and calculating the direction angle distribution value between the comparison pixel point and the geometric center coordinate point; S533: Take the comparison pixel point corresponding to the maximum direction angle distribution value as a feature point.

2. The target tracking method according to claim 1, It is characterized in that The formula for calculating the direction angle distribution value of the comparison pixel point and the geometric center coordinate point in S532 is: , , , where γ is the directional angle distribution value, e is a natural constant, and θ c To compare the direction angle between the pixel point and the geometric center coordinate point, θ m To compare the direction angle between the mth pixel point in the pixel point neighborhood and the geometric center coordinate point, x c is the horizontal coordinate of the comparison pixel, y c is the ordinate of the comparison pixel, x o is the horizontal coordinate of the geometric center coordinate point, y o is the ordinate of the geometric center coordinate point, x m is the horizontal coordinate of the mth pixel in the neighborhood of the comparison pixel on the cluster contour, y m is the ordinate of the mth pixel in the neighborhood of the comparison pixel on the cluster contour, | | is the absolute value operation, m is a positive integer, M is the number of neighborhood pixels, and arctan is the inverse tangent function.