Behavior detection method and system based on target detection model

Abstract

The invention discloses a behavior detection method and a behavior detection system based on a target detection model, which are applied to the technical field of behavior recognition and are characterized in that an improved YOLOv10 model is utilized to fuse two one-to-one detection heads and one one-to-many detection head in a Head part, rich supervision signals of the one-to-many detection heads are utilized to train and optimize model parameters to ensure the precision of the model, and the accuracy of the model is improved. And one-to-one detection heads are used for inferring to avoid NMS post-processing, so that the inferring speed is increased. Wherein two one-to-one detection heads are utilized to obtain a background candidate frame and a target candidate frame respectively, the background candidate frame and the target candidate frame are processed through a behavior recognition model to obtain a first prediction map corresponding to a first prediction network, and the position of a first tracking target is marked on the background candidate frame according to the first prediction map; and determining the attitude of the first tracking target according to the change of the position of the first tracking target in the continuous frames. Therefore, the tracking of the tracking target on the background candidate box is realized.

Description

Technical Field

[0001] The present invention relates to the field of behavior recognition technology, and more specifically to a behavior detection method and system based on a target detection model. Background Art

[0002] In modern society, most young people need to work during the day, which may result in only elderly people over 65 years old being left at home. There are generally two reasons why the elderly are prone to falling. The first is due to their own illness, such as sudden coma caused by heart disease, or sudden coma caused by cerebral infarction, or amyotrophic lateral sclerosis. The other is physical injury caused by being tripped by objects at home due to not paying attention while walking.

[0003] However, the existing fall detection methods delay the fall detection of the monitored person due to the complex background environment, and may fail to identify the fall; even when the visual detection environment is not good, such as detecting light changes, the standing angle of the elderly, etc., the skeleton coordinates are prone to extraction errors or cannot be extracted. At this time, the behavior detection effect will be seriously affected, resulting in inaccurate detection results; in addition, the skeleton coordinates can only reflect static information and cannot fully reflect the fall behavior, and its detection model has not been updated in a timely and effective manner, and has always remained in the initial modeling state. The detection and recognition accuracy of the detection model needs to be further improved.

[0004] Therefore, how to provide a behavior detection method and system based on a target detection model with high detection efficiency and accuracy is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the present invention provides a behavior detection method and system based on a target detection model to solve the problems of the prior art.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] The present invention discloses a behavior detection method based on a target detection model, comprising:

[0008] Acquire a captured target video, obtain a target video to be analyzed from the target video, divide the target video to be analyzed into N segments, and randomly extract a frame of image from each segment as an image to be identified, where N is an integer greater than 1;

[0009] An improved YOLOv10 model is constructed. The improved YOLOv10 model adds a one-to-one head to the traditional YOLOv10 model. During the training process, the two one-to-one heads and the one-to-many head are optimized together with the model, allowing the backbone and the neck to enjoy the rich supervision provided by the one-to-many assignment. During the reasoning process, the one-to-many head is discarded, and two one-to-one heads are used for prediction. Two candidate regions of different scales are obtained in each frame image; they are a background candidate box and a target candidate box; the target candidate box includes a first tracking target, and the background candidate box includes an image of the position of the first tracking target to be predicted, including the target candidate box;

[0010] The background candidate frame and the target candidate frame are input into a first feature extraction network in the behavior recognition model to obtain a first feature map of the background candidate frame and a second feature map of the target candidate frame;

[0011] Inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model to obtain a first prediction map corresponding to the first prediction network, wherein the first prediction network includes a series of M key point prediction units, and the first prediction map is used to indicate the predicted position of the first tracking target on the background candidate frame, where M is an integer greater than 1; marking the position of the first tracking target on the background candidate frame according to the first prediction map;

[0012] The posture of the first tracking target is determined according to the change of the position of the first tracking target in the continuous frames.

[0013] Preferably, in the above-mentioned behavior detection method based on the target detection model, the improved YOLOv10 model includes: adding a one-to-one detection head to the head network based on the traditional YOLOv10 model; the head network includes a first one-to-one detection head, a second one-to-one detection head and a one-to-many detection head;

[0014] The first one-to-one detection head and the second one-to-one detection head retain the same structure as the original one-to-many branch and adopt the same optimization target; during the training process, the three detection heads are optimized together with the model; during the reasoning process, the one-to-many detection head is discarded, and the first one-to-one detection head and the second one-to-one detection head are used for prediction to obtain the background candidate box and the target candidate box.

[0015] Preferably, in the above-mentioned behavior detection method based on the target detection model, the first feature map and the second feature map are input into the first prediction network in the behavior recognition model to obtain the first prediction map corresponding to the first prediction network, including:

[0016] Using each key point prediction unit in the first prediction network to process the input data, a third feature map of the target candidate frame and a fusion map corresponding to the background candidate frame are obtained, wherein the input data of the l+1th key point prediction unit is the fusion map and the third feature map obtained by the lth key point prediction unit, l∈{1,2,…M}, and the fusion map is used to indicate the predicted key points of the first tracking target on the background candidate frame;

[0017] An operation of reducing the number of channels is performed on the feature maps obtained by the Mth key point prediction unit to obtain the first prediction map.

[0018] Preferably, in the above-mentioned behavior detection method based on the target detection model, each key point prediction unit in the first prediction network is used to process the input data to obtain the third feature map of the target candidate box and the fusion map corresponding to the background candidate box, including:

[0019] Using the first convolution operation unit of the first key point prediction unit of the first prediction network to perform a convolution operation on the first target image to obtain a third feature map of the target candidate box;

[0020] Using the second convolution operation unit of the first key point prediction unit to perform a convolution operation on the second target image to obtain a fourth feature map of the background candidate box;

[0021] Using the cross-correlation operation unit in the first key point prediction unit, performing a cross-correlation operation on the third feature map and the fourth feature map to obtain a fusion map;

[0022] Among them, the first key point prediction unit is the sth key point prediction unit of the first prediction network, s∈{1,2,…M}, when s=1, the first target map is the first feature map, and the second target map is the second feature map; when s≠1, the first target map is the fusion map input to the first key point prediction unit, and the second target map is the third feature map input to the first key point prediction unit.

[0023] Preferably, in the above-mentioned behavior detection method based on the target detection model, before the cross-correlation operation unit in the first key point prediction unit is used to perform a cross-correlation operation on the third feature map and the fourth feature map to obtain a fusion map, the method further includes:

[0024] Using the first convolution operation unit, inputting the third feature map into a third convolution operation unit in the first key point prediction unit;

[0025] Using the third convolution operation unit to adjust the data size of the third feature map, and make the ratio between the adjusted data size of the third feature map and the data size of the fourth feature map within a predetermined range;

[0026] The third feature map after data size adjustment is input into the cross-correlation operation unit.

[0027] Preferably, in the above-mentioned behavior detection method based on the target detection model, a first one-to-one detection head and a second one-to-one detection head are added to the head network, a CIB module is introduced, and the kernel size of the second 3×3 depth convolution in the CIB module of the first one-to-one detection head is increased to 7×7 to obtain background candidate boxes.

[0028] On the other hand, the present invention discloses a behavior detection system based on a target detection model, and the method is applied, including:

[0029] An acquisition module acquires a captured target video, obtains a target video to be analyzed from the target video, divides the target video to be analyzed into N segments, and randomly extracts a frame of image from each segment as an image to be identified, wherein N is an integer greater than 1;

[0030] A construction module is used to construct an improved YOLOv10 model. The improved YOLOv10 model adds a one-to-one head to the traditional YOLOv10 model. During the training process, two one-to-one heads and one one-to-many head are optimized together with the model, allowing the backbone and the neck to enjoy the rich supervision provided by the one-to-many assignment. During the reasoning process, the one-to-many head is discarded, and two one-to-one heads are used for prediction. Two candidate regions of different scales are obtained in each frame image; they are a background candidate frame and a target candidate frame; the target candidate frame includes a first tracking target, and the background candidate frame includes an image of the position of the first tracking target to be predicted, including the target candidate frame;

[0031] An input module, wherein the background candidate frame and the target candidate frame are input into a first feature extraction network in the behavior recognition model to obtain a first feature map of the background candidate frame and a second feature map of the target candidate frame;

[0032] A processing module, inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model, obtaining a first prediction map corresponding to the first prediction network, wherein the first prediction network includes a series of M key point prediction units, and the first prediction map is used to indicate the predicted position of the first tracking target on the background candidate frame, where M is an integer greater than 1; marking the position of the first tracking target on the background candidate frame according to the first prediction map;

[0033] The output module determines the posture of the first tracking target according to the change of the position of the first tracking target in the continuous frames.

[0034] Preferably, in the above-mentioned behavior detection system based on the target detection model, in the construction module, the head network includes a first one-to-one detection head, a second one-to-one detection head and a one-to-many detection head;

[0035] The first one-to-one detection head and the second one-to-one detection head retain the same structure as the original one-to-many branch and adopt the same optimization target; during the training process, the three detection heads are optimized together with the model; during the reasoning process, the one-to-many detection head is discarded, and the first one-to-one detection head and the second one-to-one detection head are used for prediction to obtain the background candidate box and the target candidate box.

[0036] Preferably, in the above-mentioned behavior detection system based on the target detection model, a first one-to-one detection head and a second one-to-one detection head are added to the head network, a CIB module is introduced, and the kernel size of the second 3×3 depth convolution in the CIB module of the first one-to-one detection head is increased to 7×7 to obtain background candidate boxes.

[0037] It can be known from the above technical solution that compared with the prior art, the present invention discloses a behavior detection method and system based on a target detection model, which utilizes an improved YOLOv10 model in the Head part, creatively integrates two one-to-one detection heads and one-to-many detection heads, utilizes the rich supervision signals of the one-to-many detection heads to train and optimize the model parameters to ensure the accuracy of the model, and adopts the one-to-one detection head for inference to avoid NMS post-processing and thus improve the inference speed. Among them, two one-to-one detection heads are used to obtain the background candidate frame and the target candidate frame respectively, and the background candidate frame and the target candidate frame are processed by the behavior recognition model to obtain the first prediction map corresponding to the first prediction network. According to the first prediction map, the position of the first tracking target is marked on the background candidate frame; according to the change of the position of the first tracking target in the continuous frame, the posture of the first tracking target is determined. Thereby, the tracking target is tracked on the background candidate frame. In the face of different scenes, there is no need to determine different anchor box selection schemes, and the applicability of the target tracking method is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0039] Figure 1The improved YOLOv10 model structure diagram provided by the present invention;

[0040] Figure 2 This is a structural diagram of the behavior recognition model provided by the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The embodiment of the present invention discloses a behavior detection method based on a target detection model, such as Figure 1 As shown, including:

[0043] Acquire a captured target video, obtain a target video to be analyzed from the target video, divide the target video to be analyzed into N segments, and randomly extract a frame of image from each segment as an image to be identified, where N is an integer greater than 1;

[0044] An improved YOLOv10 model is constructed. The improved YOLOv10 model adds a one-to-one head to the traditional YOLOv10 model. During the training process, the two one-to-one heads and the one-to-many head are optimized together with the model, allowing the backbone and the neck to enjoy the rich supervision provided by the one-to-many assignment. During the reasoning process, the one-to-many head is discarded, and two one-to-one heads are used for prediction. Two candidate regions of different scales are obtained in each frame image; they are a background candidate box and a target candidate box; the target candidate box includes a first tracking target, and the background candidate box includes an image of the position of the first tracking target to be predicted, including the target candidate box;

[0045] The background candidate frame and the target candidate frame are input into a first feature extraction network in the behavior recognition model to obtain a first feature map of the background candidate frame and a second feature map of the target candidate frame;

[0046] Inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model to obtain a first prediction map corresponding to the first prediction network, wherein the first prediction network includes a series of M key point prediction units, and the first prediction map is used to indicate the predicted position of the first tracking target on the background candidate frame, where M is an integer greater than 1; marking the position of the first tracking target on the background candidate frame according to the first prediction map;

[0047] The posture of the first tracking target is determined according to the change of the position of the first tracking target in the continuous frames.

[0048] Furthermore, the key points of the posture of the first tracking target are detected to obtain the candidate box information and 17 key points, each of which consists of three parameters: x, y and c; if the confidence of a key point is lower than the preset confidence threshold, the key point is considered to be occluded; the coordinates of the candidate box, the coordinates of the 17 key points, and the occlusion of the key points are stored in the database of the trajectory; if other trajectories are identified, the current trajectory is matched with the other trajectories, and if the current trajectory matches the other trajectories successfully, the two trajectories are merged; if the trajectory fails to match the other trajectories, the current trajectory is saved in the historical trajectory; according to actual needs, the 17 key points of the first tracking target detected and the corresponding confidence levels saved in the trajectory are used to realize key action recognition.

[0049] Based on human posture information (skeleton key points), the problem of trajectory mismatch caused by pedestrian occlusion during trajectory tracking is solved. At the same time, the amount of calculation is reduced, the running speed is improved, and a more robust, real-time and scalable method is provided.

[0050] In order to further optimize the above technical solution, the improved YOLOv10 model includes: adding a one-to-one detection head to the head network based on the traditional YOLOv10 model; the head network includes a first one-to-one detection head, a second one-to-one detection head and a one-to-many detection head;

[0051] The first one-to-one detection head and the second one-to-one detection head retain the same structure as the original one-to-many branch and adopt the same optimization target; during the training process, the three detection heads are optimized together with the model; during the reasoning process, the one-to-many detection head is discarded, and the first one-to-one detection head and the second one-to-one detection head are used for prediction to obtain the background candidate box and the target candidate box.

[0052] In order to further optimize the above technical solution, the first feature map and the second feature map are input into the first prediction network in the behavior recognition model to obtain a first prediction map corresponding to the first prediction network, including:

[0053] Using each key point prediction unit in the first prediction network to process the input data, a third feature map of the target candidate frame and a fusion map corresponding to the background candidate frame are obtained, wherein the input data of the l+1th key point prediction unit is the fusion map and the third feature map obtained by the lth key point prediction unit, l∈{1,2,…M}, and the fusion map is used to indicate the predicted key points of the first tracking target on the background candidate frame;

[0054] An operation of reducing the number of channels is performed on the feature maps obtained by the Mth key point prediction unit to obtain the first prediction map.

[0055] In order to further optimize the above technical solution, each key point prediction unit in the first prediction network is used to process the input data to obtain a third feature map of the target candidate frame and a fusion map corresponding to the background candidate frame, including:

[0056] Using the first convolution operation unit of the first key point prediction unit of the first prediction network to perform a convolution operation on the first target image to obtain a third feature map of the target candidate box;

[0057] Using the second convolution operation unit of the first key point prediction unit to perform a convolution operation on the second target image to obtain a fourth feature map of the background candidate box;

[0058] Using the cross-correlation operation unit in the first key point prediction unit, performing a cross-correlation operation on the third feature map and the fourth feature map to obtain a fusion map;

[0059] Among them, the first key point prediction unit is the sth key point prediction unit of the first prediction network, s∈{1,2,…M}, when s=1, the first target map is the first feature map, and the second target map is the second feature map; when s≠1, the first target map is the fusion map input to the first key point prediction unit, and the second target map is the third feature map input to the first key point prediction unit.

[0060] In order to further optimize the above technical solution, the method further includes: using the cross-correlation operation unit in the first key point prediction unit to perform a cross-correlation operation on the third feature map and the fourth feature map to obtain a fusion map.

[0061] Using the first convolution operation unit, inputting the third feature map into a third convolution operation unit in the first key point prediction unit;

[0062] Using the third convolution operation unit to adjust the data size of the third feature map, and make the ratio between the adjusted data size of the third feature map and the data size of the fourth feature map within a predetermined range;

[0063] The third feature map after data size adjustment is input into the cross-correlation operation unit.

[0064] Specifically, the number of first prediction networks in the behavior recognition model is N, where N is a positive integer.

[0065] Let's take N=3 as an example and combine Figure 2 Describe the embodiments of the present disclosure.

[0066] like Figure 2 As shown, the target candidate box is input into the first feature extraction subnetwork to obtain three first feature maps of the target candidate box, and the background candidate box is input into the second feature extraction subnetwork to obtain three second feature maps of the background candidate box.

[0067] The three first feature maps are respectively input into the three first prediction networks, and the three second feature maps are respectively input into the three first prediction networks to obtain the first prediction maps corresponding to each first prediction network. Since there are three first prediction networks in total, three first prediction maps in total are obtained.

[0068] Then, the three first prediction heat maps are added together using an addition operation unit to obtain a first target prediction map.

[0069] Then, according to the first target prediction map, the position of the first tracking target is marked on the background candidate frame.

[0070] Using each key point prediction unit in the first prediction network to process the input data, a third feature map of the target candidate frame and a fusion map corresponding to the background candidate frame are obtained, wherein the input data of the l+1th key point prediction unit is the fusion map and the third feature map obtained by the lth key point prediction unit, l∈{1,2,…M}, and the fusion map is used to indicate the predicted key points of the first tracking target on the background candidate frame;

[0071] The fusion graph obtained by the Mth key point prediction unit is respectively subjected to operations for reducing the number of channels to obtain a first prediction graph.

[0072] The channel reduction operation unit is used to perform an operation of reducing the number of channels on the response graph to obtain a first prediction graph.

[0073] For example, if the fusion image has 512 channels, the channel reduction operation unit performs a channel reduction operation on the fusion image with 512 channels to obtain a first prediction image with 5 channels. By performing the channel reduction operation, a first prediction image with a smaller number of channels is obtained. Thus, the position of the first tracking target on the background candidate box can be obtained according to the first prediction image.

[0074] Optionally, a single channel reduction operation unit includes P convolution operation units, where P is a positive integer.

[0075] The channel reduction operation unit is used to reduce the number of channels of the fusion image to obtain the first prediction image.

[0076] The method may include: performing P convolution operations on the fusion graph to reduce the number of channels by using P convolution operation units to obtain a prediction graph, wherein the number of channels of the prediction graph may be equal to the number of channels of the label graph used when training the behavior recognition model.

[0077] In order to further optimize the above technical solution, the first one-to-one detection head and the second one-to-one detection head are added to the head network, and the CIB module is introduced. The kernel size of the second 3×3 depth convolution in the CIB module of the first one-to-one detection head is increased to 7×7 to obtain the background candidate box.

[0078] Another embodiment of the present invention discloses a behavior detection system based on a target detection model, applying the above method, including:

[0079] An acquisition module acquires a captured target video, obtains a target video to be analyzed from the target video, divides the target video to be analyzed into N segments, and randomly extracts a frame of image from each segment as an image to be identified, wherein N is an integer greater than 1;

[0080] A construction module is used to construct an improved YOLOv10 model. The improved YOLOv10 model adds a one-to-one head to the traditional YOLOv10 model. During the training process, two one-to-one heads and one one-to-many head are optimized together with the model, allowing the backbone and the neck to enjoy the rich supervision provided by the one-to-many assignment. During the reasoning process, the one-to-many head is discarded, and two one-to-one heads are used for prediction. Two candidate regions of different scales are obtained in each frame image; they are a background candidate frame and a target candidate frame; the target candidate frame includes a first tracking target, and the background candidate frame includes an image of the position of the first tracking target to be predicted, including the target candidate frame;

[0081] An input module, wherein the background candidate frame and the target candidate frame are input into a first feature extraction network in the behavior recognition model to obtain a first feature map of the background candidate frame and a second feature map of the target candidate frame;

[0082] A processing module, inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model, obtaining a first prediction map corresponding to the first prediction network, wherein the first prediction network includes a series of M key point prediction units, and the first prediction map is used to indicate the predicted position of the first tracking target on the background candidate frame, where M is an integer greater than 1; marking the position of the first tracking target on the background candidate frame according to the first prediction map;

[0083] The output module determines the posture of the first tracking target according to the change of the position of the first tracking target in the continuous frames.

[0084] In order to further optimize the above technical solution, in the construction module, the head network includes a first one-to-one detection head, a second one-to-one detection head and a one-to-many detection head;

[0085] The first one-to-one detection head and the second one-to-one detection head retain the same structure as the original one-to-many branch and adopt the same optimization target; during the training process, the three detection heads are optimized together with the model; during the reasoning process, the one-to-many detection head is discarded, and the first one-to-one detection head and the second one-to-one detection head are used for prediction to obtain the background candidate box and the target candidate box.

[0086] In order to further optimize the above technical solution, the first one-to-one detection head and the second one-to-one detection head are added to the head network, and the CIB module is introduced. The kernel size of the second 3×3 depth convolution in the CIB module of the first one-to-one detection head is increased to 7×7 to obtain the background candidate box.

[0087] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0088] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A behavior detection method based on a target detection model, characterized in that: include: Acquire a captured target video, obtain a target video to be analyzed from the target video, divide the target video to be analyzed into N segments, and randomly extract a frame of image from each segment as an image to be identified, where N is an integer greater than 1; An improved YOLOv10 model is constructed. The improved YOLOv10 model adds a one-to-one head to the traditional YOLOv10 model. During the training process, the two one-to-one heads and the one-to-many head are optimized together with the model, allowing the backbone and the neck to enjoy the rich supervision provided by the one-to-many assignment. During the reasoning process, the one-to-many head is discarded, and two one-to-one heads are used for prediction. Two candidate regions of different scales are obtained in each frame image; they are a background candidate box and a target candidate box; the target candidate box includes a first tracking target, and the background candidate box includes an image of the position of the first tracking target to be predicted, including the target candidate box; The background candidate frame and the target candidate frame are input into a first feature extraction network in the behavior recognition model to obtain a first feature map of the background candidate frame and a second feature map of the target candidate frame; Inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model to obtain a first prediction map corresponding to the first prediction network, wherein the first prediction network includes a series of M key point prediction units, and the first prediction map is used to indicate the predicted position of the first tracking target on the background candidate frame, where M is an integer greater than 1; marking the position of the first tracking target on the background candidate frame according to the first prediction map; The posture of the first tracking target is determined according to the change of the position of the first tracking target in the continuous frames.

2. The behavior detection method based on the target detection model according to claim 1, characterized in that: The improved YOLOv10 model includes: adding a one-to-one detection head to the head network based on the traditional YOLOv10 model; the head network includes a first one-to-one detection head, a second one-to-one detection head and a one-to-many detection head; The first one-to-one detection head and the second one-to-one detection head retain the same structure as the original one-to-many branch and adopt the same optimization target; during the training process, the three detection heads are optimized together with the model; during the reasoning process, the one-to-many detection head is discarded, and the first one-to-one detection head and the second one-to-one detection head are used for prediction to obtain the background candidate box and the target candidate box.

3. The behavior detection method based on the target detection model according to claim 1, characterized in that: Inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model to obtain a first prediction map corresponding to the first prediction network includes: Using each key point prediction unit in the first prediction network to process the input data, a third feature map of the target candidate frame and a fusion map corresponding to the background candidate frame are obtained, wherein the input data of the l+1th key point prediction unit is the fusion map and the third feature map obtained by the lth key point prediction unit, l∈{1,2,…M}, and the fusion map is used to indicate the predicted key points of the first tracking target on the background candidate frame; An operation of reducing the number of channels is performed on the feature maps obtained by the Mth key point prediction unit to obtain the first prediction map.

4. The behavior detection method based on the target detection model according to claim 1, characterized in that: Using each key point prediction unit in the first prediction network to process the input data to obtain a third feature map of the target candidate box and a fusion map corresponding to the background candidate box, including: Using the first convolution operation unit of the first key point prediction unit of the first prediction network to perform a convolution operation on the first target image to obtain a third feature map of the target candidate box; Using the second convolution operation unit of the first key point prediction unit to perform a convolution operation on the second target image to obtain a fourth feature map of the background candidate box; Using the cross-correlation operation unit in the first key point prediction unit, performing a cross-correlation operation on the third feature map and the fourth feature map to obtain a fusion map; Among them, the first key point prediction unit is the sth key point prediction unit of the first prediction network, s∈{1,2,…M}, when s=1, the first target map is the first feature map, and the second target map is the second feature map; when s≠1, the first target map is the fusion map input to the first key point prediction unit, and the second target map is the third feature map input to the first key point prediction unit.

5. The behavior detection method based on the target detection model according to claim 4 is characterized in that: Before performing a cross-correlation operation on the third feature map and the fourth feature map by using the cross-correlation operation unit in the first key point prediction unit to obtain a fusion map, the method further includes: Using the first convolution operation unit, inputting the third feature map into a third convolution operation unit in the first key point prediction unit; Using the third convolution operation unit to adjust the data size of the third feature map, and make the ratio between the adjusted data size of the third feature map and the data size of the fourth feature map within a predetermined range; The third feature map after data size adjustment is input into the cross-correlation operation unit.

6. The behavior detection method based on the target detection model according to claim 1, characterized in that: The first one-to-one detection head and the second one-to-one detection head are added to the head network, and the CIB module is introduced. The kernel size of the second 3×3 depth convolution in the CIB module of the first one-to-one detection head is increased to 7×7 to obtain the background candidate box.

7. A behavior detection system based on a target detection model, applying the behavior detection method based on a target detection model according to any one of claims 1 to 6, characterized in that: include: An acquisition module acquires a captured target video, obtains a target video to be analyzed from the target video, divides the target video to be analyzed into N segments, and randomly extracts a frame of image from each segment as an image to be identified, wherein N is an integer greater than 1; A construction module is used to construct an improved YOLOv10 model. The improved YOLOv10 model adds a one-to-one head to the traditional YOLOv10 model. During the training process, two one-to-one heads and one one-to-many head are optimized together with the model, allowing the backbone and the neck to enjoy the rich supervision provided by the one-to-many assignment. During the reasoning process, the one-to-many head is discarded, and two one-to-one heads are used for prediction. Two candidate regions of different scales are obtained in each frame image; they are a background candidate frame and a target candidate frame; the target candidate frame includes a first tracking target, and the background candidate frame includes an image of the position of the first tracking target to be predicted, including the target candidate frame; An input module, wherein the background candidate frame and the target candidate frame are input into a first feature extraction network in the behavior recognition model to obtain a first feature map of the background candidate frame and a second feature map of the target candidate frame; A processing module, inputting the first feature map and the second feature map into a first prediction network in the behavior recognition model, obtaining a first prediction map corresponding to the first prediction network, wherein the first prediction network includes a series of M key point prediction units, and the first prediction map is used to indicate the predicted position of the first tracking target on the background candidate frame, where M is an integer greater than 1; marking the position of the first tracking target on the background candidate frame according to the first prediction map; The output module determines the posture of the first tracking target according to the change of the position of the first tracking target in the continuous frames.

8. The behavior detection system based on target detection model according to claim 7, characterized in that: In the construction module, the head network includes a first one-to-one detection head, a second one-to-one detection head and a one-to-many detection head; The first one-to-one detection head and the second one-to-one detection head retain the same structure as the original one-to-many branch and adopt the same optimization target; during the training process, the three detection heads are optimized together with the model; during the reasoning process, the one-to-many detection head is discarded, and the first one-to-one detection head and the second one-to-one detection head are used for prediction to obtain the background candidate box and the target candidate box.

9. The behavior detection system based on target detection model according to claim 7, characterized in that: The first one-to-one detection head and the second one-to-one detection head are added to the head network, and the CIB module is introduced. The kernel size of the second 3×3 depth convolution in the CIB module of the first one-to-one detection head is increased to 7×7 to obtain the background candidate box.

Images