Football player tracking method based on improved FairMot
By improving the FairMot model and introducing Cascaded Sparse Query and Query compensation mechanisms, the problems of high time cost and low accuracy of detection algorithms in football player tracking tasks are solved, and more efficient and accurate target tracking effects are achieved.
Patent Information
- Application Number
- CN202411727838.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-04-11
AI Technical Summary
The existing target detection and target tracking methods have problems such as high time cost, waste of computing resources, low accuracy of detection algorithms and not suitable for football player tracking tasks.
Using a football player tracking method based on improved FairMot, the image scaling strategy is adjusted to maximize the initial resolution of the input frame, and the Cascaded Sparse Query mechanism and Query compensation mechanism are incorporated to improve the accuracy and efficiency of target area prediction.
Improve the speed and accuracy of the target tracking algorithm, effectively utilize high-quality football video resources, and enhance the robustness of the model, especially when players move at high speed and camera lenses.
Smart Images
Figure CN119810145A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning target detection and target tracking, and in particular to a football player tracking method based on improved FairMot. Background Art
[0002] As one of the most popular sports in the world, football has a broad mass base, huge influence and extremely high commercial value.
[0003] The football player tracking task has a wide range of application scenarios and can be used for player and team performance scoring, tactical analysis, video assistant refereeing and other related tasks.
[0004] Compared with the mainstream pedestrian target tracking task, the football player tracking task is more challenging for the following reasons: first, the targets in football scenes are more frequently and severely occluded. For example, in a corner kick attack scenario in a football match, a large number of players will gather in the small penalty area, accompanied by a lot of fierce physical confrontation, causing mutual occlusion between players; second, the movement of players on the field is unpredictable and irregular, and the camera lens may also move constantly due to changes in the game situation, which will cause the players to become extremely blurred in certain frames of the video; finally, individual players may frequently leave and reappear in the camera. Therefore, most of the existing target detection and target tracking methods are not suitable for football player tracking tasks.
[0005] Directly inputting high-resolution and high-frame-rate video resources into the tracking model is one of the most effective ways to improve tracking performance. The massive number of commercial football matches around the world also brings sufficient high-quality football video resources.
[0006] However, such high-resolution and high-frame-rate videos are often accompanied by a high time cost when used as input for target tracking algorithms, especially for small targets such as players. Such targets are usually dispersed in the image and only occupy a very small part of the image. For general target detection models, a lot of computing resources will be wasted on such meaningless areas. Introducing the Cascaded Sparse Query CSQ mechanism in QueryDet is an effective way to solve the above problems. The Cascaded Sparse Query CSQ mechanism first uses low-resolution feature maps to predict the target area, and then continues to upsample the predicted area to obtain high-resolution feature maps and generate features. The CSQ mechanism effectively utilizes the sparsity of the target in the image, while maintaining the accuracy of the detection algorithm, it speeds up the detection by avoiding meaningless calculations in the background area.
[0007] It makes it possible to efficiently utilize high-quality video resources.
[0008] However, the Cascaded Sparse Query CSQ mechanism has a serious drawback in tracking tasks. The tracking task amplifies the importance of the Cascaded Sparse Query CSQ mechanism in predicting the target area, but this is accompanied by more unfavorable factors, such as the ambiguity problem mentioned above. If the area containing the target cannot be accurately predicted, this will not only lead to the lack of part of the target detection results, but also mean that this part of the area will be absent from the subsequent process of the tracking task, ultimately resulting in the inability to effectively implement methods such as ByteTrack.
[0009] Therefore, specific techniques are needed to improve the accuracy of the predicted regions by the Cascaded Sparse Query CSQ mechanism, or to supplement its prediction results. Summary of the invention
[0010] In view of the above-mentioned problems, the present invention is proposed.
[0011] Therefore, the technical problem solved by the present invention is that the existing target detection and target tracking methods have the problems of high time cost, large waste of computing resources, low detection algorithm accuracy, and unsuitability for football player tracking tasks.
[0012] To solve the above technical problems, the present invention provides the following technical solutions: a football player tracking method based on improved FairMot, comprising collecting football player tracking data for data annotation and constructing a data set; constructing an initial FairMot football player tracking model; constructing an improved FairMot football player tracking model and training the model parameters.
[0013] As a preferred solution of the football player tracking method based on improved FairMot described in the present invention, wherein: the collecting of football player tracking data for data annotation and constructing a data set includes selecting football video clips;
[0014] Data annotation of selected video clips;
[0015] After completing the data annotation, the video clips are divided into training set, test set and validation set according to the ratio of 6:2:2.
[0016] As a preferred solution of the football player tracking method based on improved FairMot described in the present invention, wherein: the construction of the initial FairMot football player tracking model includes the initial FairMot tracking model following the tracking by detection target tracking paradigm, including an FPN backbone network, a target detection branch, and a re-identification branch;
[0017] After generating the target detection results and target appearance features of the current frame, the model will enter the similarity output and data association stage. Using the generated target detection results and target appearance features, the similarity between the output detection results and each track is calculated, and a two-stage matching strategy is adopted to match the track with the target detection results.
[0018] As a preferred solution of the football player tracking method based on improved FairMot of the present invention, wherein: said constructing an improved FairMot football player tracking model and training the model parameters includes constructing an improved FairMot football player tracking model and training the improved FairMot football player tracking model parameters;
[0019] Adjust the image scaling strategy in the data preprocessing stage of the FairMot tracking model to maximize the initial resolution of the input frame;
[0020] Integrate the Cascaded Sparse Query mechanism into the FairMot target tracking model;
[0021] A query compensation mechanism is added based on the Cascaded Sparse Query mechanism.
[0022] As a preferred solution of the football player tracking method based on improved FairMot described in the present invention, wherein: the construction of the improved FairMot football player tracking model and the training of model parameters include adding the Cascaded SparseQuery mechanism to the FairMot target tracking model, including:
[0023] The initial object detection branch consists of the heatmap subnet, the box size subnet, and the center offset subnet, and the query subnet is added;
[0024] Starting from layer N+1, the query subnet will be activated, and the original ordinary convolution calculations of the target detection branch and the re-identification branch in layer N and beyond will be replaced by sparse convolution calculations;
[0025] The query subnet generates a query map through convolution calculation. The value in the query map grid (i, j) represents the probability that the target exists in the area corresponding to the (i, j) coordinate in the input image.
[0026] The grid coordinates (i, j) greater than the set threshold of 0.8 will be transmitted to the next layer. After being supplemented by the query supplement mechanism, they will be used as active sites to transform the feature map generated by the backbone network into a sparse feature map and participate in subsequent sparse convolution calculations.
[0027] As a preferred solution of the football player tracking method based on improved FairMot described in the present invention, wherein: the construction of the improved FairMot football player tracking model and the training of model parameters include a Query supplement mechanism including:
[0028] In the Kth frame, collect the grid coordinates whose Query subnet values are greater than 0.9;
[0029] In the K+1th frame, the values stored in the Kth frame are used to supplement the results predicted by the Query subnet, and a red-black tree is needed to filter out duplicate coordinates.
[0030] As a preferred solution of the football player tracking method based on improved FairMot of the present invention, the step of constructing the improved FairMot football player tracking model and training the model parameters includes using a red-black tree to filter repeated coordinates, including:
[0031] In the K+1 frame, the results predicted by the Query subnet and the coordinates collected in the Kth frame are inserted into the red-black tree in two stages;
[0032] In the first stage, the results predicted by the Query subnet are inserted into the red-black tree;
[0033] In the second stage, the coordinates collected by the query supplement mechanism are inserted into the red-black tree in sequence. Before each insertion, it is necessary to query whether the red-black tree has already stored the current coordinates;
[0034] If it exists, choose to skip this insertion;
[0035] After completing the second stage of insertion, the coordinates stored in the red-black tree are the final results of the coordinates.
[0036] Another object of the present invention is to provide a football player tracking system based on an improved FairMot, which can solve the problem that the current xxxx technology contains xxx through one of the solutions xxxxx.
[0037] As a preferred solution of the football player tracking system based on improved FairMot described in the present invention, it includes: an initialization module, a model building module, and a performance evaluation module; the initialization module is used to build a data set for football player tracking; the model building module is used to build an initial FairMot football player tracking model and an improved FairMot football player tracking model; the performance evaluation module is used to perform performance evaluation and continuously adjust and improve the model according to the results.
[0038] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a football player tracking method based on an improved FairMot.
[0039] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a football player tracking method based on improved FairMot.
[0040] Beneficial effects of the present invention: The football player tracking method based on the improved FairMot provided by the present invention provides richer visual information through high-resolution images. Selecting high-resolution images as model input is the most effective and simple means to improve the detection of small targets such as players. Therefore, the present invention constructs a high-resolution data set and adjusts the image scaling strategy in the data preprocessing stage of the FairMot tracking model to maximize the initial resolution of the input frame, rather than greatly compressing the initial resolution of the input image for the sake of maintaining the model speed. The Cascaded Sparse Query mechanism is added, and the low-resolution feature map is used to predict the area where the target exists. The predicted area is continuously sampled to obtain a high-resolution feature map and generate features. In this way, it is possible to avoid wasting computing resources in a large number of background areas where there are no targets. While maintaining the accuracy of the target tracking algorithm, the speed of the target tracking algorithm is improved, making it possible to use high-quality football video resources. The Query compensation mechanism is added to supplement the coordinates predicted by the Cascaded Sparse Query mechanism. By making full use of the potential information of past frames in the target tracking process to supplement the coordinates predicted in the current frame, the accuracy of the target position prediction by the Cascaded Sparse Query mechanism is improved at a very low time cost. This alleviates the target blur problem caused by camera lens movement and high-speed player motion to a certain extent, and enhances the robustness of the target tracking model. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0042] Figure 1 An overall flow chart of a football player tracking method based on improved FairMot provided in the first embodiment of the present invention.
[0043] Figure 2A structural diagram of an initial FairMot football player tracking model of a football player tracking method based on an improved FairMot is provided in a second embodiment of the present invention.
[0044] Figure 3 A structural diagram of an improved FairMot football player tracking model of a football player tracking method based on an improved FairMot is provided in accordance with a second embodiment of the present invention.
[0045] Figure 4 An overall flow chart of a football player tracking system based on improved FairMot provided in the third embodiment of the present invention. DETAILED DESCRIPTION
[0046] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.
[0047] Example 1, reference Figure 1 Figure 3 , as an embodiment of the present invention, provides a football player tracking method based on improved FairMot, comprising:
[0048] S1: Collect football player tracking data for data annotation and build a dataset.
[0049] Furthermore, football player tracking data is collected, and a sufficient number of football video clips with a resolution of (1920, 1080), a host broadcasting position and a duration of 60 seconds are selected.
[0050] The selected football video clips are labeled with data. In addition to labeling general information such as frame number and target category, event types are also added; there are a total of six event types: "substitute", "foul", "offside", "corner kick", "free kick" and "others"; due to changes in the situation on the field, different frames of the same video clip may have different labeled events.
[0051] Thanks to the widespread popularity of football, a large number of commercial football matches around the world ensure sufficient football video resources; without significantly reducing the input image resolution, choosing high-resolution and high-frame rate images as model input can effectively improve model performance.
[0052] High-resolution images provide richer visual information and are the most effective and simple way to improve the detection of small targets such as players.
[0053] Labeling different event types can enable researchers to have a more comprehensive understanding of the performance of the target tracking model, thereby adjusting the model parameters or further improving the model.
[0054] The annotated video clips are divided into training set, test set and validation set according to the ratio of 6:2:2.
[0055] S2: Build the initial FairMot soccer player tracking model.
[0056] Furthermore, we constructed the initial FairMot football player tracking model, referring to Figure 2 The initial FairMot tracking model includes the FPN backbone network, the target detection branch, and the re-identification branch;
[0057] The backbone network is responsible for extracting input image features and outputting feature maps with different resolutions to the target detection branch and the re-identification branch.
[0058] The target detection branch consists of a heatmap subnet, a box size subnet, and a center offset subnet, and is responsible for generating target detection results, including the location and category of the target.
[0059] The re-identification branch is responsible for generating the target appearance features of the current frame.
[0060] After obtaining the target detection results and target appearance features of the current frame, the model will enter the similarity calculation and data association stage; the FairMot tracking model adopts a two-stage association strategy similar to MOTDT.
[0061] In the first stage of association, the model uses a Kalman filter to predict the state of the trajectory in the current frame.
[0062] Calculate the Mahalanobis distance Dm between the Kalman filter prediction result and the target detection result.
[0063] The Mahalanobis distance Dm is fused with the appearance feature Dr, and the formula is as follows:
[0064] D=fDr+(1-f)Dm
[0065] Among them, f is the preset weight, which is 0.98.
[0066] The Hungarian algorithm was used for the first-stage association, and the threshold for the first-stage association was 0.4.
[0067] Start the second stage of association, which is based on the overlap between the target detection result and the trajectory; GIOU is selected as the metric to measure the overlap, and the threshold of the second stage association is 0.5.
[0068] The second stage association is based on the overlap between the target detection result and the trajectory, and GIOU is selected as the metric to measure the overlap.
[0069] Updates track appearance characteristics.
[0070] Initialize the unmatched detection results as new tracks.
[0071] Save the unmatched trajectories for 30 frames.
[0072] S3: Build an improved FairMot football player tracking model and train the model parameters.
[0073] Furthermore, we build an improved FairMot football player tracking model, referring to Figure 2 .
[0074] Adjust the image scaling strategy in the data preprocessing stage of the FairMot tracking model to maximize the initial resolution of the input frame. In order to maintain the model running speed, most models limit the size of the input image by default. If the size of the input image exceeds the preset upper limit, the input image will be reduced. Therefore, adjusting the image scaling strategy in the model preprocessing stage can avoid the high-resolution images input during the training and inference stages of the model from being significantly reduced.
[0075] Integrate the Cascaded Sparse Query mechanism into the FairMot target tracking model.
[0076] Add the Query subnet to the target detection branch in FairMot;
[0077] Starting from the N+1 layer, the Query subnet will be activated and responsible for predicting the area where the target exists. At the same time, the ordinary convolution calculations in the target detection branch and the re-identification branch of the Nth layer and later will be replaced by sparse convolution calculations. The Query subnet generates a Query map through convolution calculations (the N+1 layer is ordinary convolution calculations, and the N+1 layer and later are sparse convolution calculations). The value in the Query map grid (i, j) represents the probability that the target exists in the position corresponding to (i, j) in the input image. The grid coordinates (i, j) greater than the set threshold of 0.8 will be transmitted to the next layer. After being supplemented by the Query compensation mechanism, they will be used as active sites to convert the feature map generated by FPN into a sparse feature map and participate in subsequent sparse convolution calculations. The Cascaded Sparse Query mechanism can avoid wasting computing resources in a large number of background areas where there are no targets. Especially for high-resolution input images, while maintaining the accuracy of the target detection algorithm, it improves the speed of the target detection algorithm, making it possible to efficiently use high-quality football video resources.
[0078] After the coordinates predicted by the previous layer are transmitted to the next layer, they are first supplemented using the Query compensation mechanism; the current frame is marked as the K+1 frame, and grid coordinates with values greater than 0.9 are collected in the K-th frame; in the K+1 frame, the results predicted by the Query subnet and the coordinates collected in the K-th frame are inserted into the red-black tree in two stages; in the first stage, the results predicted by the Query subnet are inserted into the red-black tree; in the second stage, the coordinates collected by the Query supplement mechanism are inserted into the red-black tree in sequence. In this stage, before each insertion, it is necessary to query whether the red-black tree has already stored the current coordinates. If so, choose to skip this insertion; after completing the second stage of insertion, the coordinates stored in the red-black tree are the final results of the coordinates.
[0079] The purpose of adding the query compensation mechanism is to make full use of the potential information of past frames in the target tracking process to supplement the prediction results of the query subnet of the current frame. This can alleviate the target blur problem caused by camera movement, high-speed movement of players, etc. to a certain extent, and improve the accuracy of the Cascaded Sparse Query mechanism in predicting the target position at a very low time cost, thereby obtaining more accurate target detection results and enhancing the robustness of the target tracking model.
[0080] At the same time, as an efficient search tree, the red-black tree can query the existing duplicate coordinates more quickly, reducing the time cost of the query compensation mechanism to a certain extent.
[0081] After completing the construction of the improved FairMot football player tracking model, the improved FairMot football player tracking model is trained;
[0082] It should be noted that the trained parameters are loaded and the test set is input into the target tracking algorithm model for performance evaluation, and the model is continuously adjusted and improved based on the results.
[0083] Example 2, an embodiment of the present invention, provides a football player tracking method based on improved FairMot. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.
[0084] First of all, this experiment is divided into three models. The first model is the initial version of FairMot. The second model is the initial version of FairMot with Cascaded Sparse Query mechanism added. The third model includes Cascaded Sparse Query mechanism and Query compensation mechanism. The image scaling strategies in the preprocessing stage of the three models have been adjusted so that the initial resolution of the input frame can be maintained to the maximum extent.
[0085] Then, each model was trained. After pre-training on the COCO dataset, it was put into the constructed football player tracking training set for a total of 40 rounds of training to fine-tune the parameters. The final three models were obtained, in which the batch size was set to 12 and the initial learning rate was 10-4, which was gradually reduced to 10-5 in the first 20 rounds of training.
[0086] The present invention evaluates the performance of the invention through MOTA, HOTA, IDF1 and FPS indicators, where MOTA, HOTA and IDF1 all reflect the accuracy of the model, and FPS reflects the speed of the model. Table 1 shows the experimental results of the final generated model in the test set of the data set:
[0087] Table 1 Experimental data table
[0088]
[0089]
[0090] Comparing the second and first rows in the experimental results, by introducing the Cascaded Sparse Query mechanism, the FPS has been significantly improved, and the accuracy of the model has also been slightly improved. This shows that the Cascaded Sparse Query mechanism can avoid calculations that do not include the target area, greatly reducing the time cost of the FairMot algorithm. Compared with traditional algorithms, the FairMot model with Cascaded Sparse Query is more efficient in processing high-resolution images. Comparing the second and first rows in the experimental results, the Query compensation mechanism only reduces the FPS by 0.2, but brings stable performance improvements in various accuracy indicators. This shows that the Query compensation mechanism can effectively supplement the prediction results of FairMot during the tracking process.
[0091] Example 3, reference Figure 4 , which is an embodiment of the present invention, provides a football player tracking system based on improved FairMot, including an initialization module, a model building module, and a performance evaluation module.
[0092] The initialization module is used to construct a dataset for football player tracking; the model building module is used to build the initial FairMot football player tracking model and the improved FairMot football player tracking model; the performance evaluation module is used to perform performance evaluation and continuously adjust and improve the model based on the results.
[0093] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0094] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.
[0095] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.
[0096] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A soccer player tracking method based on improved FairMot, characterized in that: include: Collect football player tracking data for data annotation and build a dataset; Build the initial FairMot football player tracking model; Build an improved FairMot soccer player tracking model and train the model parameters.
2. The soccer player tracking method based on improved FairMot as claimed in claim 1, characterized in that: The collecting of football player tracking data, data annotation and constructing a data set includes selecting football video clips; Data annotation of selected video clips; After completing the data annotation, the video clips are divided into training set, test set and validation set according to the ratio of 6:2:
2.
3. The soccer player tracking method based on improved FairMot as claimed in claim 2, characterized in that: The constructing of the initial FairMot football player tracking model includes the initial FairMot tracking model following a tracking by detection target tracking paradigm, including an FPN backbone network, a target detection branch, and a re-identification branch; After generating the target detection results and target appearance features of the current frame, the model will enter the similarity output and data association stage. Using the generated target detection results and target appearance features, the similarity between the output detection results and each track is calculated, and a two-stage matching strategy is adopted to match the track with the target detection results.
4. The soccer player tracking method based on improved FairMot as claimed in claim 3, characterized in that: The constructing an improved FairMot football player tracking model and training model parameters includes constructing an improved FairMot football player tracking model and training the improved FairMot football player tracking model parameters; Adjust the image scaling strategy in the data preprocessing stage of the FairMot tracking model to maximize the initial resolution of the input frame; Integrate the Cascaded Sparse Query mechanism into the FairMot target tracking model; A query compensation mechanism is added based on the Cascaded Sparse Query mechanism.
5. The soccer player tracking method based on improved FairMot as claimed in claim 4, characterized in that: The construction of the improved FairMot football player tracking model and training of the model parameters includes adding the Cascaded Sparse Query mechanism to the FairMot target tracking model, including: The initial object detection branch consists of the heatmap subnet, the box size subnet, and the center offset subnet, and the query subnet is added; Starting from layer N+1, the query subnet will be activated, and the original ordinary convolution calculations of the target detection branch and the re-identification branch in layer N and beyond will be replaced by sparse convolution calculations; The query subnet generates a query map through convolution calculation. The value in the query map grid (i, j) represents the probability that the target exists in the area corresponding to the (i, j) coordinate in the input image. The grid coordinates (i, j) greater than the set threshold of 0.8 will be transmitted to the next layer. After being supplemented by the query supplement mechanism, they will be used as active sites to transform the feature map generated by the backbone network into a sparse feature map and participate in subsequent sparse convolution calculations.
6. The soccer player tracking method based on improved FairMot as claimed in claim 5, characterized in that: The construction of the improved FairMot football player tracking model and the training of model parameters include a Query supplement mechanism including: In the Kth frame, collect the grid coordinates whose Query subnet values are greater than 0.9; In the K+1th frame, the values stored in the Kth frame are used to supplement the results predicted by the Query subnet, and a red-black tree is needed to filter out duplicate coordinates.
7. The soccer player tracking method based on improved FairMot as claimed in claim 6, characterized in that: The construction of the improved FairMot football player tracking model and the training of model parameters include using a red-black tree to filter duplicate coordinates, including: In the K+1 frame, the results predicted by the Query subnet and the coordinates collected in the Kth frame are inserted into the red-black tree in two stages; In the first stage, the results predicted by the Query subnet are inserted into the red-black tree; In the second stage, the coordinates collected by the query supplement mechanism are inserted into the red-black tree in sequence. Before each insertion, it is necessary to query whether the red-black tree has already stored the current coordinates; If it exists, choose to skip this insertion; After completing the second stage of insertion, the coordinates stored in the red-black tree are the final results of the coordinates.
8. A system using the improved FairMot-based soccer player tracking method as claimed in any one of claims 1 to 7, characterized in that: Includes initialization module, model building module, and performance evaluation module; The initialization module is used to construct a dataset for football player tracking; The model building module is used to build an initial FairMot football player tracking model and an improved FairMot football player tracking model; The performance evaluation module is used to perform performance evaluation and continuously adjust and improve the model based on the results.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the football player tracking method based on improved FairMot are implemented as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the football player tracking method based on improved FairMot are implemented as described in any one of claims 1 to 7.