Multi-view athlete tracking identification method for Kabadi competition

Abstract

The invention discloses a multi-view athlete tracking and identifying method for a Kabadi competition, which comprises the following steps of: synchronously acquiring videos of the Kabadi competition by adopting a plurality of camera devices with different view angles, and performing character target detection on each frame of video image after time alignment; tracking objects in the effective competition area are screened out, an SORT tracking algorithm is adopted to process the tracking objects, and a target frame and a short-time tracking ID of current frame data are obtained; inquiring whether an athlete is mapped in the short-time tracking ID in a mapping table, if so, marking the tracking object as the athlete, and returning to the first step when the competition is not finished; otherwise, entering the next step; extracting features of the corresponding image in the target frame by adopting a feature extraction network to obtain a feature vector, screening k matched identities in the athlete feature database based on the feature vector and returning confidence, then performing identity matching on the tracking object corresponding to the short-time tracking ID, and returning to the first step after the identity matching is finished and the competition is not finished.

Description

Technical Field

[0001] The present invention relates to a target tracking technology, in particular to a multi-viewing athlete tracking and identification method for a kabaddi game. Background Art

[0002] Kabaddi is a high-intensity, competitive team sport characterized by rapid and intense movements, often accompanied by strong physical collisions and occlusions. This makes it difficult for traditional computer vision methods to accurately and continuously track and identify people in these scenarios. To meet the needs of functional modules such as real-time match analysis, tactical replays, and violation detection, the system urgently needed an underlying technology that could stably identify and track athletes from multiple perspectives, ensuring high-precision tracking capabilities across the entire field.

[0003] In the field of multi-target tracking, the classic SORT algorithm uses Kalman filtering and the Hungarian algorithm for position-based inter-frame target matching. This algorithm offers high real-time performance, but is prone to tracking instability in scenarios with overlapping or occluded targets. To improve robustness, the DeepSORT algorithm incorporates appearance features to effectively mitigate ID drift caused by occlusion and short-term target disappearance. Although subsequent algorithms such as MOTDT have further optimized tracking accuracy, stable and continuous tracking of a specific individual remains difficult in environments with dense targets, rapid interactions, and frequent occlusions. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the multi-perspective athlete tracking and identification method for kabaddi competitions provided by the present invention solves the problem that the prior methods are difficult to stably and continuously track a specified individual.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A multi-view athlete tracking and identification method for a kabaddi game is provided, comprising the steps of: S1. Using multiple cameras with different viewing angles to synchronously capture videos of a kabaddi match, temporally aligning the video frames of the multiple cameras, and performing human target detection on each aligned video frame. S2. Filter the tracking objects within the valid competition area based on the human target detection results, and use the SORT tracking algorithm to process the tracking objects to obtain the target frame and short-term tracking ID of the current frame data; S3. Check in the mapping table whether the short-time tracking ID is mapped to an athlete. If so, mark the tracking object as the athlete and return to step S1 if the game is not over. Otherwise, go to step S4. S4. Using a feature extraction network to extract features of the corresponding image in the target frame, obtaining a feature vector, and screening k matching identities and corresponding confidence levels in the athlete feature database based on the feature vector; S5. Perform identity matching on the tracking object corresponding to the short-time tracking ID according to the k matching identities, corresponding confidence levels, and the mapping table. After the identity matching is completed and the game is not over, return to step S1.

[0006] The beneficial effects of the present invention are as follows: this solution fully analyzes and utilizes the deep features of athlete images captured by cameras with different perspectives, thereby improving the accuracy and stability of tracking and identification; when the athlete mapped to the short-term tracking ID is not found in the mapping table, this solution extracts feature vectors, combines them with the athlete feature database to determine multiple possible identities, and then further combines the confidence and mapping table for identity matching. Since the athlete feature database is an online updated, feature-generalized and fused athlete identity feature vector database, it takes into account short-term efficiency tracking, feature extraction and multi-perspective tracking with cross-perspective information fusion, effectively improving the recognition accuracy and tracking efficiency of the tracking and identification method in complex sports environments.

[0007] Furthermore, step S5 further includes: S51: Select the matching identity corresponding to the maximum confidence and check whether it matches the existing short-term tracking ID in the mapping table. If so, proceed to step S52; otherwise, mark the tracked object as a matching identity. S52: Determine whether the maximum confidence is greater than the confidence of the athlete corresponding to the short-term tracking ID. If so, proceed to step S53; otherwise, delete the matching identity corresponding to the maximum confidence and return to step S51. S53: Mark the tracking object as a matching identity, delete the mapping relationship between the existing short-term tracking ID in the mapping table, and return to step S1 when the game is not settled.

[0008] The beneficial effect of this technical solution is that the mapping table establishes a dynamic, one-to-one correspondence between the random, auto-incrementing ID generated by the short-term tracking algorithm and the athlete's stable identity ID, thereby achieving accurate identification and continuous tracking of athletes while ensuring identity consistency. Furthermore, this mechanism exhibits excellent scalability and anti-interference capabilities, making it particularly suitable for complex situations such as frequent occlusion of targets and rapid switching of perspectives in multi-player sports scenarios.

[0009] Furthermore, the multi-view athlete tracking and identification method for kabaddi competitions further includes determining whether an ID drift occurs in a tracking target corresponding to the camera device, and correcting the ID drift if it occurs: A1. Acquire time-aligned multi-view athlete tracking data and convert the image coordinates from the multi-view into athlete coordinate data in a standardized top-view coordinate system using a homography transformation method; A2. Collect the standardized coordinate data of each athlete from all perspectives, remove outliers, and weight the coordinate data of the athlete observed from multiple perspectives to obtain the athlete's current fused coordinates. A3. Determine whether the difference between the coordinate data under a single view and the fused coordinates exceeds a preset tolerance threshold. If so, proceed to step A4. Otherwise, the tracking target corresponding to the camera device does not experience ID drift. A4. The tracking target corresponding to the current camera device's viewing angle experiences ID drift, and the mapping relationship between the short-term tracking ID corresponding to the current camera device and the athlete in the mapping table is deleted.

[0010] The beneficial effects of the above technical solution are: by using the homography transformation method to uniformly map the coordinates of multi-camera data and designing an error correction feedback mechanism, the accuracy of athlete tracking is further improved. The ID drift is then determined by combining the fused coordinates, which solves the problem of identity loss and drift of athletes in rapidly changing confrontation scenarios, and further effectively improves the recognition accuracy and tracking efficiency of the tracking system in complex sports environments.

[0011] Furthermore, the expression converted into the standardized athlete coordinate data in step A1 is: , in, and These are the coordinate data of the athletes before and after normalization; is the homography transformation matrix; 、 and for The three elements in the first row of 、 and for The three elements in the second row; 、 and for The three elements in the third row; The expression of the athlete's fusion coordinates at the current moment is: in, and is the fusion coordinate of athlete i at the current moment; For athlete i at the current moment Standardized coordinate data of visual observation; For athlete i at the current moment The weight corresponding to the viewing angle; For athlete i at the current moment The coordinates of the camera device of the viewing angle; is the correction parameter; n is the total number of cameras.

[0012] The beneficial effects of the above technical solution are: by uniformly converting the image coordinates under multiple perspectives into a standardized top-view coordinate system, and fusing the multi-perspective information for weighted calculation, the weight distribution is based on the distance between the camera and the athlete. Cameras with closer distances are given higher weights due to their higher recognition accuracy. This not only significantly improves the accuracy of athlete position estimation, but also effectively identifies and locates identity drift problems caused by factors such as occlusion and perspective changes, and promptly feeds back abnormal information to trigger the lower-level camera system to perform identity correction.

[0013] Furthermore, after step S2 and before step S4, the player feature database is updated: B1. Based on the current real-time athlete tracking data, dynamically capture athlete image frames from the kabaddi game footage, extract time-sensitive appearance feature vectors based on the image frames, and store them in the athlete feature database; B2. Based on the appearance feature vector and the time-sensitive appearance feature vector in the basic appearance feature data, a weighted fusion algorithm is used to perform feature generalization processing and construct a fused feature representation vector: in, is the fusion feature representation vector of the athlete; m is the total number of appearance feature vectors in the basic appearance feature data; is the oth appearance feature vector of the athlete; for The corresponding weight; N is the total number of the time-sensitive appearance feature vectors of the athlete at the current moment; is the p-th temporal appearance feature vector of the athlete; for The corresponding weight; B3. Perform L2 normalization on the fusion feature representation vector of the athlete: in, is the quadratic norm; B4. The L2-normalized fusion feature representation vector of the athlete is stored in the athlete feature database.

[0014] The beneficial effects of the above technical solution are: this solution updates the database in the above manner, and uses the real-time athlete image data as timely athlete image data; this method effectively solves the static defects of the basic appearance feature database, and the use of timely feature data can effectively improve the accuracy of athlete recognition in complex sports; the update process supports continuous operation, realizing real-time supplementation and updating of athlete appearance features.

[0015] Furthermore, the method for updating the athlete feature database further includes: Calculate the weight of each time-sensitive appearance feature vector in the athlete feature database at the current moment: in, is the weight of each time-sensitive appearance feature vector at the current moment, is the hyperparameter of the decay rate; is the hyperparameter of the global temporal feature weight; For the current moment; The moment when each time-sensitive appearance feature vector is stored in the athlete feature database; When the weight of each time-sensitive appearance feature vector at the current moment is less than a preset removal threshold, it will be removed from the athlete feature database.

[0016] The beneficial effect of the above technical solution is: by gradually reducing the weight value of old feature data over time, and automatically removing data from the feature database when the feature weight value is lower than the preset clearing threshold, redundant occupation of system memory resources is prevented.

[0017] Furthermore, the method of screening k matching identities and corresponding confidence levels in the athlete feature database based on the feature vector includes: Calculate the cosine similarity between the feature vector and each fusion feature representation vector in the athlete feature database; The k fusion feature representation vectors with the largest cosine similarity are selected, and the corresponding athletes are used as the matching identities of the feature vectors, and the cosine similarity is used as the confidence of the matching identity.

[0018] Furthermore, the method for obtaining basic appearance feature data is as follows: In the initialization phase, image samples of each athlete at different positions in the kabaddi competition venue are manually pre-collected. The image samples include multi-scale perspective images of close distance, medium distance and long distance; The corresponding appearance feature vector is extracted for each image sample, and multiple appearance feature vectors of all athletes are used to form basic appearance feature data.

[0019] Furthermore, in the mapping table, the confidence level corresponding to each athlete with a short-term tracking ID decays linearly with the recording time. With this configuration, as the algorithm runs, the confidence level of each mapping item gradually decreases, making it more likely to be replaced by a new database recognition result in step S52, reducing persistent recognition conflicts and achieving the goal of regularly clearing long-term inactive mapping data.

[0020] Furthermore, the method of screening out tracking objects within the effective playing area according to the human target detection result includes: Constructing a valid playing area mask for kabaddi matches : in, is the coordinate of the tracked object; Each detected tracking object Calculate the intersection rate with the valid competition area mask: in, is the intersection rate; When the intersection rate is less than the empirical threshold, the corresponding tracking object Delete and filter the tracking objects within the valid competition area.

[0021] The beneficial effects of the above-mentioned further method are: based on the prior knowledge of the competition venue, it can eliminate the misdetected targets outside the venue or in the fuzzy boundary area in the image, improve the spatial consistency and target effectiveness of the detection results, reduce the misrecognition and mismapping phenomena in subsequent recognition and tracking, and improve the recognition accuracy and operation efficiency of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flowchart of the multi-view athlete tracking and recognition method for kabaddi competition.

[0023] Figure 2 This is a top view of the fixed position of the camera under an embodiment of a kabaddi game.

[0024] Figure 3 Detailed flow chart of the multi-view fusion tracking and identification method for athletes in the Kabaddi competition in this embodiment.

[0025] Figure 4 An example diagram of the mapping table designed for this solution.

[0026] Figure 5 This is a schematic diagram of the Kabaddi field. DETAILED DESCRIPTION

[0027] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0028] refer to Figure 1 , Figure 1 1 shows a flow chart of a multi-view athlete tracking and identification method for a kabaddi match; Figure 1 As shown, the method S includes steps S1 to S5. To facilitate understanding of this solution, Figure 3 A detailed flow chart of a multi-view fusion tracking and identification method for athletes in a Kabaddi game (ie, steps S1 to S5) is shown.

[0029] In step S1, a plurality of cameras with different viewing angles are used to synchronously capture a video of a kabaddi match, the video frames of the plurality of cameras are time-aligned, and human target detection is performed on each aligned video frame.

[0030] Before capturing kabaddi match videos, the camera parameters of all cameras were unified, the shooting positions of the cameras were fixed as required to capture videos from different perspectives, and the system time of all cameras was synchronized using the NTP protocol.

[0031] When capturing video of a kabaddi match, the acquisition timestamp of each frame is recorded synchronously, and the timestamp information corresponding to each frame is read. Based on the principle of nearest time matching, the video frames from different cameras are time-aligned to achieve time synchronization of multi-view images.

[0032] In this embodiment, the camera device (camera, video camera, etc. described below all refer to the camera device) is installed at the kabaddi competition site. Figure 2 Specifically, multiple cameras are evenly distributed around the kabaddi competition field to cover the entire competition area and ensure no blind spots. Each camera's shooting angle is optimized in advance to minimize data loss or recognition errors caused by player occlusion or lighting differences.

[0033] In this solution, the preferred camera device location and data configuration are as follows: All cameras are installed at a height of approximately 2.5 to 3 meters, avoiding the athletes' activity paths and ensuring clear and stable images; the cameras are fixedly installed and cannot be moved at will to ensure the stability of subsequent image calibration.

[0034] All cameras have built-in or external embedded computing units that support the NTP protocol. Their system time accesses the NTP server through the network. After synchronization is completed, each frame of image captured by the camera device is bound to the current system timestamp and recorded in the standard Unix time format (for example, millisecond precision timestamp = 1703487745193).

[0035] Figure 5 A schematic diagram of a kabaddi competition field is given. Based on the fixed layout of the competition field, a field prior mask region containing boundary information such as the side line and the baseline is constructed, and the effective competition area mask M is defined.

[0036] In step S2, the tracking objects within the effective game area are screened out according to the human target detection results, and the tracking objects are processed using the SORT tracking algorithm to obtain the target frame and short-term tracking ID of the current frame data.

[0037] like Figure 3 As shown in the figure, a batch of image data is taken from the "frame queue" image cache queue, and the "target detection" operation is first performed on the image. After basic screening such as non-maximum suppression, category, confidence, and target box aspect ratio, the target box data that basically meets the athlete's attributes is output; the target box data is then sent to the "valid area screening" process. The specific implementation process is as follows: Constructing a valid playing area mask for kabaddi matches : in, is the coordinate of the tracked object; Each detected tracking object Calculate the intersection rate with the valid competition area mask: in, is the intersection rate; When the intersection rate is less than the empirical threshold, the corresponding tracking object Delete and filter the tracking objects within the valid competition area.

[0038] This embodiment uses a target detection algorithm to mark the locations of all entities in the scene. After screening out targets that meet the athlete characteristics, the mask intersection rate is calculated. A low-time-complexity calculation method is used to efficiently screen out athlete targets within the effective field area, providing high-quality intermediate data output for subsequent long-term target tracking and recognition.

[0039] like Figure 3The "SORT tracker" module shown is the SORT tracking algorithm of this solution. It is a relatively mature SORT multi-target tracking algorithm. It is used as a short-term tracking module. For each tracked object screened out in step S2, the Kalman filter is used to predict its position in the current frame. Then, the IoU matrix is ​​constructed and the Hungarian algorithm is used to match the optimal target box and the predicted box to update the tracking status.

[0040] The SORT tracking algorithm automatically assigns a self-increasing "short-term tracking ID" to each target on the field. This allows for efficient tracking of the target's trajectory when it appears continuously, clearly, and stably within the field of view. However, significant ID drift occurs when the target is obscured or disappears and reappears, making stable tracking of the player difficult.

[0041] In step S3, the mapping table is queried to see if the short-term tracking ID is mapped to an athlete. If so, the tracking object is marked as the athlete, and the process returns to step S1 before the game ends; otherwise, the process goes to step S4.

[0042] The mapping table of this solution is as follows Figure 4 As shown in the figure, the mapping table of this scheme is an abbreviation of the injective function mapping table, where "short-term tracking ID" is the tracking ID output by the short-term target tracking module, which is self-incremental and unstable; "athlete ID" is the athlete identity ID pre-input before the game, which is invariant and stable. It is worth mentioning that it contains special Unknown elements for temporary mapping of characters whose confidence level is lower than the threshold or whose features have not been extracted and identified.

[0043] The "Confidence Record" on the right records the recognition confidence data corresponding to the mapped athlete IDs. This table retains the mapping entries with higher confidence levels to resolve identification conflicts. Furthermore, in the mapping table, the confidence level for each athlete with a short-term tracking ID decays linearly with the recording time. This prevents occasional high-confidence identification data from occupying the mapping entry for a long period of time, ensuring the table's updatability.

[0044] In order to maintain the information in the mapping table, the tracking status of all targets is divided into five standard states: "unidentified state", "conflict state", "tracking state", "ID drift state", and "lost state". The five states are defined as follows: Unidentified state: the short-time target tracking module has tracked it, but no mapping with the actual athlete identity has been established; if no athlete is mapped to the short-time tracking ID in step S3, the state at this time is unidentified.

[0045] Tracking status: Tracked by the short-term target tracking module (SORT tracking algorithm), the identity mapping with the actual athlete has been established, and long-term stable tracking is in progress; that is, the status of the athlete is mapped to the short-term tracking ID in step S3.

[0046] Conflict state: The target has been tracked by the short-term target tracking module, and a mapping conflict occurs when establishing the identity mapping, waiting for conflict resolution; in step S51, the matching identity corresponding to the maximum confidence is matched with the existing short-term tracking ID in the mapping table, and the state at this time is the conflict state.

[0047] Lost state: The short-term target tracking module loses tracking and waits for the target to reappear.

[0048] ID drift state: The top-level module reports ID drift and waits for the system to clear the mapping. This is the state when the difference between the coordinate data under a single perspective and the fused coordinates exceeds the preset tolerance threshold.

[0049] In step S4, a feature extraction network is used to extract features of the corresponding image in the target frame to obtain a feature vector, and k matching identities and corresponding confidence levels are selected from the athlete feature database based on the feature vector; During implementation, the method for selecting k matching identities and corresponding confidence levels preferably includes: Calculate the cosine similarity between the feature vector and each fusion feature representation vector in the athlete feature database; The k fusion feature representation vectors with the largest cosine similarity are selected, and the corresponding athletes are used as the matching identities of the feature vectors, and the cosine similarity is used as the confidence of the matching identity.

[0050] In step S5, identity matching is performed on the short-term tracking ID corresponding to the tracking object based on the k matching identities, corresponding confidence levels and the mapping table. After the identity matching is completed and the game is not over, the process returns to step S1.

[0051] In one embodiment of the present invention, step S5 further includes: S51. Select the matching identity corresponding to the maximum confidence and check whether it matches an existing short-term tracking ID in the mapping table. If so, proceed to step S52. Otherwise, mark the tracked object as a matching identity and update the mapping table. S52: Determine whether the maximum confidence is greater than the confidence of the athlete corresponding to the short-term tracking ID. If so, proceed to step S53; otherwise, delete the matching identity corresponding to the maximum confidence and return to step S51. S53: Mark the tracking object as a matching identity, delete the mapping relationship between the existing short-term tracking ID in the mapping table, and return to step S1 when the game is not settled.

[0052] In one embodiment of the present invention, the multi-view athlete tracking and identification method for kabaddi competition further includes determining whether an ID drift occurs in a tracking target corresponding to a camera device, and correcting the ID drift if it occurs. A1. Obtain time-aligned multi-view athlete tracking data and use homography to convert the image coordinates from the multi-view into the athlete coordinate data in a standardized top-view coordinate system: , in, and These are the coordinate data of the athletes before and after normalization; is the homography transformation matrix; 、 and for The three elements in the first row of 、 and for The three elements in the second row; 、 and for The three elements in the third row.

[0053] A2. Collect the standardized coordinate data of each athlete from all perspectives, remove outliers, and weight the coordinate data of the athlete observed from multiple perspectives to obtain the fused coordinates of the athlete at the current moment: in, and is the fusion coordinate of athlete i at the current moment; For athlete i at the current moment Standardized coordinate data of visual observation; For athlete i at the current moment The weight corresponding to the viewing angle; For athlete i at the current moment The coordinates of the camera device of the viewing angle; is the correction parameter; n is the total number of cameras.

[0054] A3. Determine whether the difference between the coordinate data under a single view and the fused coordinates exceeds a preset tolerance threshold. If so, proceed to step A4. Otherwise, the tracking target corresponding to the camera device does not experience ID drift. A4. The tracking target corresponding to the current camera device's viewing angle experiences ID drift, and the mapping relationship between the short-term tracking ID corresponding to the current camera device and the athlete in the mapping table is deleted.

[0055] After calculating the fusion coordinates, this method will check the coordinates of the players in each perspective, and calculate the two-dimensional distance between the coordinates of the players located in each sub-perspective and the fusion coordinates. If the difference exceeds the preset tolerance threshold, feedback data will be sent to the camera device; Figure 3 The "queue to be corrected" module shown is used to cache feedback data, waiting for correction in step A4. In this way, the advantages of the multi-view system are effectively utilized, and the recognition error caused by the limitation of a single view and the resulting ID drift are alleviated.

[0056] After step S2 and before step S4, the player feature database is also updated: B1. Based on the current real-time athlete tracking data, dynamically capture athlete image frames from the kabaddi game footage, extract time-sensitive appearance feature vectors based on the image frames, and store them in the athlete feature database; B2. Based on the appearance feature vector and the time-sensitive appearance feature vector in the basic appearance feature data, a weighted fusion algorithm is used to perform feature generalization processing and construct a fused feature representation vector: in, is the fusion feature representation vector of the athlete; m is the total number of appearance feature vectors in the basic appearance feature data; is the oth appearance feature vector of the athlete; for The corresponding weight; N is the total number of the time-sensitive appearance feature vectors of the athlete at the current moment; is the p-th temporal appearance feature vector of the athlete; for The corresponding weight; The method for obtaining basic appearance feature data is as follows: In the initialization phase, image samples of each athlete at different positions in the kabaddi competition venue are manually pre-collected. The image samples include multi-scale perspective images of close distance, medium distance and long distance; The corresponding appearance feature vector is extracted for each image sample, and multiple appearance feature vectors of all athletes are used to form basic appearance feature data.

[0057] B3. Perform L2 normalization on the fusion feature representation vector of the athlete: in, is the quadratic norm; B4. The L2-normalized fusion feature representation vector of the athlete is stored in the athlete feature database.

[0058] Through the above steps, this solution achieves efficient identification and dynamic management of athlete identities in kabaddi competitions. It has good generalization ability and real-time adaptability, and effectively improves the recognition accuracy in complex scenarios.

[0059] During implementation, the method for updating the athlete feature database preferably further includes: Calculate the weight of each time-sensitive appearance feature vector in the athlete feature database at the current moment: in, is the weight of each time-sensitive appearance feature vector at the current moment, is the hyperparameter of the decay rate; is the hyperparameter of the global temporal feature weight; For the current moment; The moment when each time-sensitive appearance feature vector is stored in the athlete feature database; When the weight of each time-sensitive appearance feature vector at the current moment is less than a preset removal threshold, it will be removed from the athlete feature database.

[0060] In summary, our method combines the SORT tracking algorithm to obtain short-term tracking IDs and features extracted by the feature extraction network, which can achieve long-term tracking management of athletes; constructs an athlete feature database with adaptive update capabilities, which can realize continuous and accurate identification of athlete targets in different motion states, occlusion levels and shooting perspectives; and realizes the athlete coordinate fusion and error correction algorithm of overlapping perspectives through the mathematical method of homography transformation, achieving stable identity tracking consistent with multiple cameras.

Claims

1. A multi-view athlete tracking and recognition method for kabaddi competition, characterized in that: Including steps: S1. Using multiple cameras with different viewing angles to synchronously capture videos of a kabaddi match, temporally aligning the video frames of the multiple cameras, and performing human target detection on each aligned video frame. S2. Filter the tracking objects within the valid competition area based on the human target detection results, and use the SORT tracking algorithm to process the tracking objects to obtain the target frame and short-term tracking ID of the current frame data; S3. Check in the mapping table whether the short-time tracking ID is mapped to an athlete. If so, mark the tracking object as the athlete and return to step S1 if the game is not over. Otherwise, go to step S4. S4. Using a feature extraction network to extract features of the corresponding image in the target frame, obtaining a feature vector, and screening k matching identities in the athlete feature database based on the feature vector and returning the corresponding confidence score; S5. Perform identity matching on the tracking object corresponding to the short-time tracking ID according to the k matching identities, corresponding confidence levels, and the mapping table. After the identity matching is completed and the game is not over, return to step S1.

2. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 1, wherein: Step S5 further comprises: S51: Select the matching identity corresponding to the maximum confidence and check whether it matches the existing short-term tracking ID in the mapping table. If so, proceed to step S52; otherwise, mark the tracked object as a matching identity and update the mapping table. S52: Determine whether the maximum confidence is greater than the confidence of the athlete corresponding to the short-term tracking ID. If so, proceed to step S53; otherwise, delete the matching identity corresponding to the maximum confidence and return to step S51. S53: Mark the tracking object as a matching identity, delete the mapping relationship between the existing short-term tracking ID in the mapping table, and return to step S1 when the game is not settled.

3. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 1, wherein: It also includes determining whether the tracking target corresponding to the camera device has ID drift, and correcting it if ID drift occurs: A1. Obtain time-aligned multi-view athlete tracking data and convert the image coordinates from the multi-view into athlete coordinate data in a standardized top-view coordinate system using a homography transformation method; A2. Collect the standardized coordinate data of each athlete from all perspectives, remove outliers, and weight the coordinate data of the athlete observed from multiple perspectives to obtain the athlete's current fused coordinates. A3. Determine whether the difference between the coordinate data under a single view and the fused coordinates exceeds a preset tolerance threshold. If so, proceed to step A4. Otherwise, the tracking target corresponding to the camera device does not experience ID drift. A4. The tracking target corresponding to the current camera device's viewing angle experiences ID drift, and the mapping relationship between the short-term tracking ID corresponding to the current camera device and the athlete in the mapping table is deleted.

4. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 3, wherein: The expression converted into the normalized athlete coordinate data in step A1 is: ， in, and These are the coordinate data of the athletes before and after normalization; is the homography transformation matrix; 、 and for The three elements in the first row of 、 and for The three elements in the second row; 、 and for The three elements in the third row; The expression of the athlete's fusion coordinates at the current moment is: in, and is the fusion coordinate of athlete i at the current moment; For athlete i at the current moment Standardized coordinate data of visual observation; For athlete i at the current moment The weight corresponding to the viewing angle; For athlete i at the current moment The coordinates of the camera device of the viewing angle; is the correction parameter; n is the total number of cameras.

5. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 1, characterized in that: After step S2 and before step S4, the player feature database is also updated: B1. Based on the current real-time athlete tracking data, dynamically capture athlete image frames from the kabaddi game footage, extract time-sensitive appearance feature vectors based on the image frames, and store them in the athlete feature database; B2. Based on the appearance feature vector and the time-sensitive appearance feature vector in the basic appearance feature data, a weighted fusion algorithm is used to perform feature generalization processing and construct a fused feature representation vector: in, is the fusion feature representation vector of the athlete; m is the total number of appearance feature vectors in the basic appearance feature data; is the oth appearance feature vector of the athlete; for The corresponding weight; N is the total number of the time-sensitive appearance feature vectors of the athlete at the current moment; is the p-th temporal appearance feature vector of the athlete; for The corresponding weight; B3. Perform L2 normalization on the fusion feature representation vector of the athlete: in, is the quadratic norm; B4. The L2-normalized fusion feature representation vector of the athlete is stored in the athlete feature database.

6. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 5, characterized in that: The method for updating the athlete feature database further includes: Calculate the weight of each time-sensitive appearance feature vector in the athlete feature database at the current moment: in, is the weight of each time-sensitive appearance feature vector at the current moment, is the hyperparameter of the decay rate; is the hyperparameter of the global temporal feature weight; For the current moment; The moment when each time-sensitive appearance feature vector is stored in the athlete feature database; When the weight of each time-sensitive appearance feature vector at the current moment is less than a preset removal threshold, it will be removed from the athlete feature database.

7. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 6, characterized in that: The method for screening k matching identities and corresponding confidences in the athlete feature database based on the feature vector includes: Calculate the cosine similarity between the feature vector and each fusion feature representation vector in the athlete feature database; The k fusion feature representation vectors with the largest cosine similarity are selected, and the corresponding athletes are used as the matching identities of the feature vectors, and the cosine similarity is used as the confidence of the matching identity.

8. The multi-viewpoint athlete tracking and identification method for kabaddi competition according to claim 6, characterized in that: The method for obtaining basic appearance feature data is: In the initialization phase, image samples of each athlete at different positions in the kabaddi competition venue are manually pre-collected. The image samples include multi-scale perspective images of close distance, medium distance and long distance; The corresponding appearance feature vector is extracted for each image sample, and multiple appearance feature vectors of all athletes are used to form basic appearance feature data.

9. The multi-view athlete tracking and identification method for kabaddi competition according to any one of claims 1 to 8, characterized in that: In the mapping table, the confidence corresponding to each athlete mapped with a short-term tracking ID decays linearly with the recording time.

10. The multi-view athlete tracking and identification method for kabaddi competition according to any one of claims 1 to 8, characterized in that: Methods for filtering tracking objects within the effective competition area based on the human target detection results include: Constructing a valid playing area mask for kabaddi matches : in, is the coordinate of the tracked object; Each detected tracking object Calculate the intersection rate with the valid competition area mask: in, is the intersection rate; When the intersection rate is less than the empirical threshold, the corresponding tracking object Delete and filter the tracking objects within the valid competition area.

Images