A multi-target tracking method, medium and product for soccer game scenarios

Abstract

This invention provides a method, medium, and product for multi-target tracking in football scenarios. The method includes acquiring video frames as input to an ECA_DLAseg backbone network to obtain multi-target tracking results. The ECA_DLAseg backbone network includes a joint detection head that outputs feature map detection results containing heatmaps, center offsets, bounding box sizes, and intersection-over-union (IoU); a ReID module for generating a unique identifier for each detected target based on the feature map detection results containing heatmaps, center offsets, bounding box sizes, and IoU; a Kalman filter for predicting the trajectory of each detected target and obtaining the predicted trajectory; and a data association module for matching the trajectory predicted by the Kalman filter with the unique identifier of each detected target through data association to obtain the final multi-target tracking result. This invention achieves accurate and efficient tracking of multiple targets in football scenarios.

Description

Technical Field

[0001] This invention relates to a multi-target tracking method, medium, and product for football scenarios, belonging to the fields of computer vision and deep learning technology. Background Technology

[0002] With the rapid development of computer vision and deep learning technologies, multi-object tracking algorithms have been widely applied in various scenarios, such as intelligent monitoring, autonomous driving, and motion analysis. However, multi-object tracking still faces many challenges in complex scenarios, especially in dynamic environments such as football matches, involving the rapid movement of multiple targets, complex occlusion, and frequent interactions. Traditional multi-object tracking algorithms mainly rely on a detection-association framework, which first detects targets in each frame and then concatenates these detection results to form a trajectory through data association. In recent years, end-to-end deep learning methods, such as the real-time multi-object tracking algorithm FairMOT, have gradually become mainstream. FairMOT combines target detection and re-identification tasks into a unified framework, achieving higher tracking accuracy and efficiency.

[0003] However, existing multi-target tracking algorithms still face several challenges when applied to football scenarios. First, the rapid changes in a football match and the complex, ever-changing trajectories of targets make it difficult for traditional methods to handle these dynamic changes. Second, occlusion between players is frequent, especially in densely populated playing areas, and existing methods often perform poorly in handling occlusion, leading to tracking failures. Finally, the frequent interactions between players in football matches increase the difficulty of tracking, making it challenging for existing methods to accurately distinguish and track each player.

[0004] To address the aforementioned issues, this invention proposes an improved multi-target tracking method specifically designed for multi-target tracking in football scenarios. This method improves upon the real-time multi-target tracking algorithm FairMOT in several ways, including introducing an efficient channel attention mechanism (ECA) to enhance the deep layer aggregation segmentation network DLAseg, developing a new joint detection head, and introducing a polynomial cross-entropy (PolyCE) loss function in the re-identification module. These improvements aim to enhance the detection accuracy of small targets, optimize detection performance in dense scenes, and improve tracking accuracy and robustness. Training and evaluation on the SoccerNet Tracking dataset validates the superior performance of the proposed method in rapidly changing, complex occlusion, and multi-player interaction scenarios, demonstrating its potential and effectiveness in real-world football match applications.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to combine target detection and re-identification tasks to solve the problem of multi-target occlusion, and to track and accurately distinguish the trajectory of each player in real time.

[0007] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0008] In a first aspect, the present invention provides a multi-target tracking method for a football game scenario, comprising: Acquire video frame images of a football action scene; Video frame images of a football game scene are input into the ECA_DLAseg backbone network to obtain multi-target tracking results; The ECA_DLAseg backbone network includes a joint detection head, a ReID re-identification module, a Kalman filter, and a data association module. The joint detection head is used to output a feature map containing the target to a feature map detection result that includes a heatmap, center offset, bounding box size and intersection-union ratio. The ReID module uses the PolyCE cross-entropy loss function as the re-identification loss to balance the ratio of positive and negative samples; it is used to generate a unique identifier for each detected target based on the feature map detection results, which include heatmap, center offset, bounding box size, and intersection-union ratio. The Kalman filter is used to predict the trajectory of each detected target to obtain the predicted trajectory; The data association module is used to match the trajectory predicted by the Kalman filter with the unique identifier of each detected target through data association to obtain the final multi-target tracking result; The training method for the ECA_DLAseg backbone network includes: The ECA_DLAseg backbone network is trained based on the training set and the total loss function to obtain the trained ECA_DLAseg backbone network.

[0009] Furthermore, the method for constructing the ECA_DLAseg backbone network includes: The upsampling ratio of the input feature map by the deep layer aggregation segmentation network DLAseg is increased from 1 / 4 to 1 / 2 of the original feature map size; An efficient channel attention mechanism (ECA) is embedded in each block of the deep layer aggregation segmentation network DLAseg to obtain the ECA_DLAseg backbone network.

[0010] Furthermore, the method for constructing the joint detection head includes: The generalized cross-union ratio (CIU) loss function is introduced into the CIU calculation module to obtain an improved CIU calculation module; An improved cross-union ratio calculation module is added to the detection head to obtain a joint detection head.

[0011] Furthermore, a method for generating a unique identifier for each detected target based on the feature map detection results, which include heatmaps, center offsets, bounding box dimensions, and intersection-union ratios, includes: Based on the feature map detection results including heatmap, center offset, bounding box size and cross-union ratio, the bounding box of each detected target is decoded, non-maximum suppression (NMS) is used to filter the bounding boxes, and the appearance feature map of each detected target is extracted from the filtered bounding boxes. The appearance feature map of each detected target is processed by two layers of convolution and ReLU activation function to reduce the number of channels of the appearance feature map of each detected target to 128, thus obtaining the processed appearance feature map; The classifier converts the embedded features in the processed appearance feature map into a score for each detected target, which serves as a unique identifier for each detected target.

[0012] Furthermore, the total loss obtained from the total loss function of the ECA_DLAseg backbone network includes the detection loss and re-identification loss of the joint detection head. The total loss of the ECA_DLAseg backbone network is expressed as follows: ; Where L represents the total loss of the ECA_DLAseg backbone network, This represents the weighting adjustment term for the detection loss of the combined detection head. This represents the weight adjustment term for the re-identification loss. This indicates the re-identification loss detection loss. This indicates the re-identification loss. This represents the learnable weights of the detection. This represents the learnable weights of the re-identification loss.

[0013] Furthermore, the detection loss of the joint detection head includes heatmap loss, size loss, offset loss, and cross-union ratio (IoU) loss; the detection loss is obtained by weighting and summing the heatmap loss, size loss, offset loss, and IoU loss according to their respective weights.

[0014] Furthermore, a Kalman filter is used to predict the trajectory of each detected target to obtain a predicted trajectory, which includes the trajectory of confirmed detected targets and the trajectory of unconfirmed detected targets.

[0015] Further, the step of matching the trajectory predicted by the Kalman filter with the unique identifier of each detected target through data association to obtain the final multi-target tracking result includes: The confirmed target trajectory and the unique identifier of each target are concatenated and matched to obtain the first non-matching detection result, the first non-matching trajectory, and the first matching trajectory; The first mismatch detection result and the first mismatch trajectory are matched with the unconfirmed detection target trajectory using Hungarian matching to obtain the second mismatch detection result, the second mismatch trajectory, and the second matching trajectory. The second matched trajectory is input into the Kalman filter to update the parameters of the first matched trajectory, thus obtaining the updated first matched trajectory; The second mismatch detection result is input into the Kalman filter and initialized as a new trajectory; If the second mismatched trajectory and the new trajectory can match the confirmed detection target trajectory in three consecutive frames, they are designated as the determined trajectories and fed into the Kalman filter to update the parameters of the first matching trajectory, thus obtaining the updated first matching trajectory. If the second mismatched trajectory and the new trajectory fail to match the confirmed detection target trajectory for several consecutive frames, and the number of consecutive frames in which the second mismatched trajectory and the new trajectory fail to match is less than a set threshold, the second mismatched trajectory and the new trajectory are retained, and the parameters of the first matching trajectory are updated by inputting them into the Kalman filter to obtain the updated first matching trajectory. If the second mismatched trajectory and the new trajectory fail to match the feature map detection result for several consecutive frames, and the number of consecutive mismatched frames of the new trajectory is greater than or equal to the set threshold, the new trajectory is deleted. Output all updated first-match trajectories as the multi-target tracking result.

[0016] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the above-described method.

[0017] Thirdly, the present invention also provides a computer program product, including a computer program / instructions, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the above-described method.

[0018] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention provides a multi-target tracking method for football scenarios, achieving accurate, efficient, and robust tracking of multiple targets in football scenes. This technology not only effectively extracts and identifies multiple targets in video frames but also predicts their trajectories. Furthermore, a re-identification module assigns a unique identifier to each target, ensuring stable tracking performance even in complex environments. This provides strong technical support for real-time analysis and decision-making in football scenarios.

[0019] The ECA_DLAseg backbone network provided by this invention significantly improves the detection accuracy of the deep layer aggregation segmentation network DLAseg by introducing the efficient channel attention mechanism ECA, thereby enhancing the detection accuracy of the ECA_DLAseg backbone network for long-distance targets in football scenes and allowing for better preservation and utilization of detailed information. Secondly, the joint detection head of the ECA_DLAseg backbone network integrates an Intersection over Union (IoU) calculation module and introduces a generalized IoU loss function into the IoU calculation module, effectively improving detection performance under complex occlusion and multi-player interaction conditions, ensuring accurate detection and localization in dense scenes. Furthermore, the ReID module of the ECA_DLAseg backbone network introduces a polynomial cross-entropy (PolyCE) loss function to balance the ratio of positive and negative samples, reducing overfitting of the ECA_DLAseg backbone network to frequently occurring targets and enhancing its ability to distinguish similar-looking targets, thereby improving the accuracy and robustness of tracking. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall network architecture of a multi-target tracking method for a football game scenario provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the network architecture of the deep layer aggregation and segmentation network DLAseg provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the network architecture of the ECA_DLAseg backbone network provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the workflow of the efficient channel attention mechanism ECA provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the ECA_DLAseg backbone network structure provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the combined detection head structure provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the ReID (Re-identification Module) workflow provided in an embodiment of the present invention; Figure 8This is a schematic diagram illustrating the data association process of a multi-target tracking method for a football game scenario provided by an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0022] Example 1 This embodiment introduces a multi-target tracking method for football scenarios, including: Acquire video frame images of a football action scene; Video frame images of a football game scene are input into the ECA_DLAseg backbone network to obtain multi-target tracking results; The overall network architecture of a multi-target tracking method for football scenarios is as follows: Figure 1 As shown.

[0023] The ECA_DLAseg backbone network includes a joint detection head, a ReID re-identification module, a Kalman filter, and a data association module. The joint detection head is used to output a feature map containing the target to a feature map detection result that includes a heatmap, center offset, bounding box size and intersection-union ratio. The ReID module uses the PolyCE cross-entropy loss function as the re-identification loss to balance the ratio of positive and negative samples; it is used to generate a unique identifier for each detected target based on the feature map detection results, which include heatmap, center offset, bounding box size, and intersection-union ratio. The Kalman filter is used to predict the trajectory of each detected target to obtain the predicted trajectory; The data association module is used to match the trajectory predicted by the Kalman filter with the unique identifier of each detected target through data association to obtain the final multi-target tracking result.

[0024] This invention enhances the deep layer aggregation segmentation network DLAseg by introducing the efficient channel attention mechanism ECA, resulting in the ECA_DLAseg backbone network. This significantly improves the detection accuracy of the ECA_DLAseg backbone network for distant targets in football scenes, allowing for better preservation and utilization of detailed information. The network architecture of the enhanced deep layer aggregation segmentation network DLAseg is as follows: Figure 2 As shown, the network structure of the ECA_DLAseg backbone network is as follows: Figure 3As shown in the diagram, the workflow of the Efficient Channel Attention (ECA) mechanism is illustrated below. Figure 4 As shown.

[0025] The method for constructing the ECA_DLAseg backbone network includes: The upsampling ratio of the input feature map by the deep layer aggregation segmentation network DLAseg is increased from 1 / 4 to 1 / 2 of the original feature map size; An efficient channel attention mechanism (ECA) is embedded in each block of the deep layer aggregation segmentation network DLAseg to obtain the ECA_DLAseg backbone network.

[0026] Efficient Channel Attention (ECA) is a lightweight channel attention method. It enhances the adaptability of convolutional neural networks to features through weighted aggregation, thereby improving the network's performance on specific tasks.

[0027] For example, the input feature map of the ECA_DLAseg backbone network is ,in, This represents the number of channels in the feature map. and Let R represent the height and width of the feature map, respectively, and let R represent the real number field.

[0028] First, perform global average pooling (GAP) on each channel of the input feature map to obtain a channel descriptor. Where C represents the dimension of the channel descriptor, i.e., the number of channels, and its c-th element is calculated as follows: (1); In the formula, z c This represents the average value of the c-th channel, where c represents a channel of the feature map. and These represent the height and width of the input feature map, respectively. i This represents the height dimension index of the feature map. j X represents the width dimension index of the feature map. cij This represents the pixel value of feature map X at channel c, height i, and width j.

[0029] The efficient channel attention mechanism, ECA, captures global information for each channel of the input feature map and then processes the channel descriptors using one-dimensional convolution. Based on the number of channels in the input feature map, it uses an adaptive mechanism to calculate the convolution kernel size to obtain the input feature map data. Figure 1 The output after 3D convolution enables the efficient channel attention mechanism ECA to dynamically adjust its internal parameters according to different input features.

[0030] The input features Figure 1The output after 3D convolution is represented as: (2); In the formula, Representing input features Figure 1 The output after 3D convolution, represents a one-dimensional convolution operation, and z represents a channel descriptor.

[0031] Based on input features Figure 1 The output of the 3D convolution calculates the channel weights of each channel of the input feature map, and then recalibrates the input feature map using these channel weights to obtain the output for each channel of the recalibrated input feature map. The sigmoid function is used to calculate the channel weights, ensuring that the values ​​are between 0 and 1.

[0032] The channel weight of each channel of the input feature map is expressed as follows: (3); In the formula, This represents the channel weight of each channel in the input feature map. Representing input features Figure 1 The output after 3D convolution, This represents the Sigmoid function, which normalizes the output value to a range between 0 and 1.

[0033] The output of each channel of the recalibrated input feature map is represented as follows: (4); In the formula, y c w represents the output of the c-th channel of the recalibrated input feature map. c x represents the channel weight of the c-th channel of the input feature map. c This represents the c-th channel of the input feature map.

[0034] The constructed ECA_DLAseg backbone network, through adjustments to the upsampling strategy, more effectively preserves the detailed information of the input. Furthermore, the introduced efficient channel attention mechanism, ECA, significantly enhances the network's sensitivity to global and local image details, particularly in color and morphology. This mechanism allows the ECA_DLAseg backbone network to focus on processing the feature information of small targets on higher-resolution feature maps, thereby significantly improving the detection accuracy for athletes with small target areas. These improvements ensure the efficiency and accuracy of the ECA_DLAseg backbone network in capturing crucial motion details.

[0035] The schematic diagram of the ECA_DLAseg backbone network structure is shown below. Figure 5 As shown.

[0036] The training method for the ECA_DLAseg backbone network includes: The ECA_DLAseg backbone network is trained based on the training set and the total loss function to obtain the trained ECA_DLAseg backbone network.

[0037] The training process of the ECA_DLAseg backbone network includes the following steps:

[0038] Step 1: Use the ECA_DLAseg backbone network to extract all feature maps containing the target to be detected from the training set. The network architecture of the ECA_DLAseg backbone network is as follows: Figure 3 As shown.

[0039] The SoccerNet Tracking dataset is used as the training set for this invention. The SoccerNet Tracking dataset is a multi-object tracking dataset specifically designed for football match scenarios. It provides rich football movement scenarios and challenges, such as fast movement, occlusion, and intense interactions between players, making it ideal for training multi-object tracking algorithms for football scenes. The SoccerNet Tracking dataset contains key moments from 12 important football matches captured by the main camera, totaling 200 video clips. The video data is stored as video frame images, with each video clip lasting 30 seconds and containing 750 frames. Each video clip provides detailed tracking data, including frame ID, tracking ID, bounding box coordinates X, Y, bounding box width W, bounding box height H, and bounding box confidence score.

[0040] The overall annotation format of the SoccerNet Tracking dataset is consistent with that of the MOT20 dataset.

[0041] Step 2: Input the feature map into the joint detection head to obtain the feature map detection result, which includes heatmap, center offset, bounding box size, and intersection-over-union ratio. A schematic diagram of the joint detection head structure is shown below. Figure 6 As shown, n represents the number of categories.

[0042] This invention introduces a generalized cross-union ratio (GIoU) ​​loss function into the cross-union ratio calculation module to obtain an improved cross-union ratio calculation module. By adding the improved cross-union ratio calculation module to the existing detection head, a joint detection head is proposed.

[0043] The joint detection head is responsible for performing target detection tasks to identify targets in a football scene.

[0044] The feature map containing the target to be detected is input into the joint detection head. Through a series of convolution operations and batch normalization operations, the feature map detection results containing heatmap, center offset, bounding box size and intersection-over-union (IoU) are obtained.

[0045] Among them, the heatmap is used to represent the probability that each position in the feature map containing the target is the center of the target; the center offset is used to correct the prediction error of the target center point due to discretization; the bounding box size is used to predict the width and height of the target; and the intersection-union calculation module is used to evaluate the degree of overlap between the predicted bounding box of the target and the true bounding box of the target.

[0046] In the intersection-over-union (CIU) calculation module, this invention introduces a generalized CIU loss function to measure the geometric relationship between bounding boxes. The generalized CIU loss function is expressed as: (5); In the formula, GIoU represents the generalized intersection-union ratio. The traditional intersection-union ratio (CUIR) is calculated as the intersection area between the predicted bounding box and the ground truth bounding box of the detected target divided by their union area. E represents the area of ​​the smallest closed region containing both the predicted and ground truth bounding boxes of the detected target. This represents the area of ​​the union region.

[0047] The introduction of the Generalized Intersection over Union (GIoU) ​​loss function helps to more accurately optimize the position and size of bounding boxes during bounding box regression, especially when there is occlusion or overlap between detected targets, where the GIoU loss function can provide better performance.

[0048] The detection loss function of the joint detection head includes heatmap loss, size loss, offset loss, and cross-union ratio (CUP) loss. The heatmap loss is calculated using a Focal Loss-based loss function, while the size loss, offset loss, and IoU loss are calculated using an L1 loss function. The detection loss is calculated by weighted summation of the heatmap loss, size loss, offset loss, and CUP loss. The weights for heatmap loss and offset loss are set to 1, and the weights for size loss and IoU loss are set to 0.1.

[0049] Step 3: Based on the feature map detection results including heatmap, center offset, bounding box size and intersection-union ratio, decode the bounding box of each detected target, use non-maximum suppression (NMS) to filter the bounding boxes, and extract the appearance feature map of each detected target from the filtered bounding boxes.

[0050] Step 4: Input the appearance feature map of each detected target into the ReID re-identification module. Use the Polynomial Cross-Entropy (PolyCE) loss function as the re-identification loss of the ReID module to balance the ratio of positive and negative samples and generate a unique identifier for each detected target. The workflow of the ReID module is illustrated in the diagram below. Figure 7 As shown.

[0051] The ReID (Re-identification) module is responsible for generating re-identification features. In the ReID module, each detected target is assigned a unique identifier.

[0052] The specific steps include: processing the appearance feature map of each detected target through two layers of convolution and ReLU activation function to reduce the number of channels of the appearance feature map of each detected target to 128, thereby obtaining the processed appearance feature map.

[0053] The classifier transforms the embedded features in the processed appearance feature map into a score for each detected target, serving as a unique identifier for each target. The score for each detected target has the dimension (B, nID), where B represents the batch size (the number of samples processed at one time) and nID represents the total number of different individuals that the ECA_DLAseg backbone network can identify or track.

[0054] This invention introduces the Polynomial Cross-Entropy (PolyCE) loss function as the loss function during the training process of the ECA_DLAseg backbone network. The overall loss of the ECA_DLAseg backbone network includes the detection loss of the joint detection head and the re-identification loss of the ReID module.

[0055] By imposing additional penalties on samples that are difficult to distinguish, the ECA_DLAseg backbone network is able to learn to distinguish similar-looking targets more effectively, so that the identity of each target can be accurately identified in frequently changing environments, thereby enhancing the ECA_DLAseg backbone network's ability to cope with occlusion, fast movements, and complex scene changes.

[0056] The polynomial cross-entropy PolyCE loss function is expressed as follows: (6); Here, PolyCE represents the polynomial cross-entropy loss function. A vector representing the actual label. This represents the label vector predicted by the model. The standard cross-entropy loss function is expressed by the following formula: This represents a small positive constant used to adjust the balance of the loss function. Represents the first element in the predicted label vector. Each element specifically points to a mispredicted item. This indicates the power of a polynomial term.

[0057] In summary, the total loss of the ECA_DLAseg backbone network includes the detection loss of the joint detection head and the re-identification loss of the ReID re-identification module. The total loss of the ECA_DLAseg backbone network is expressed as follows: (7); Where L represents the total loss of the ECA_DLAseg backbone network, This represents the weighting adjustment term for the detection loss of the combined detection head. This represents the weight adjustment term for the re-identification loss. This indicates the re-identification loss detection loss. This indicates the re-identification loss. This represents the learnable weights of the detection. This represents the learnable weights of the re-identification loss.

[0058] Step 5: Use a Kalman filter to predict the trajectory of each detected target to obtain the predicted trajectory.

[0059] The trajectory of each detected target is predicted using a Kalman filter to obtain the predicted trajectory, which includes the trajectory of confirmed detected targets and the trajectory of unconfirmed detected targets.

[0060] Step 6: Match the predicted trajectory with the unique identifier of each detected target through data association to obtain the final multi-target tracking result. A flowchart illustrating the data association process of a multi-target tracking method for a football scenario provided in this embodiment of the invention is shown below. Figure 8 As shown.

[0061] The confirmed target trajectory and the unique identifier of each target are concatenated and matched to obtain the first non-matching detection result, the first non-matching trajectory, and the first matching trajectory; The first mismatch detection result and the first mismatch trajectory are matched with the unconfirmed detection target trajectory using Hungarian matching to obtain the second mismatch detection result, the second mismatch trajectory, and the second matching trajectory. The second matched trajectory is input into the Kalman filter to update the parameters of the first matched trajectory, thus obtaining the updated first matched trajectory; The second mismatch detection result is input into the Kalman filter and initialized as a new trajectory; If the second mismatched trajectory and the new trajectory can match the confirmed detection target trajectory in three consecutive frames, they are designated as the determined trajectories and fed into the Kalman filter to update the parameters of the first matching trajectory, thus obtaining the updated first matching trajectory. If the second mismatched trajectory and the new trajectory fail to match the confirmed detection target trajectory for several consecutive frames, and the number of consecutive frames in which the second mismatched trajectory and the new trajectory fail to match is less than a set threshold, the second mismatched trajectory and the new trajectory are retained, and the parameters of the first matching trajectory are updated by inputting them into the Kalman filter to obtain the updated first matching trajectory. If the second mismatched trajectory and the new trajectory fail to match the feature map detection result for several consecutive frames, and the number of consecutive mismatched frames of the new trajectory is greater than or equal to the set threshold, the new trajectory is deleted. Output all updated first-match trajectories as the multi-target tracking result.

[0062] Example 2

[0063] A computer-readable storage medium storing computer instructions thereon, characterized in that the computer instructions, when executed by a processor, implement the steps of the method described in Embodiment 1 above.

[0064] Example 3

[0065] The present invention also provides a computer program product, including computer instructions, characterized in that the computer instructions, when executed by a processor, implement the steps of the method described in Embodiment 1 above.

[0066] In summary, this invention, through the use of an ECA_DLAseg backbone network, a joint detection head, a ReID (Re-identification) module, and a Kalman filter, successfully achieves accurate, efficient, and robust tracking of multiple targets in a football scenario. This technology not only effectively extracts and identifies multiple targets in video frames but also predicts their trajectories. Furthermore, the ReID module assigns a unique identifier to each target, ensuring stable tracking even in complex environments. This provides strong technical support for real-time analysis and decision-making in football scenarios. It solves the problems of existing technologies in handling multi-target occlusion and accurately distinguishing the trajectories of each player in real-time.

[0067] This invention provides a multi-target tracking method for football scenes. It enhances the deep layer aggregation segmentation network DLAseg by introducing an efficient channel attention mechanism (ECA), significantly improving the detection accuracy of the ECA_DLAseg backbone network for distant targets in football scenes, allowing for better preservation and utilization of detailed information. Secondly, the joint detection head integrates an Intersection over Union (IoU) calculation module and introduces a generalized IoU loss function within this module, effectively improving detection performance under complex occlusion and multi-player interaction conditions, ensuring accurate detection and localization in dense scenes. Furthermore, this invention introduces a polynomial cross-entropy (PolyCE) loss function into the ReID module to balance the ratio of positive and negative samples, reducing overfitting of the ECA_DLAseg backbone network to frequently occurring targets and enhancing its ability to distinguish between similar-looking targets, thereby improving tracking accuracy and robustness.

[0068] This invention was trained and evaluated on the SoccerNet Tracking dataset, verifying the superior performance of the multi-target tracking method provided by this invention for football scenarios in rapidly changing, complex occlusion, and multi-player interaction scenarios, demonstrating its potential and effectiveness in actual football match applications.

[0069] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0070] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A multi-target tracking method for football scenarios, characterized in that, include: Acquire video frame images of a football action scene; Video frame images of a football game scene are input into the ECA_DLAseg backbone network to obtain multi-target tracking results; The ECA_DLAseg backbone network includes a joint detection head, a ReID re-identification module, a Kalman filter, and a data association module. The joint detection head is used to output a feature map containing the target to a feature map detection result that includes a heatmap, center offset, bounding box size and intersection-union ratio. The ReID module uses the PolyCE cross-entropy loss function as the re-identification loss to balance the ratio of positive and negative samples. Used to generate a unique identifier for each detected target based on the feature map detection results, which include heatmap, center offset, bounding box size, and intersection-union ratio; The Kalman filter is used to predict the trajectory of each detected target to obtain the predicted trajectory; The data association module is used to match the trajectory predicted by the Kalman filter with the unique identifier of each detected target through data association to obtain the final multi-target tracking result; The training method for the ECA_DLAseg backbone network includes: The ECA_DLAseg backbone network is trained based on the training set and the total loss function to obtain the trained ECA_DLAseg backbone network. The method for constructing the ECA_DLAseg backbone network includes: The upsampling ratio of the input feature map by the deep layer aggregation segmentation network DLAseg is increased from 1 / 4 to 1 / 2 of the original feature map size; Efficient channel attention mechanism ECA is embedded in each block of the deep layer aggregation segmentation network DLAseg to obtain the ECA_DLAseg backbone network; The method for constructing the joint detection head includes: The generalized cross-union ratio (CIU) loss function is introduced into the CIU calculation module to obtain an improved CIU calculation module; An improved cross-union ratio calculation module is added to the detection head to obtain a joint detection head.

2. The multi-target tracking method for a football game scenario according to claim 1, characterized in that, A method for generating a unique identifier for each detected target based on feature map detection results including heatmaps, center offsets, bounding box dimensions, and intersection-union ratios includes: Based on the feature map detection results including heatmap, center offset, bounding box size and cross-union ratio, the bounding box of each detected target is decoded, non-maximum suppression (NMS) is used to filter the bounding boxes, and the appearance feature map of each detected target is extracted from the filtered bounding boxes. The appearance feature map of each detected target is processed by two layers of convolution and ReLU activation function to reduce the number of channels of the appearance feature map of each detected target to 128, thus obtaining the processed appearance feature map; The classifier converts the embedded features in the processed appearance feature map into a score for each detected target, which serves as a unique identifier for each detected target.

3. The multi-target tracking method for a football game scenario according to claim 1, characterized in that, The total loss obtained from the total loss function of the ECA_DLAseg backbone network includes the detection loss of the joint detection head and the re-identification loss. The total loss of the ECA_DLAseg backbone network is expressed as follows: ； Where L represents the total loss of the ECA_DLAseg backbone network, This represents the weighting adjustment term for the detection loss of the combined detection head. This represents the weight adjustment term for the re-identification loss. This indicates the re-identification loss detection loss. This indicates the re-identification loss. This represents the learnable weights of the detection. This represents the learnable weights of the re-identification loss.

4. The multi-target tracking method for a football game scenario according to claim 3, characterized in that, The detection loss of the joint detection head includes heatmap loss, size loss, offset loss, and cross-union ratio (IoU) loss; the detection loss is obtained by weighting and summing the heatmap loss, size loss, offset loss, and IoU loss according to their respective weights.

5. The multi-target tracking method for a football game scenario according to claim 1, characterized in that, The trajectory of each detected target is predicted using a Kalman filter to obtain the predicted trajectory, which includes the trajectory of confirmed detected targets and the trajectory of unconfirmed detected targets.

6. The multi-target tracking method for a football game scenario according to claim 5, characterized in that, The steps of matching the trajectory predicted by the Kalman filter with the unique identifier of each detected target through data association to obtain the final multi-target tracking result include: The confirmed target trajectory and the unique identifier of each target are concatenated and matched to obtain the first non-matching detection result, the first non-matching trajectory, and the first matching trajectory; The first mismatch detection result and the first mismatch trajectory are matched with the unconfirmed detection target trajectory using Hungarian matching to obtain the second mismatch detection result, the second mismatch trajectory, and the second matching trajectory. The second matched trajectory is input into the Kalman filter to update the parameters of the first matched trajectory, thus obtaining the updated first matched trajectory; The second mismatch detection result is input into the Kalman filter and initialized as a new trajectory; If the second mismatched trajectory and the new trajectory can match the confirmed detection target trajectory in three consecutive frames, they are designated as the determined trajectories and fed into the Kalman filter to update the parameters of the first matching trajectory, thus obtaining the updated first matching trajectory. If the second mismatched trajectory and the new trajectory fail to match the confirmed detection target trajectory for several consecutive frames, and the number of consecutive frames in which the second mismatched trajectory and the new trajectory fail to match is less than a set threshold, the second mismatched trajectory and the new trajectory are retained, and the parameters of the first matching trajectory are updated by inputting them into the Kalman filter to obtain the updated first matching trajectory. If the second mismatched trajectory and the new trajectory fail to match the feature map detection result for several consecutive frames, and the number of consecutive mismatched frames of the new trajectory is greater than or equal to the set threshold, the new trajectory is deleted. Output all updated first-match trajectories as the multi-target tracking result.

7. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.

Images