Badminton action tracking method
The badminton motion tracking method combining YOLOv11 and DeepOCSort algorithms solves the problems of high cost and low accuracy, and realizes low-cost, high-precision badminton motion tracking and real-time tactical analysis, which is suitable for general venues and small and medium-sized training units.
Patent Information
- Application Number
- CN202511326232.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-17
- Publication Date
- 2025-11-21
AI Technical Summary
Existing badminton detection and tracking technologies are expensive and lack specialized training for their models, making them difficult to apply in general venues and competitions. Furthermore, their accuracy is insufficient to support coaches' precise tactical analysis.
By employing the YOLOv11 deep learning target detection model and the DeepOCSort tracking algorithm, combined with image preprocessing and an interactive analysis platform, high-precision tracking of badminton movements and ID consistency assurance are achieved.
It reduces system costs, improves badminton detection accuracy and tracking stability, supports real-time tactical analysis, provides quantified motion trajectories and landing point distribution, and provides coaches with real-time decision-making support.
Smart Images
Figure CN120997250A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of badminton tactical analysis, and particularly relates to a tracking method of badminton action. BACKGROUND
[0002] Badminton has been widely popular in the world and has become one of the most popular sports. With the improvement of competition level, the demand for tactical analysis and auxiliary judgment of competition results is increasing; Existing badminton detection and tracking technology relies on high-order photography equipment and special systems, which has high accuracy, but the construction and maintenance cost is high, and cannot be widely applied to schools, training venues and general events, in addition, some research has tried to apply deep learning model for badminton tracking, but most models are designed for general use, lack of specialized training for badminton scene, resulting in insufficient accuracy in actual application, and cannot ensure consistent tracking across frames, making it difficult to support coaches to conduct accurate tactical analysis, therefore, improvement is needed; Therefore, it is necessary to invent a tracking method of badminton action. SUMMARY
[0003] Therefore, the present application provides a tracking method of badminton action to solve the problems in the background art.
[0004] In order to achieve the above purpose, the present application provides the following technical scheme: a tracking method of badminton action, a tracking method of badminton action, the specific steps are as follows: S1, image data extraction and preprocessing: extract each frame of image from the original data, and adjust it to 1920x1080 pixels, and unify the spatial resolution; S2, labeling and semi-automatic labeling process: first manually label the badminton, racket and athlete in each image, label 500 images for each object category, select a small amount of manually labeled data to fine-tune the YOLOv11 model, and use the fine-tuned model to automatically label the remaining images and manually check; S3, introducing the latest object detection model: adopting YOLOv11 deep learning target detection model, using its architecture design and training strategy advantages to improve the identification accuracy and reasoning speed of high-speed moving badminton and related targets in the image; S4, model training and verification: adopting the YOLOv11 target detection model trained specially, based on 70% training set, 10% verification set and 20% test set to optimize the model weight; S5, object tracking and ID consistency guarantee: combined with DeepOCSort tracking algorithm, a dynamic exponential moving average (EMA) mechanism based on detection confidence score is specially introduced to handle the motion blur problem caused by high-speed movement of badminton. This method can automatically adjust the weight of appearance features according to the detection confidence of each frame, suppress the interference of unstable features, significantly improve the ID retention rate in cross-frame tracking, and solve the ID switching phenomenon commonly seen in traditional trackers in badminton movement.
[0005] S6, action trajectory confirmation and quantity statistics: combined with the target bounding box detected by YOLOv11 and the tracking result of DeepOCSort, the continuous action trajectory of badminton from starting to landing point is generated, the number of occurrences of each object is counted, and the space classification is carried out according to the landing position to generate quantitative data. S7, interactive analysis platform design: design an interactive platform based on data analysis to facilitate coaches and technical analysts to intuitively understand the action distribution of the opponent in the game.
[0006] Preferably, the interactive platform comprises: a function button module for users to filter and query according to specific analysis requirements, such as landing area distribution and high-frequency landing points; a real-time visualization module for dynamically displaying the continuous action trajectory of badminton, the number of landing points and the hot spot distribution; a data update module for real-time refreshing of analysis results, so that changes in the game process can be quickly reflected on the interface, providing dynamic decision-making basis.
[0007] Preferably, the original data in step S1 is derived from world-class badminton match videos, and each extracted frame of image ensures to contain key actions such as serving, hitting and swinging.
[0008] The beneficial effects of the present application are: Low cost threshold: no high-level hardware equipment is needed for implementation, suitable for general venues or small and medium-sized training units, with high feasibility and promotion potential.
[0009] Improved detection accuracy: through retraining of YOLOv11 model, the recognition ability of high-speed moving badminton can be significantly improved, effectively supporting key analysis requirements such as landing point judgment and ball detection.
[0010] Strong tracking stability: combined with DeepOCSort tracking algorithm, effectively solves the problem of inconsistent object ID and tracking interruption, ensuring the continuity and correctness of data analysis.
[0011] Real-time application can be realized: the system has high running efficiency, real-time data updating and interactive interface display can be realized through the design of the interactive platform, and the coach and the analyst can quickly adjust the tactics according to the on-site data.
[0012] Strengthen the tactical auxiliary function: finally, accurate motion trajectory and landing point hot area distribution can be output, providing quantitative basis for post-match analysis, and further improving the training quality and competition performance. BRIEF DESCRIPTION OF DRAWINGS
[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed in the following embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other implementation drawings can be derived from the provided drawings without creative labor.
[0014] The structure, proportion, size, etc. shown in the specification are only used to cooperate with the content disclosed in the specification, so that those skilled in the art can understand and read, and are not used to limit the implementation conditions of the present application, so they do not have technical significance. Any modification of structure, change of proportion relationship or adjustment of size, without affecting the effect and purpose that the present application can produce, should still fall within the scope of the technical content disclosed by the present application.
[0015] Figure 1 The step flowchart provided by the present application.
[0016] Figure 2 The Yolov11 model schematic diagram in the embodiment provided by the present application.
[0017] Figure 3 The Yolov11 model schematic diagram in the embodiment provided by the present application.
[0018] Figure 4 is a schematic diagram of Yolov11 tracking classification head in the embodiment of the present application.
[0019] Figure 5 The DeepOcsort tracking algorithm schematic diagram in the embodiment of the present application.
[0020] Figure 6 Actual image in the database. DETAILED DESCRIPTION
[0021] The preferred embodiments of the present application will be described below in conjunction with the drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.
[0022] Referring to the drawings Figure 1The present invention provides a method for tracking badminton movements, the specific steps of which are as follows: S1. Image data extraction and preprocessing: Extract each frame of image from the raw data and adjust it to 1920×1080 pixels, while unifying the spatial resolution. S2. Labeling and semi-automatic labeling process: First, manually label the badminton shuttlecock, racket and athlete in each image. Label 500 images for each object category. Select a small amount of manually labeled data to fine-tune the YOLOv11 model. Use the fine-tuned model to automatically label the remaining images and then manually review them. S3. Introducing the latest object detection model: The YOLOv11 deep learning object detection model is adopted, leveraging its architectural design and training strategy advantages to improve the accuracy and inference speed of identifying high-speed moving badminton shuttlecocks and related targets in images. The YOLOv11 deep learning target detection model is adopted. Its core architecture achieves feature extraction and target prediction through forward propagation, where the forward propagation formula is: ; in, For the first The output of the layer, and The first Layer weights and biases For activation functions; The model is trained using the Complete Intersection over Union (CIOU) loss, defined as follows: ; Where w is the width of the predicted bounding box, h is the height of the predicted bounding box, w^gt is the width of the ground truth bounding box, h^gt is the height of the ground truth bounding box, α is the weight variable, v is the aspect ratio, b is the center point of the predicted bounding box, and b^gt is the center point of the ground truth bounding box.
[0023] S4. Model Training and Validation: A specialized YOLOv11 object detection model was used, and the model weights were optimized based on 70% of the training set, 10% of the validation set, and 20% of the test set. During training, the cross-entropy loss function is used to optimize the classification task. The formula is the loss function (cross-entropy) updated by gradient descent: ; in, For the first The output of the layer, and The first Layer weights and biases This is the activation function.
[0024] S5, object tracking and ID consistency guarantee: combined with DeepOCSort tracking algorithm, especially introduced dynamic exponential moving average (EMA) mechanism based on detection confidence score to handle the motion blur problem caused by high-speed movement of badminton, this method can automatically adjust the weight of appearance features according to the detection confidence of each frame, suppress the interference of unstable features, significantly improve the ID retention rate in cross-frame tracking, and solve the ID switching phenomenon commonly seen in traditional trackers in badminton movement; Combined with DeepOCSort tracking algorithm, cross-frame ID consistency tracking is realized by dynamically adjusting the weight of exponential moving average (EMA). The EMA weight update formula is: DA dynamically adjusts the weight of Exponential Moving Average (EMA), and determines whether the appearance feature of the current frame should be integrated into the trajectory model according to the detection confidence score.
[0025] ; : the trajectory appearance embedding feature at the current time.
[0026] : the trajectory embedding feature of the previous frame.
[0027] : the appearance embedding feature of the newly detected target.
[0028] : the fixed weight parameter of EMA, which determines the fusion ratio of new and old features.
[0029] ; : detection confidence score (Detection Confidence Score).
[0030] : detection confidence score threshold, minimum confidence score requirement.
[0031] When =σ, =1, indicating completely ignoring the current appearance embedding .
[0032] =1, = , indicating that the current embedding is added to the model with a weight (1− ).
[0033] Fixed base value (paper set 0.95) Embedding is a multi-dimensional vector, for example , used to numerically and compress the appearance information of objects.
[0034] ReID matches targets through appearance embedding features, and combines with motion models to form a complete association cost matrix.
[0035] ; : Appearance cost of track m and detection box n (the smaller the value, the more similar).
[0036] : Appearance feature vector of track m.
[0037] : Appearance feature vector of detection box n.
[0038] : Cosine similarity, used to measure the similarity of two feature vectors.
[0039] ; The cost calculated by the motion model (such as IoU or Kalman filter), used to measure the matching degree of track and detection box in space.
[0040] Final association cost (matrix) C): ; C[m,n]: Comprehensive association cost of track m and detection box n.
[0041] : Global appearance cost weight, adjusts the relative importance of appearance and motion information.
[0042] : Distinctiveness weighting, dynamically adjusts the influence of appearance cost according to distinctiveness.
[0043] Calculate the appearance distinctiveness of each track and detection box (the difference between the first and second highest scores), and improve the matching priority of high distinctiveness.
[0044] S6, action track confirmation and quantity statistics: combine the target bounding box detected by YOLOv11 and the tracking result of DeepOCSort to generate the continuous action track of badminton from starting to landing point, count the appearance times of each object, and classify the space according to the landing position to generate quantitative data. S7, interactive analysis platform design: design an interactive platform based on data analysis to facilitate coaches and technical analysts to intuitively understand the action distribution of opponents in the game.
[0045] The interactive platform comprises: a function button module for users to filter and query according to specific analysis requirements, the specific analysis requirements being landing area distribution and high-frequency landing points; a real-time visualization module for dynamically displaying continuous action trajectories of badminton, landing point frequencies and hotspot distribution; a data updating module for real-time refreshing of analysis results, so that changes in the process of the game can be quickly reflected on the interface, and dynamic decision-making basis is provided.
[0046] The original data in step S1 is derived from world-class badminton game videos, and each extracted frame of image ensures to contain key actions of serving, hitting and swinging.
[0047] The use process of the application is as follows: first, image data extraction and preprocessing are performed, world-class badminton game videos are used as the original training data source, each frame of image is extracted from the selected high-quality game videos, and key actions in the game process are ensured to be captured completely in each video frame, after extraction, the size of each image is adjusted, and 1920x1080 pixels are uniformly set to ensure that the subsequent model input image has consistent spatial resolution and picture ratio, thereby helping to improve the accuracy and stability of object detection; Then, the labeling and semi-automatic marking process is performed, each image obtained in the above step is marked with an object bounding box according to different targets (such as a badminton, a racket, an athlete, etc.) appearing in the image, at least 500 images are marked for each object category to ensure the richness of the data set, then a small amount of manually labeled image data is selected to fine-tune the latest Yolov11 object detection model, the model weight capable of recognizing various targets is preliminarily obtained, then the fine-tuned Yolov11 model is used to quickly detect the remaining images, and a preliminary object labeling result is generated, after completion, the automatic labeling result is carefully reviewed and corrected by manual work to ensure that all object bounding boxes are accurate, and high-quality labeling data is provided for subsequent large-scale training; Subsequently, the latest object detection model is introduced, after the data labeling work is completed, the most advanced deep learning object detection model Yolov11 is used in this step, this model has significant advantages in architecture design and training strategy, can more accurately recognize various targets in the image, and has high inference speed and accuracy; Then model training and verification are carried out, after having sufficient labeled data set, overall model training is carried out using Yolov11, the specific process is as follows: all labeled data is divided according to the proportion of 70% training, 10% verification, 20% test, 70% data is used as the training set, and the model weight is updated gradually, in each training iteration (epoch), the performance of the model is evaluated in real time using 10% of the verification set, the best weight parameter is selected according to the verification result, finally the remaining 20% data is used as an independent test set to quantitatively test the performance of the final model, and ensure that the model can also maintain excellent object detection accuracy on unseen data; At the same time, object tracking and ID consistency are ensured, in order to ensure that the same object can be accurately identified and tracked in continuous frames based on object detection, it is necessary to introduce the DeepOCsort tracking algorithm, which can solve the ID switching phenomenon of traditional tracking of badminton, and use DeepOCsort to solve it, the algorithm can maintain ID consistency according to the feature information of the object in different frames, prevent repeated or missed tracking problems; the motion trajectory of the object can be automatically constructed, the landing point and action path are tracked; accurate space-time position information can also be provided for subsequent statistical action data and analysis of competition strategy; Then the action trajectory is confirmed and the number is counted, combining the object classification bounding box detected by Yolov11 and the tracking result of DeepOCsort, the continuous action trajectory of the badminton is confirmed by using the tracking algorithm, so that the tracking of the badminton is realized, and the landing point frequency and spatial distribution are counted; Finally, the interactive platform is designed to facilitate coaches and technical analysts to intuitively understand the action distribution of the opponent in the game, wherein the interactive platform comprises: a function button module, which is used for users to select and query according to specific analysis requirements, the specific analysis requirements are landing area distribution and high-frequency landing point; a real-time visualization module, which is used for dynamically displaying the continuous action trajectory of the badminton, the landing point frequency and the hotspot distribution; a data updating module, which is used for refreshing the analysis result in real time, so that the changes in the game process can be quickly reflected on the interface, and dynamic decision basis is provided; Among them, the core technology of the badminton action tracking method is the retrained YOLOv11 model and the advanced tracking algorithm DeepOCSort, the mathematical model is used to support the operation principle of each step in the system, for example, the forward propagation in the neural network, the convolution feature extraction, the data error calculation, and the state estimation in the dynamic target tracking, all have clear mathematical expressions, the specific formula is recorded in the above, which will not be described here.
Claims
1. A method for tracking badminton movements, characterized in that, The specific steps are as follows: S1. Image data extraction and preprocessing: Extract each frame of image from the raw data and adjust it to 1920×1080 pixels, while unifying the spatial resolution. S2. Labeling and semi-automatic labeling process: First, manually label the badminton shuttlecock, racket and athlete in each image. Label 500 images for each object category. Select a small amount of manually labeled data to fine-tune the YOLOv11 model. Use the fine-tuned model to automatically label the remaining images and then manually review them. S3. Introducing the latest object detection model: The YOLOv11 deep learning object detection model is adopted, leveraging its architectural design and training strategy advantages to improve the accuracy and inference speed of identifying high-speed moving badminton shuttlecocks and related targets in images. S4. Model Training and Validation: A specialized YOLOv11 object detection model was used, and the model weights were optimized based on 70% of the training set, 10% of the validation set, and 20% of the test set. S5. Object Tracking and ID Consistency Guarantee: Combining the DeepOCSort tracking algorithm, a dynamic exponential moving average (EMA) mechanism based on detection confidence scores is specially introduced to deal with the motion blur problem caused by the high speed of the shuttlecock. This method can automatically adjust the weight of appearance features according to the detection confidence of each frame, suppress the interference of unstable features, significantly improve the ID retention rate in cross-frame tracking, and solve the ID switching phenomenon commonly seen in traditional trackers in badminton. S6. Motion trajectory confirmation and quantity statistics: Combine the target bounding box detected by YOLOv11 with the tracking results of DeepOCSort to generate the continuous motion trajectory of the badminton shuttlecock from start to landing point, count the number of times each object appears, and perform spatial classification according to the landing point position to generate quantitative data. S7. Interactive Analysis Platform Design: Design an interactive platform based on data analysis to enable coaches and technical analysts to intuitively understand the distribution of opponents' actions during the game.
2. The method for tracking badminton movements according to claim 1, characterized in that: The interactive platform includes: The function button module allows users to filter and query based on specific analysis needs, such as the distribution of landing points and high-frequency landing points. The real-time visualization module is used to dynamically display the continuous trajectory of the badminton shuttlecock, the number of landing points, and the distribution of hot spots; The data update module is used to refresh the analysis results in real time, so that changes during the competition can be quickly reflected on the interface, providing a dynamic basis for decision-making.
3. The method for tracking badminton movements according to claim 1, characterized in that: In step S1, the raw data comes from world-class badminton match videos, and each extracted frame of video ensures that it contains key actions such as serving, hitting, and swinging the racket.