Sports evaluation method and system based on computer vision

Abstract

The invention discloses a sports evaluation method and system based on computer vision, and belongs to the field of sports intelligent evaluation, and the method comprises the following steps: S1, collecting a motion image; s2, matching and fusing the moving images under multiple view angles; s3, processing the motion video stream; s4, data preprocessing; s5, motion tracks of the athlete and the target are obtained, and motion parameters of key parts of the athlete are extracted; s6, comparing the motion parameters of the key part of the athlete with preset data, outputting an evaluation result according to a comparison result, and meanwhile, evaluating whether the speed change and the route of the athlete are compliant or not by analyzing a motion track; and S7, encrypting the evaluation result, and uploading the encrypted evaluation result to a user interaction interface for display. By the adoption of the sports evaluation method and system based on computer vision, evaluation accuracy can be improved, meanwhile, more visual and effective training suggestions are provided for athletes, and therefore the competitive level is promoted to be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of sports intelligent evaluation, and in particular to a sports evaluation method and system based on computer vision. Background Art

[0002] In the early days, sports teaching, training and competitions relied entirely on manual refereeing and evaluation, which was inefficient, subjective and had large errors. With the advancement of science and technology, infrared technology was introduced to assist monitoring and counting. Although it improved the degree of automation, it was still insufficient in terms of accuracy and environmental adaptability. It was greatly affected by environmental interference and could not meet the requirements of modern sports for accuracy and stability.

[0003] Nowadays, intelligent vision technology represented by visual AI is developing rapidly and is widely used in many fields such as campus sports and competitive sports. These technologies not only improve the objectivity and accuracy of evaluation, but also provide scientific training guidance for athletes. However, despite the many advances brought by intelligent vision technology, there are still some defects and deficiencies in technology and data, as follows: Defects and deficiencies in technology

[0004] 1. Limited environmental adaptability: Even advanced visual AI technology may suffer from problems such as reduced recognition accuracy and unstable tracking in extreme weather conditions (such as heavy rain, heavy snow, and dense fog) or complex lighting environments (such as strong backlight and frequent changes in stage lighting). This limits its application in outdoor and special venues.

[0005] 2. Occlusion problem to be solved: In multi-player sports scenarios, when athletes block each other or objects, the intelligent vision system may have difficulty accurately capturing and analyzing movements, resulting in missing or incorrect data. This problem is particularly prominent in team sports, affecting the accuracy of the overall evaluation.

[0006] 3. Difficulty in identifying high-speed moving objects: Intelligent vision systems may not be able to quickly and accurately track and identify high-speed moving objects in sports such as tennis and racing, affecting the judgment of the game situation and data collection. In particular, the system's response speed and accuracy become challenges in key frame extraction and standard comparison.

[0007] 4. Limitations of hardware equipment: Although high-definition camera and sensor technology continues to advance, the performance of different brands and models of equipment varies greatly, which may lead to uneven quality of collected data. In addition, factors such as the installation location and angle of the equipment will also affect the integrity and accuracy of the data.

[0008] 5. Data quality varies: Due to the accuracy of the collection equipment, environmental factors, and human operation, the collected data may be noisy, biased, or incomplete, affecting the accuracy of subsequent analysis and evaluation results. High-quality data is the basis for model training and application, so improving the quality and consistency of data collection is an urgent problem to be solved.

[0009] 6. Data labeling is difficult: Sports movements are complex and diverse, and accurate labeling of data requires professional knowledge and a lot of manpower. Labeling is costly, inefficient, and there may be inconsistent labeling, which affects the model training effect.

[0010] In summary, although intelligent vision technology has broad application prospects in the field of sports, it still needs to overcome technical challenges such as environmental adaptability, occlusion, and recognition of high-speed moving objects, and solve problems such as data security, quality, and labeling. Summary of the invention

[0011] The purpose of the present invention is to provide a sports evaluation method and system based on computer vision to solve the above technical problems.

[0012] To achieve the above object, the present invention provides a sports evaluation method based on computer vision, comprising the following steps:

[0013] S1. Collect a series of sports images of athletes and targets from multiple perspectives;

[0014] S2, matching and fusing the motion images from multiple perspectives acquired in step S1 to obtain multiple frames of motion images along the motion trajectory, wherein the multiple frames of motion images constitute a motion video stream;

[0015] S3, using the non-maximum suppression algorithm improved by Soft-NMS to process the motion video stream obtained in step S2 to remove occlusion;

[0016] S4, data preprocessing, and using adaptive illumination compensation algorithm to adjust the brightness and contrast of the motion video stream;

[0017] S5. Use the target detection algorithm in deep learning to identify the athlete and the target, then use algorithms such as Kalman filtering or particle filtering to obtain the motion trajectories of the athlete and the target, and use the OpenPose posture estimation method to extract the motion parameters of the key parts of the athlete;

[0018] S6. Compare the key motion parameters of the athlete with the preset data, and output the evaluation results according to the comparison results. At the same time, evaluate whether the athlete's speed change and route are in compliance by analyzing the motion trajectory;

[0019] S7. Encrypt the evaluation results and upload them to the user interface for display, and upload them to the cloud database for storage.

[0020] Preferably, in step S1, the motion images are captured using visible light cameras and infrared cameras arranged at multiple angles along the motion path.

[0021] Preferably, in step S4, when the illumination intensity is lower than a set threshold, the motion video stream is processed using an image enhancement technique, otherwise the background is removed using a background subtraction algorithm.

[0022] Preferably, in step S5, the target detection algorithm in deep learning is the YOLO algorithm or the Faster R-CNN algorithm;

[0023] The key parts of the athlete include the athlete’s torso attachment points, shoulder, hip, knee, ankle, and elbow joints;

[0024] Step S5 specifically includes the following steps:

[0025] S51, pre-processing each frame image in the input motion video stream and inputting it into the YOLO network, the YOLO network divides the input image into multiple grids, each grid is responsible for detecting the target in a specific area;

[0026] S52, inputting the image data with the athlete's body parts and key points labeled into the YOLO network for training, so that the YOLO network learns the feature representations of different parts and key parts; when a new image enters, the YOLO network predicts whether there is a target in each grid, and if so, outputs the category of the target and the bounding box coordinates;

[0027] S53, decoding the predicted bounding box, removing the bounding boxes with too high overlap by non-maximum suppression method, and obtaining the final accurate body part detection box;

[0028] S54, initialize the state vector and covariance matrix of the Kalman filter, where the state vector includes the position coordinates (x, y) and velocity (v x , v y );

[0029] S55. In each frame of the image, the state of the target in the current frame is predicted based on the state and motion model predicted in the previous frame:

[0030]

[0031] P k|k-1 =AP k-1|k-1 A T +Q;

[0032] In the formula, represents the state vector predicted at time k based on the state at time k-1; A represents the state transfer matrix; represents the optimal estimated state vector at time k-1; B represents the control input matrix; u k represents the control input at time k; P k|k-1 represents the covariance matrix predicted at time k based on the state at time k-1; P k-1|k-1 represents the covariance matrix of the optimal estimation state at time k-1; Q represents the process noise covariance matrix;

[0033] S56, taking the detection result obtained by the target detection algorithm as the observation value, and fusing the observation value with the prediction value, and then correcting the prediction result using the Kalman gain to obtain the target state estimation of the current frame:

[0034]

[0035] P k|k =(IK k H)P k|k-1 ;

[0036] In the formula, represents the optimal estimated state vector at time k; K k represents the Kalman gain at time k; Z k represents the observation value at time k; H represents the observation matrix; P k|k represents the covariance matrix of the optimal estimation state at time k; I represents the identity matrix;

[0037] S57, repeat steps S55-S56 to obtain the motion trajectory of the target;

[0038] S58, inputting the image containing the athlete after the target detection and trajectory tracking processing into the OpenPose model;

[0039] S59. Use the OpenPose model to extract features from the image, learn the feature representation of the human body in the image, predict the location of each key part of the athlete's body, and present the probability distribution of each key part in the image in the form of a heat map;

[0040] S510, determining the coordinates of the key parts according to the heat map, and extracting the motion parameters of the key parts of the athlete by calculating the relative position relationship and angle between the key parts.

[0041] Preferably, in step S6, the preset data is preset standard data or preset historical data, wherein the preset standard data is obtained by training with data annotated by experts using a semi-supervised learning method;

[0042] It specifically includes the following steps:

[0043] S61, comparing the motion parameters of the key parts of the athlete extracted from the extracted sports video with preset data to obtain a comparison result;

[0044] S62. Based on the comparison results of each parameter, a comprehensive evaluation of the athlete's overall performance is conducted;

[0045] S63. Generate evaluation results based on the comprehensive evaluation results;

[0046] S64. Calculate the velocity v according to the positions and time intervals of two adjacent points on the motion trajectory:

[0047]

[0048] Where (x1, y1) and (x2, y2) represent the position coordinates of two adjacent points; Δt represents the time interval between (x1, y1) and (x2, y2);

[0049] S65. Assume that the point on the standard route is (x s ,y s ), the point on the athlete’s trajectory is (x a ,y a ), then the deviation distance calculation formula is:

[0050]

[0051] In the formula, d represents the distance from the standard route;

[0052] S67, determine whether the route is compliant: set the maximum allowable deviation distance to d max , if d>d max , the route is considered non-compliant, otherwise it is considered compliant.

[0053] Preferably, in step S7, the athlete's training is guided in real time through voice broadcast.

[0054] Preferably, in step S7, the evaluation results are displayed in the form of charts, animations, or 3D models through virtual reality or augmented reality.

[0055] A sports evaluation system based on computer vision is used to execute a sports evaluation method based on computer vision, which includes a multi-angle camera module, a data processing module, an analysis and comparison module, a user interaction interface and a cloud database, wherein the multi-angle camera module is used to collect motion images of athletes and targets from multiple perspectives; the data processing module is used to process the collected motion images to obtain motion trajectories and motion parameters; the analysis and comparison module is used to compare the processed motion parameters with preset data to output evaluation results, and at the same time evaluate the speed changes and route compliance of athletes by analyzing the motion trajectory; the user interaction interface is used to display the evaluation results; and the cloud database is used to store the evaluation results.

[0056] Preferably, the data processing module includes a graphics processing unit and a programmable gate array, and the graphics processing unit and the programmable gate array are connected via a standard interface.

[0057] Therefore, the present invention adopts the above-mentioned sports evaluation method and system based on computer vision, which has the following beneficial effects:

[0058] 1. Multi-view data acquisition and fusion: (1) Multi-angle camera module: Multiple visible light and infrared cameras are used to capture the athlete's movements from different angles, ensuring the comprehensiveness and accuracy of the data. In particular, in complex environments (such as insufficient light or obstructions), infrared cameras can make up for the shortcomings of visible light cameras and provide a more reliable data source. At the same time, the non-perceptual monitoring method is more flexible in deployment. (2) Image matching and fusion technology: Accurately match and fuse images from different perspectives to generate a continuous and complete motion video stream, which helps to more accurately reconstruct the motion trajectory in three-dimensional space.

[0059] 2. Intelligent target detection and pose estimation: (1) Improved non-maximum suppression algorithm (Soft-NMS): effectively solves the problem of mutual occlusion between targets and improves the accuracy of target detection; (2) Adaptive illumination compensation and background subtraction algorithm: automatically adjusts image quality according to environmental conditions to ensure high-quality input data under various lighting conditions; (3) Advanced target detection and tracking algorithm: uses deep learning algorithms such as YOLO and Faster R-CNN combined with Kalman filtering or particle filtering to achieve high-precision recognition and stable tracking of athletes and their equipment; (4) OpenPose pose estimation method: accurately extracts the positions of key bone points of athletes, providing a solid foundation for subsequent motion analysis;

[0060] 3. Data-driven evaluation mechanism: (1) Comparison between preset standard data and historical data: Not only can evaluation be performed based on standard data annotated by industry experts, but also personalized analysis can be performed based on the athlete's own historical performance, allowing the athlete to obtain accurate feedback on the action in a timely manner, thereby improving the quality of training; for example, after a gymnast completes a move, the system can give instant feedback on the center of gravity, speed, and coordination, helping the athlete to make timely adjustments and improvements, reduce the risk of injury, improve the competitive level, and make the training more scientific and targeted; (2) Real-time feedback and guidance: Evaluation information is instantly conveyed to athletes through voice broadcasts, chart displays, etc., helping them to adjust their training strategies in a timely manner; it also supports the presentation of results in the form of virtual reality (VR) or augmented reality (AR), which enhances interactivity and fun and reduces interference from human factors;

[0061] 4. Safe and efficient data management: (1) Encrypted transmission and cloud storage: The evaluation results are encrypted and uploaded to the user interface for display, and are simultaneously saved to the cloud database, which not only ensures information security but also facilitates long-term archiving and query; (2) Modular design: The entire system consists of multiple independent but closely coordinated functional modules, which facilitates maintenance and upgrades, and also improves the flexibility and scalability of the system.

[0062] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The present invention is a flow chart of a sports evaluation method based on computer vision. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not used to limit the embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0065] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0066] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0067] like Figure 1 As shown, a sports evaluation method based on computer vision includes the following steps:

[0068] S1. Collect a series of sports images of athletes and targets from multiple perspectives;

[0069] In step S1, visible light cameras and infrared cameras arranged at multiple angles along the motion path are used to collect motion images. In bad weather, infrared cameras can be used to assist in imaging.

[0070] S2, matching and fusing the motion images from multiple perspectives acquired in step S1 to ensure the consistency and accuracy of the data, and obtaining multiple frames of motion images along the motion trajectory, which constitute a motion video stream;

[0071] S3, using the non-maximum suppression algorithm improved by Soft-NMS to process the motion video stream obtained in step S2 to remove occlusion;

[0072] S4, data preprocessing (cleaning, denoising, correction, rotation, flipping, cropping), and using adaptive illumination compensation algorithm to adjust the brightness and contrast of the motion video stream;

[0073] In step S4, when the light intensity is lower than a set threshold, the motion video stream is processed using image enhancement technology, otherwise the background is removed using a background subtraction algorithm.

[0074] S5. Use the target detection algorithm in deep learning to identify the athlete and the target, then use algorithms such as Kalman filtering or particle filtering to obtain the motion trajectories of the athlete and the target, and use the OpenPose posture estimation method to extract the motion parameters of the key parts of the athlete;

[0075] In step S5, the target detection algorithm in deep learning is the YOLO algorithm or the Faster R-CNN algorithm. The target detection algorithm in this embodiment is obtained by training data under multiple sports scenes to improve the accuracy and robustness of detection;

[0076] The key parts of the athlete include the athlete’s torso attachment points, shoulder, hip, knee, ankle, and elbow joints;

[0077] Step S5 specifically includes the following steps:

[0078] S51, pre-processing each frame image in the input motion video stream and inputting it into the YOLO network, the YOLO network divides the input image into multiple grids, each grid is responsible for detecting the target in a specific area;

[0079] S52, inputting the image data with the athlete's body parts and key points labeled into the YOLO network for training, so that the YOLO network learns the feature representations of different parts and key parts; when a new image enters, the YOLO network predicts whether there is a target in each grid, and if so, outputs the category of the target and the bounding box coordinates;

[0080] S53, decoding the predicted bounding box, removing the bounding boxes with too high overlap by non-maximum suppression method, and obtaining the final accurate body part detection box;

[0081] S54, initialize the state vector and covariance matrix of the Kalman filter, where the state vector includes the position coordinates (x, y) and velocity (v x , v y );

[0082] S55. In each frame of the image, the state of the target in the current frame is predicted based on the state and motion model predicted in the previous frame:

[0083]

[0084] P k|k-1 =AP k-1|k-1 A T +Q;

[0085] In the formula, represents the state vector predicted at time k based on the state at time k-1; A represents the state transfer matrix; represents the optimal estimated state vector at time k-1; B represents the control input matrix; u k represents the control input at time k; P k|k-1 represents the covariance matrix predicted at time k based on the state at time k-1; P k-1|k-1 represents the covariance matrix of the optimal estimation state at time k-1; Q represents the process noise covariance matrix;

[0086] S56, taking the detection result obtained by the target detection algorithm as the observation value, and fusing the observation value with the prediction value, and then correcting the prediction result using the Kalman gain to obtain the target state estimation of the current frame:

[0087]

[0088] P k|k =(IK k H)P k|k-1 ;

[0089] In the formula, represents the optimal estimated state vector at time k; K krepresents the Kalman gain at time k; Z k represents the observation value at time k; H represents the observation matrix; P k|k represents the covariance matrix of the optimal estimation state at time k; I represents the identity matrix;

[0090] S57, repeat steps S55-S56 to obtain the motion trajectory of the target;

[0091] S58, inputting the image containing the athlete after the target detection and trajectory tracking processing into the OpenPose model;

[0092] S59. Use the OpenPose model to extract features from the image, learn the feature representation of the human body in the image, predict the location of each key part of the athlete's body, and present the probability distribution of each key part in the image in the form of a heat map;

[0093] S510, determining the coordinates of the key parts according to the heat map, and extracting the motion parameters of the key parts of the athlete by calculating the relative position relationship and angle between the key parts.

[0094] S6. Compare the key motion parameters of the athlete with the preset data, and output the evaluation results according to the comparison results. At the same time, evaluate whether the athlete's speed change and route are in compliance by analyzing the motion trajectory;

[0095] In step S6, the preset data is preset standard data or preset historical data, wherein the preset standard data is obtained by training with data annotated by experts using a semi-supervised learning method, that is, training unlabeled data with a small amount of labeled data to obtain a large amount of labeled data;

[0096] It specifically includes the following steps:

[0097] S61, comparing the motion parameters of the key parts of the athlete extracted from the extracted sports video with preset data to obtain a comparison result;

[0098] S62. Based on the comparison results of each parameter, a comprehensive evaluation of the athlete's overall performance is conducted;

[0099] S63. Generate evaluation results based on the comprehensive evaluation results;

[0100] S64. Calculate the velocity v according to the positions and time intervals of two adjacent points on the motion trajectory:

[0101]

[0102] Where (x1, y1) and (x2, y2) represent the position coordinates of two adjacent points; Δt represents the time interval between (x1, y1) and (x2, y2);

[0103] S65. Assume that the point on the standard route is (x s ,y s ), the point on the athlete’s trajectory is (x a ,y a ), then the deviation distance calculation formula is:

[0104]

[0105] In the formula, d represents the distance from the standard route;

[0106] S67, determine whether the route is compliant: set the maximum allowable deviation distance to d max , if d>d max , the route is considered non-compliant, otherwise it is considered compliant.

[0107] In this embodiment, the rules can be learned from the key parts such as the center of gravity, hip, knee, and ankle or the empirical values ​​of the athletes' 100-meter run, standing jump, 100-meter run, standing jump, triple jump, marathon, and volleyball videos, and the pixel index of the key points can be obtained to determine the optimal range of the actual distance of the parameters such as angle, angular velocity, support time, support ratio, foot lift angle, foot landing angle, stride length, and calf length. Then, the obtained parameters such as angle, angular velocity, support time, support ratio, and stride length are compared with the actual video, and the key frames are automatically marked (first step off, first step touchdown; second step off, second step touchdown; third step off, third step touchdown; fourth step off, fourth step touchdown; fifth step off, fifth step touchdown), and the comparison results and training suggestions are given through the key frames.

[0108] The specific steps of automatically marking key frames are as follows:

[0109] Step 1: Parameter extraction and calculation: From the athletes’ 100-meter sprint, standing jump, triple jump, marathon, volleyball and other sports videos, use the target detection algorithm and posture estimation method to track the center of gravity, hip, knee, ankle and other key parts. And based on the pixel coordinates of the key parts, calculate the angle, angular velocity, support time, support-to-lift ratio, foot lift angle, foot landing angle, stride length, calf length and other parameters;

[0110] Step 2: Set the key frame judgment threshold: Based on the rules learned from a large number of motion videos and experience values ​​in the early stage, determine the reasonable range of each parameter at the key frame moment, and use this as the judgment threshold.

[0111] Step 3: Traverse the video frames to perform key frame detection: Analyze the calculated parameters frame by frame according to the time sequence of the video. In each frame, compare the extracted parameters with the set key frame judgment threshold. For example, for the running project, when detecting the first step, when the detected parameters meet the threshold conditions for the first step push-off (such as the center of gravity moving up quickly, the leg joint angle changes meeting the push-off characteristics, etc.), then mark the frame as the first step push-off key frame;

[0112] Step 4. Keyframe verification and correction: To avoid misjudgment, verify the initially marked keyframes: Analyze the parameter change trends of the frames before and after the keyframes to ensure that the parameter changes are consistent with the logic of the key actions. For example, after marking the ground contact keyframe of running, check whether the leg joint angles in the subsequent frames change according to the normal running action rules. If not, re-evaluate the accuracy of the keyframe. If a misjudgment is found, correct the keyframe according to the actual situation to ensure the accuracy of the marking;

[0113] Step 5: Label output and sorting: Label and output the verified key frames in a unified format, such as recording the frame number of the key frame and the corresponding action type (such as the first step of push-off, the second step of touchdown, etc.). The labeled key frames are sorted into a list or database to facilitate subsequent comparison with the preset optimal parameter range, and then provide comparison results and targeted training suggestions.

[0114] S7. The evaluation results are encrypted and uploaded to the user interaction interface for display, and uploaded to the cloud database for storage, wherein the encryption operation is convenient for protecting the privacy of the athletes.

[0115] In step S7, the athlete's training is guided in real time through voice broadcast.

[0116] In step S7, the evaluation results are displayed in the form of charts, animations, or 3D models through virtual reality or augmented reality.

[0117] A sports evaluation system based on computer vision is used to execute a sports evaluation method based on computer vision, which includes a multi-angle camera module, a data processing module, an analysis and comparison module, a user interaction interface and a cloud database, wherein the multi-angle camera module is used to collect sports images of athletes and targets from multiple perspectives; the data processing module is used to process the collected sports images to obtain motion trajectories and motion parameters; the analysis and comparison module is used to compare the processed motion parameters with preset data to output evaluation results, and at the same time evaluate the speed changes and route compliance of athletes by analyzing the motion trajectory; the user interaction interface is used to display the evaluation results; and the cloud database is used to store the evaluation results. The data processing module includes a graphics processing unit and a programmable gate array, and the graphics processing unit and the programmable gate array are connected via a standard interface to enable seamless connection between different devices and systems; at the same time, a modular system architecture design is adopted to facilitate the addition of new functional modules or evaluation items when needed.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A sports evaluation method based on computer vision, characterized in that: The following steps are involved: S1. Collect a series of sports images of athletes and targets from multiple perspectives; S2, matching and fusing the motion images from multiple perspectives acquired in step S1 to obtain multiple frames of motion images along the motion trajectory, wherein the multiple frames of motion images constitute a motion video stream; S3, using the non-maximum suppression algorithm improved by Soft-NMS to process the motion video stream obtained in step S2 to remove occlusion; S4, data preprocessing, and using adaptive illumination compensation algorithm to adjust the brightness and contrast of the motion video stream; S5. Use the target detection algorithm in deep learning to identify the athlete and the target, then use algorithms such as Kalman filtering or particle filtering to obtain the motion trajectories of the athlete and the target, and use the OpenPose posture estimation method to extract the motion parameters of the key parts of the athlete; S6. Compare the key motion parameters of the athlete with the preset data, and output the evaluation results according to the comparison results. At the same time, evaluate whether the athlete's speed change and route are in compliance by analyzing the motion trajectory; S7. Encrypt the evaluation results and upload them to the user interface for display, and upload them to the cloud database for storage.

2. A sports evaluation method based on computer vision according to claim 1, characterized in that: In step S1 , motion images are captured using visible light cameras and infrared cameras arranged at multiple angles along a motion path.

3. A sports evaluation method based on computer vision according to claim 1, characterized in that: In step S4, when the light intensity is lower than a set threshold, the motion video stream is processed using image enhancement technology, otherwise the background is removed using a background subtraction algorithm.

4. A sports evaluation method based on computer vision according to claim 1, characterized in that: In step S5, the target detection algorithm in deep learning is the YOLO algorithm or the Faster R-CNN algorithm; The key parts of the athlete include the athlete’s torso attachment points, shoulder, hip, knee, ankle, and elbow joints; Step S5 specifically includes the following steps: S51, pre-processing each frame image in the input motion video stream and inputting it into the YOLO network, the YOLO network divides the input image into multiple grids, each grid is responsible for detecting the target in a specific area; S52, inputting the image data with the athlete's body parts and key points labeled into the YOLO network for training, so that the YOLO network learns the feature representations of different parts and key parts; when a new image enters, the YOLO network predicts whether there is a target in each grid, and if so, outputs the category of the target and the bounding box coordinates; S53, decoding the predicted bounding box, removing the bounding boxes with too high overlap by non-maximum suppression method, and obtaining the final accurate body part detection box; S54, initialize the state vector and covariance matrix of the Kalman filter, where the state vector includes the position coordinates (x, y) and velocity (v x , v y ); S55. In each frame of the image, the state of the target in the current frame is predicted based on the state and motion model predicted in the previous frame: P k|k-1 =AP k-1|k-1 From T +Q; In the formula, represents the state vector predicted at time k based on the state at time k-1; A represents the state transfer matrix; represents the optimal estimated state vector at time k-1; B represents the control input matrix; u k represents the control input at time k; P k|k-1 represents the covariance matrix predicted at time k based on the state at time k-1; P k-1|k-1 represents the covariance matrix of the optimal estimation state at time k-1; Q represents the process noise covariance matrix; S56, taking the detection result obtained by the target detection algorithm as the observation value, and fusing the observation value with the prediction value, and then correcting the prediction result using the Kalman gain to obtain the target state estimation of the current frame: P k | k =(I-K k H)P k | k-1 ; In the formula, represents the optimal estimated state vector at time k; K k represents the Kalman gain at time k; Z k represents the observation value at time k; H represents the observation matrix; P k|k represents the covariance matrix of the optimal estimation state at time k; I represents the identity matrix; S57, repeat steps S55-S56 to obtain the motion trajectory of the target; S58, inputting the image containing the athlete after the target detection and trajectory tracking processing into the OpenPose model; S59. Use the OpenPose model to extract features from the image, learn the feature representation of the human body in the image, predict the location of each key part of the athlete's body, and present the probability distribution of each key part in the image in the form of a heat map; S510, determining the coordinates of the key parts according to the heat map, and extracting the motion parameters of the key parts of the athlete by calculating the relative position relationship and angle between the key parts.

5. A sports evaluation method based on computer vision according to claim 1, characterized in that: In step S6, the preset data is preset standard data or preset historical data, wherein the preset standard data is obtained by training with data annotated by experts using a semi-supervised learning method; It specifically includes the following steps: S61, comparing the motion parameters of the key parts of the athlete extracted from the extracted sports video with preset data to obtain a comparison result; S62. Based on the comparison results of each parameter, a comprehensive evaluation of the athlete's overall performance is conducted; S63. Generate evaluation results based on the comprehensive evaluation results; S64. Calculate the velocity v according to the positions and time intervals of two adjacent points on the motion trajectory: Where (x1, y1) and (x2, y2) represent the position coordinates of two adjacent points; Δt represents the time interval between (x1, y1) and (x2, y2); S65. Assume that the point on the standard route is (x s ,y s ), the point on the athlete’s trajectory is (x a ,y a ), then the deviation distance calculation formula is: In the formula, d represents the distance from the standard route; S67. Determine whether the route is compliant: Set the maximum allowable deviation distance to d max , if d>d max , the route is considered non-compliant, otherwise it is considered compliant.

6. A sports evaluation method based on computer vision according to claim 1, characterized in that: In step S7, the athlete's training is guided in real time through voice broadcast.

7. A sports evaluation method based on computer vision according to claim 1, characterized in that: In step S7, the evaluation results are displayed in the form of charts, animations, or 3D models through virtual reality or augmented reality.

8. A sports evaluation system based on computer vision, characterized by: A computer vision-based sports evaluation method for executing any one of claims 1 to 7, comprising a multi-angle camera module, a data processing module, an analysis and comparison module, a user interaction interface and a cloud database, wherein the multi-angle camera module is used to collect motion images of athletes and targets from multiple perspectives; the data processing module is used to process the collected motion images to obtain motion trajectories and motion parameters; the analysis and comparison module is used to compare the processed motion parameters with preset data to output evaluation results, and at the same time evaluate the speed changes and route compliance of athletes by analyzing the motion trajectory; The user interface is used to display the evaluation results; The cloud database is used to store the evaluation results.

9. A sports evaluation system based on computer vision according to claim 8, characterized in that: The data processing module includes a graphics processing unit and a programmable gate array, and the graphics processing unit and the programmable gate array are connected via a standard interface.

Images