Basketball action intelligent identification and learning effect prediction system based on AI

Abstract

The invention relates to the technical field of artificial intelligence and physical education, and discloses an AI-based basketball motion intelligent identification and learning effect prediction system, which comprises: a data acquisition device acquires a training video in a basketball motion training process, processes the training video through an edge computing device, and transmits the processed training video to a server; the edge computing device deploys a lightweight model, and carries out real-time processing on the data acquired by the data acquisition device; the server carries out model training and optimization and comprises a data preprocessing module, an action detection module and an effect prediction module, the action detection module comprises an action recognition unit, an action classification unit and an anomaly detection unit, and training videos are input into a trained action recognition model; and outputting action recognition, classification and anomaly detection results in the basketball training process. According to the invention, motion capture and analysis are realized through an AI algorithm architecture, the motion recognition accuracy is improved, and the method is suitable for physical education, exercise training and personalized skill assessment scenes.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and sports education technology, and in particular to an AI-based basketball action intelligent recognition and learning effect prediction system. Background Art

[0002] Currently, traditional basketball movement assessment relies primarily on visual inspection by coaches, a method characterized by significant subjectivity and an error rate exceeding 30%. Existing technical solutions also suffer from numerous shortcomings: wearable device solutions rely on sensor data, which is not only costly but also incapable of comprehensive full-body movement analysis; competitive-grade motion capture systems require specialized venues and markers, hindering widespread adoption in educational settings; and other related technologies generally lack professional motion algorithm analysis support. This current state of the art leads to significant problems and consequences: traditional manual assessment cycles can take up to 72 hours, resulting in significant lags and inability to provide real-time feedback to learners. Furthermore, under standardized training models, student sports injury rates are increasing by an average of 12% annually, leaving the need for personalized training unmet. Summary of the Invention

[0003] The present invention provides an AI-based basketball motion intelligent recognition and learning effect prediction system, which realizes motion capture and analysis through AI algorithm architecture, improves the accuracy of motion recognition, and is suitable for physical education teaching, sports training and personalized skill assessment scenarios.

[0004] The present invention provides an AI-based basketball action intelligent recognition and learning effect prediction system, which is characterized by comprising a data acquisition device, an edge computing device and a server, wherein the data acquisition device is connected to the server via the edge computing device, and the server comprises a data preprocessing module, a motion detection module and an effect prediction module;

[0005] The data acquisition device is used to collect training videos during basketball training, and transmit the training videos to the server after processing through the edge computing device;

[0006] The edge computing device deploys a lightweight model for real-time processing of data collected by the data collection device;

[0007] The server is used for model training and optimization, and includes a data preprocessing module, a motion detection module and an effect prediction module, wherein the motion detection module includes a motion recognition unit, a motion classification unit and an anomaly detection unit, which is used to input the training video into the trained motion recognition model to output motion recognition, classification and anomaly detection results during basketball training.

[0008] Furthermore, the data acquisition equipment includes a visual sensor and an inertial measurement unit. The visual sensor uses a Hikvision DS-2CD8A87F camera to collect training videos during basketball training. The inertial measurement unit is used to capture the motion acceleration and angular velocity of key parts of the player's wrist and ankle during basketball.

[0009] Furthermore, the edge computing device stores the training video of the basketball training process, and decomposes the training video to be detected into a set of image frames to be identified, and transmits the set of image frames to be identified to the data preprocessing module.

[0010] Furthermore, the data preprocessing module uses the cv2.undistort() function of OpenCV to process each frame of the image frame set to eliminate wide-angle distortion based on the calibration parameters of the Hikvision DS-2CD8A87F camera.

[0011] Furthermore, in the action detection module, the action recognition unit adopts the MMPose model to output the coordinates of the player's motion key points of the image frame set to realize action recognition during basketball training, specifically including:

[0012] Create a Python virtual environment, install PyTorch and torchvision, install MMPose and its dependent libraries, and install visualization tools;

[0013] Select the pre-trained PoseNet as the 3D pose estimation model, determine the model configuration file and pre-trained weights, initialize the model and load the dataset information, and obtain the joint name and index mapping;

[0014] Read the image frame set, input the image frame set into the 3D posture estimation model, the model output includes the 3D joint coordinates of each human body, extract the required joint coordinate data, where the joint coordinates are represented as

[0015] [x,y,z,confidence], where x, y, and z are three-dimensional coordinates and confidence is the confidence score.

[0016] Furthermore, in the action detection module, the action classification unit connects the key point coordinates of each frame in the image frame set, and sorts the image frames in the image frame set in chronological order, so as to identify the eight core actions of basketball according to the sorted connection results and the preset standard action range, so as to realize action classification in the basketball training process.

[0017] Furthermore, in the action detection module, the abnormality detection unit connects the key point coordinates of each image frame in the image frame set, calculates the angles of multiple joint angles, and determines the changes in multiple joint angles of the player according to the time sequence of the image frames, and sets an abnormal angle threshold for each joint angle change. When the joint angle change exceeds the range set by the corresponding threshold, the force error at that location is marked to realize abnormality detection during basketball training.

[0018] Furthermore, the calculation of the angles of multiple joint angles specifically includes: determining the joint combination for which the angle is to be calculated, wherein an angle is defined by three joint points, extracting the three-dimensional coordinates of three related joints for each joint angle, and calculating the two vectors formed by the intermediate joint points, and finally using the vector dot product to calculate the angle between the two vectors to convert radians into angles.

[0019] Furthermore, the effect prediction module evaluates the player's action completion according to the angle change of the joint angle calculated by the key point connection line, and sets the scoring calculation rules: split and set the weight according to each action,

[0020] If the upper arm is at 90° with the body, 7 points will be awarded. The allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 2 points will be deducted. If it is lower than 80° or higher than 100°, 5 points will be deducted.

[0021] If the forearm and upper arm are 90°, 7 points will be awarded; the allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 2 points will be deducted; if it is lower than 80° or higher than 100°, 5 points will be deducted;

[0022] If the wrist and forearm are at 90°, 3 points will be awarded. The allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 1 point will be deducted. If it is below 80° or above 100°, 3 points will be deducted.

[0023] If the distance between the palm and the ball is 2-3 fingers, 3 points will be awarded. If the distance is outside this range, 1-3 points will be deducted.

[0024] Four weights are set according to the importance of the above four actions, and the sum of the four weights is equal to 1. Finally, the player's action completion score is calculated based on the four action scoring results and their corresponding weights.

[0025] Furthermore, the server is also connected to multiple terminals such as mobile phones, tablets and PCs to view the training and detection data in real time and synchronously.

[0026] The beneficial effects of the present invention are:

[0027] The data acquisition device of the present invention collects training videos during basketball training and transmits them to the server; the edge computing device deploys a lightweight model to perform real-time data processing on the data acquisition device; the server performs model training and optimization, inputs the training video into the action recognition model, and outputs the action recognition, classification and anomaly detection results during basketball training. High-precision action recognition is achieved, the recognition accuracy of core actions (such as shooting, dribbling, layups, etc.) is improved, and real-time synchronous viewing of training data by multiple terminals such as mobile phones, tablets, and PCs is supported; based on the mmpose model, the training data of athletes can be analyzed, the trend of skill improvement can be predicted, and personalized training suggestions can be provided (such as adjusting the shooting angle and optimizing the dribbling rhythm). Combined with the force trajectory analysis, it can effectively reduce the risk of sports injuries and is suitable for physical education, sports training and personalized skill assessment scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a structural diagram of the AI-based basketball action intelligent recognition and learning effect prediction system of the present invention.

[0029] Figure 2 This is a schematic diagram of connecting key points for on-site identification in the present invention.

[0030] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0031] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0032] like Figure 1-2 As shown, the present invention provides an AI-based basketball action intelligent recognition and learning effect prediction system, including a data acquisition device, an edge computing device and a server. The data acquisition device is connected to the server through the edge computing device. The server includes a data preprocessing module, an action detection module and an effect prediction module; the server is also connected to multiple terminals such as mobile phones, tablets and PCs to view training detection data in real time and synchronously.

[0033] (1) Data acquisition equipment

[0034] The data acquisition device is used to collect training videos during basketball training, and transmit the training videos to the server after processing through the edge computing device.

[0035] The data acquisition equipment includes a visual sensor and an inertial measurement unit. The visual sensor, a Hikvision DS-2CD8A87F camera (1280×720 resolution, 30fps), supports simultaneous RGB-D data acquisition and a field of view of ≥86° (horizontal) × 75° (vertical). Multi-camera networking can achieve coverage of ≥110°. The visual sensor captures training videos from basketball training sessions, employing image recognition and correction algorithms to ensure stability in complex environments.

[0036] The Inertial Measurement Unit (IMU) is used to capture the motion acceleration and angular velocity of the player's wrist and ankle key parts during basketball movement.

[0037] (2) Edge computing devices

[0038] The edge computing device deploys a lightweight model to process the data collected by the data acquisition device in real time. It stores training videos of the basketball training process, breaks down the training videos to be detected into a set of image frames to be identified, and transmits the set of image frames to be identified to the data preprocessing module. Furthermore, model quantization technology can be employed to enable the system to operate on the edge computing device.

[0039] (3) Server

[0040] The server uses Intel I9-14900KF for model training and optimization, and includes a data preprocessing module, a motion detection module, and an effect prediction module. The motion detection module includes a motion recognition unit, a motion classification unit, and an anomaly detection unit, which is used to input the training video into the trained motion recognition model to output motion recognition, classification, and anomaly detection results during basketball training.

[0041] 1) Data preprocessing module

[0042] The data preprocessing module uses the cv2.undistort() function of OpenCV to process each frame of the image frame set to eliminate wide-angle distortion based on the calibration parameters of the Hikvision DS-2CD8A87F camera.

[0043] 2) Motion detection module

[0044] The action detection module inputs the training video into a trained action recognition model to output action recognition, classification and anomaly detection results during basketball training.

[0045] a. Action recognition unit

[0046] The action recognition unit uses the MMPose model to output the coordinates of the player's motion key points of the image frame set to realize action recognition during basketball training, specifically including:

[0047] Create a Python virtual environment, install PyTorch and torchvision, install MMPose and its dependent libraries (including MMCV and MMDetection), and install visualization tools; build a joint point detection model for 8 core actions (shooting, dribbling, etc.) based on PyTorch.

[0048] Select the pre-trained PoseNet as the 3D pose estimation model, determine the model configuration file and pre-trained weights, initialize the model and load the dataset information, and obtain the joint name and index mapping;

[0049] Read the image frame set, input the image frame set into the 3D posture estimation model, the model output includes the 3D joint coordinates of each human body, extract the required joint coordinate data, where the joint coordinates are represented as

[0050] [x,y,z,confidence], where x, y, and z are three-dimensional coordinates and confidence is the confidence score.

[0051] b. Action classification unit

[0052] The action classification unit connects the key point coordinates of each frame in the image frame set and sorts the image frames in the image frame set in chronological order, so as to identify eight core actions of basketball (such as shooting and cross dribbling) according to the sorted connection results and the preset standard action range, so as to realize action classification in the basketball training process.

[0053] c. Anomaly detection unit

[0054] After connecting the key point coordinates of each image frame in the image frame set, the anomaly detection unit calculates the angles of multiple joint angles, and determines the changes in multiple joint angles of the player according to the time sequence of the image frames, and sets an abnormal angle threshold for each joint angle change. When the joint angle change exceeds the range set by its corresponding threshold, the force error at that location is marked to realize anomaly detection during basketball training; at the same time, a force trajectory deviation report is generated in real time (such as elbow joint abduction angle error ±10°).

[0055] The calculation of the angles of multiple joint angles specifically includes: determining the joint combination for which the angle is to be calculated, wherein an angle is defined by three joint points, extracting the three-dimensional coordinates of three related joints for each joint angle, and calculating two vectors formed by the intermediate joint points, and finally using the vector dot product to calculate the angle between the two vectors to convert radians into angles.

[0056] Finally, by overlaying the 2D pose estimation results on the image frames, we can create a 3D visualization that shows the human skeletal structure, annotate the calculated joint angles in the 3D visualization, and analyze the angle data for subsequent applications (such as motion analysis, posture assessment, etc.).

[0057] 3) Effect prediction module

[0058] The effect prediction module evaluates the player's action completion based on the angle change of the joint angle calculated by the key point connection line, and generates a score based on the action completion (such as the shooting angle ±5°) and smoothness (action continuity score). Set the action completion score calculation rules: split and set the weight according to each action,

[0059] If the upper arm is at 90° with the body, 7 points will be awarded. The allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 2 points will be deducted. If it is lower than 80° or higher than 100°, 5 points will be deducted.

[0060] If the forearm and upper arm are 90°, 7 points will be awarded; the allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 2 points will be deducted; if it is lower than 80° or higher than 100°, 5 points will be deducted;

[0061] If the wrist and forearm are at 90°, 3 points will be awarded. The allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 1 point will be deducted. If it is below 80° or above 100°, 3 points will be deducted.

[0062] If the distance between the palm and the ball is 2-3 fingers, 3 points will be awarded. If the distance is outside this range, 1-3 points will be deducted.

[0063] Four weights are set according to the importance of the above four actions, and the sum of the four weights is equal to 1. Finally, the player's action completion score is calculated based on the four action scoring results and their corresponding weights.

[0064] This invention can be used in elementary and middle school basketball classes. Deployment: Four Hikvision DS-2CD8A87F cameras are installed on both sides of the court to provide full field coverage. Usage: After students complete shooting training, the system displays data such as hand angle and ball speed in real time. The teacher's app receives analysis reports (e.g., "Student A's elbow is not straightened when shooting. Correction suggestion: Strengthen triceps training").

[0065] The present invention achieves the following technical effects:

[0066] 1. High-precision action recognition. Utilizing deep learning algorithm technology, the recognition accuracy of core actions (such as shooting, dribbling, and layups) is ≥85%, significantly lower than the 30% error rate of traditional manual evaluation.

[0067] 2. Real-time feedback and low latency. End-to-end latency from motion capture to analysis report generation is ≤ 5 seconds, providing instant feedback compared to traditional manual evaluation (which typically takes 72 hours), significantly improving training efficiency. Support for real-time synchronization across multiple devices (such as mobile phones, tablets, and PCs) allows coaches and students to access training data at any time.

[0068] 3. Low cost and high adaptability. Deployment is simple with standard cameras (such as the Hikvision DS-2CD8A87F), eliminating the need for expensive optical motion capture equipment and reducing hardware costs. The system is compatible with cameras with a field of view of 110° or greater, suitable for standard basketball court environments, without the need for additional markers or special venue modifications.

[0069] 4. Personalized training optimization. Based on the mmpose model, it can analyze athletes' training data, predict skill improvement trends, and provide personalized training suggestions (such as adjusting shooting angles and optimizing dribbling rhythm). Combined with force trajectory analysis (such as elbow abduction angle error of ±10°), it can effectively reduce the risk of sports injuries (pilot data shows a 40% reduction in training injury rates).

[0070] 5. Scalability and industrial value. The system architecture supports modular expansion and can be adapted to other sports (such as football, badminton, gymnastics, etc.) in the future.

[0071] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0072] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An AI-based basketball action intelligent recognition and learning effect prediction system, characterized by: It includes a data acquisition device, an edge computing device and a server, wherein the data acquisition device is connected to the server through the edge computing device, and the server includes a data preprocessing module, an action detection module and an effect prediction module; The data acquisition device is used to collect training videos during basketball training, and transmit the training videos to the server after processing through the edge computing device; The edge computing device deploys a lightweight model for real-time processing of data collected by the data collection device; The server is used for model training and optimization, and includes a data preprocessing module, a motion detection module and an effect prediction module, wherein the motion detection module includes a motion recognition unit, a motion classification unit and an anomaly detection unit, which is used to input the training video into the trained motion recognition model to output motion recognition, classification and anomaly detection results during basketball training.

2. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 1 is characterized in that: The data acquisition equipment includes a visual sensor and an inertial measurement unit. The visual sensor uses a Hikvision DS-2CD8A87F camera to collect training videos during basketball training. The inertial measurement unit is used to capture the motion acceleration and angular velocity of key parts of the player's wrist and ankle during basketball.

3. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 2 is characterized in that: The edge computing device stores the training video during the basketball training process, breaks down the training video to be detected into a set of image frames to be identified, and transmits the set of image frames to be identified to the data preprocessing module.

4. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 3 is characterized in that: The data preprocessing module uses the cv2.undistort() function of OpenCV to process each frame of the image frame set to eliminate wide-angle distortion based on the calibration parameters of the Hikvision DS-2CD8A87F camera.

5. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 4 is characterized in that: In the action detection module, the action recognition unit adopts the MMPose model to output the coordinates of the player's motion key points of the image frame set to realize action recognition during basketball training, specifically including: Create a Python virtual environment, install PyTorch and torchvision, install MMPose and its dependent libraries, and install visualization tools; Select the pre-trained PoseNet as the 3D pose estimation model, determine the model configuration file and pre-trained weights, initialize the model and load the dataset information, and obtain the joint name and index mapping; The image frame set is read and input into the 3D pose estimation model. The model output includes the 3D joint coordinates of each human body, and the required joint coordinate data is extracted, where the joint coordinates are expressed as [x, y, z, confidence], where x, y, and z are three-dimensional coordinates and confidence is a confidence score.

6. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 5 is characterized in that: In the action detection module, the action classification unit connects the key point coordinates of each frame in the image frame set and sorts the image frames in the image frame set in chronological order, so as to identify the eight core actions of basketball according to the sorted connection results and the preset standard action range, so as to realize action classification in the basketball training process.

7. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 6 is characterized in that: In the action detection module, the anomaly detection unit connects the key point coordinates of each image frame in the image frame set, calculates the angles of multiple joint angles, determines the changes in the player's multiple joint angles according to the time sequence of the image frames, and sets an abnormal angle threshold for each joint angle change. When the joint angle change exceeds the range set by its corresponding threshold, the force error at that location is marked to achieve anomaly detection during basketball training.

8. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 7 is characterized in that: The calculation of the angles of multiple joint angles specifically includes: determining the joint combination for which the angle is to be calculated, wherein an angle is defined by three joint points, extracting the three-dimensional coordinates of three related joints for each joint angle, and calculating two vectors formed by the intermediate joint points, and finally using the vector dot product to calculate the angle between the two vectors to convert radians into angles.

9. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 7, characterized in that: The effect prediction module evaluates the player's action completion according to the angle change of the joint angle calculated by the key point connection line, and sets the scoring calculation rules: split and set the weight according to each action, If the upper arm is at 90° with the body, 7 points will be awarded. The allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 2 points will be deducted. If it is lower than 80° or higher than 100°, 5 points will be deducted. If the forearm and upper arm are 90°, 7 points will be awarded; the allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 2 points will be deducted; if it is lower than 80° or higher than 100°, 5 points will be deducted; If the wrist and forearm are at 90°, 3 points will be awarded. The allowable error range is between 85° and 95°. If it is between 80° and 85° or between 95° and 100°, 1 point will be deducted. If it is below 80° or above 100°, 3 points will be deducted. If the distance between the palm and the ball is 2-3 fingers, 3 points will be awarded. If the distance is outside this range, 1-3 points will be deducted. Four weights are set according to the importance of the above four actions, and the sum of the four weights is equal to 1. Finally, the player's action completion score is calculated based on the four action scoring results and their corresponding weights.

10. The AI-based basketball action intelligent recognition and learning effect prediction system according to claim 1, characterized in that: The server is also connected to multiple terminals such as mobile phones, tablets and PCs to view training and detection data in real time and synchronously.

Images