Intelligent management system for sports basketball training risk data

Abstract

The invention discloses an intelligent management system for sports basketball training risk data, which relates to the technical field of sports data analysis and comprises an acquisition module for generating three-dimensional tracks of athletes, an identification module for identifying risk actions, an analysis module for analyzing kinematic characteristics of the risk actions and a feedback module. According to the invention, data acquisition is carried out through multi-view video capture and a three-dimensional reconstruction technology, through real-time analysis of features such as postures and joint angles of athletes, the system can provide customized action suggestions for each athlete, the training quality is effectively improved, iterative updating of a deep learning model is carried out according to an evaluation result of a feedback module, and the training efficiency is improved. Along with the promotion of the training process, the recognition precision and risk judgment capability of the system are continuously improved, and more accurate risk action recognition is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of sports data analysis, and specifically to an intelligent management system for sports basketball training risk data. Background Art

[0002] In the field of intelligent sports training, sports training data analysis based on intelligent recognition is becoming an important research topic for optimizing athlete training and improving athlete performance. This system covers multiple aspects such as the collection and analysis of motion data. In high-intensity sports training projects such as basketball training, although traditional training methods rely on the experience of coaches, they lack quantitative data support and are difficult to accurately judge the technical details and potential risk actions of athletes during training. For example, high-intensity confrontation in basketball training is likely to cause athletes to get injured, especially in high-intensity training, athletes are prone to perform non-standard actions or sudden physical collisions.

[0003] Most current intelligent recognition systems rely on image recognition technology to analyze the training performance of athletes by recognizing actions in images. However, traditional methods have problems such as image occlusion and pose estimation errors when facing high-dynamic environments, such as rapid position changes and multi-person interactions in basketball training, and cannot accurately capture the real actions of athletes. In addition, the existing systems have limited ability to identify and classify potential risk actions of athletes, such as excessive turning, rapid direction change, excessive knee bending, etc., and are difficult to effectively predict and timely feedback actions that may cause injuries. Summary of the Invention

[0004] I) Technical Problems to be Solved

[0005] The present invention provides an intelligent management system for sports basketball training risk data, which can identify potential risk actions in athlete training.

[0006] II) Technical Solutions

[0007] To achieve the above object, the present invention provides the following technical solutions: An intelligent management system for sports basketball training risk data, comprising:

[0008] A collection module, based on visual capture technology, simultaneously performs two-dimensional video capture on athletes from different shooting angles within the training ground, constructs a three-dimensional space coordinate based on the training ground, performs three-dimensional reconstruction on the captured two-dimensional video, and generates a three-dimensional trajectory of the athlete;

[0009] The recognition module, based on the trained deep learning model, performs pose analysis on the three-dimensional trajectory of the athlete, identifies the risk actions pre-labeled by the learning model therefrom, and takes the two-dimensional image frames captured from different shooting angles when such risk actions occur as key frames, and traces back a plurality of set numbers of image frames forward according to the time series based on the key frames. The image frames traced back forward and the key frames together constitute a risk image frame set;

[0010] The analysis module is used to perform pose analysis on the risk image frame set and the three-dimensional trajectory at the corresponding moment to extract kinematic features. The kinematic features include the joint angles of the athlete, the movement trajectories of the limbs, the movement speed, and the acceleration;

[0011] The feedback module receives the extracted kinematic features, compares them with the corresponding preset standard ranges, respectively obtains the feature deviations between the joint angle detection value, the movement speed detection value, the acceleration detection and the corresponding preset standard ranges, and performs weighted processing on each of the feature deviations according to the preset weights to evaluate the risk level of the athlete's current risk action.

[0012] Further, the acquisition module captures real-time videos of the athlete from different angles in the training ground, extracts features from the captured multiple two-dimensional images, then constructs a three-dimensional coordinate space of the training ground, and calculates the depth information of the athlete in the training ground through feature matching and triangulation of each two-dimensional image;

[0013] Based on the depth information calculated from different perspectives, the three-dimensional coordinates of all the matched feature points are integrated to construct a three-dimensional point cloud in the training ground;

[0014] Based on the three-dimensional point cloud data reconstructed in three dimensions, track the position change of the athlete in the three-dimensional space; among them, the three-dimensional coordinates corresponding to each frame of two-dimensional image can be used to deduce the movement path of the athlete by comparing the point cloud data of adjacent frames, so as to obtain the three-dimensional trajectory of the athlete.

[0015] Further, the recognition module identifies the defined risk actions from the three-dimensional trajectory through the trained deep learning model. Among them, the labeled risk actions are defined as actions that pose a threat to the safety of the athlete, and the deep learning model is trained to identify the labeled risk actions from the three-dimensional trajectory through a temporal convolutional network.

[0016] Further, when the recognition module is processing the three-dimensional data of the athlete in real time, the deep learning model detects and judges the risk actions by extracting the kinematic features in the three-dimensional trajectory of the athlete and comparing them with the preset feature recognition threshold.

[0017] Further, when the recognition module detects a risky action, the multi-perspective two-dimensional image frames captured at that moment are defined as the key frames; among them, the key frames are two-dimensional images captured based on multiple perspectives, and each of the two-dimensional images provides visual information of the athlete in different directions;

[0018] Taking the key frame as the time reference, a set number of image frames are traced back forward; among them, the traced-back image frames are two-dimensional images captured from different perspectives at the same moment;

[0019] According to the time sequence, the key frame and the set number of traced-back image frames form the risky image frame set.

[0020] Further, the recognition module extracts the joint positions of the athlete in each frame from the risky image frame set, and calculates the joint angles based on the extracted joint positions;

[0021] Based on the time sequence, the dynamic angle changes of the joints during the occurrence of the risky action are analyzed using the calculated joint angles.

[0022] Further, the recognition module calculates the movement speed and acceleration of the athlete based on the three-dimensional trajectory data corresponding to the risky image frame set; specifically, the movement speed and acceleration are calculated based on the time change of the positions of each joint of the athlete in three-dimensional space.

[0023] Further, the feedback module receives the kinematic features extracted from the analysis module, including the joint angles, movement speed, and acceleration at the moment when the athlete performs a risky action recognized by the deep learning model, and combines the preset kinematic criteria, including the thresholds of different joint angles and the upper limit speed of movement, to calculate the risk score of the current action. If the risk score is greater than the upper limit value of the set range, the risky action is evaluated as a high risk; if the risk score is less than the lower limit value of the preset range, the risky action is evaluated as a low risk.

[0024] Further, the deep learning model generates a prediction result for the current risky action of the athlete based on the previous training data, compares the received risk score with the prediction result, and if there is a deviation beyond the preset range, the weights of the network are adjusted using the backpropagation algorithm according to the kinematic features extracted from the evaluation of the analysis module, and the parameters of the model are updated.

[0025] III) Beneficial effects:

[0026] Compared with the prior art, the invention has the following beneficial effects:

[0027] The present invention captures the actions of athletes in real-time through two-dimensional videos from multiple angles. Based on visual capture technology and three-dimensional reconstruction algorithms, it generates the three-dimensional trajectories and pose information of the athletes, comprehensively and accurately capturing the motion data of the athletes without disturbing them.

[0028] Based on a deep learning model, it conducts pose analysis on the three-dimensional trajectories of the athletes to identify potential risky actions. When the system detects a risky action, it captures the current multi-view two-dimensional image frames as key frames and backtracks the multi-view image frames forward to form a set of image frames for subsequent analysis. It conducts pose analysis on the backtracked set of two-dimensional image frames to extract kinematic features such as joint angles, motion trajectories, motion speeds, and accelerations. By deeply analyzing the key actions of the athletes, it evaluates their risk levels to ensure the accurate identification and analysis of high-risk actions. At the same time, the evaluation results will be input into the deep learning model as training data to optimize the model's ability to identify risky actions and continuously improve the system performance. Brief Description of the Drawings

[0029] Figure 1 It is a schematic diagram of the scenario of an intelligent management system for sports basketball training risk data provided by an embodiment of the present invention;

[0030] Figure 2 It is a flowchart of the working process of the acquisition module in an intelligent management system for sports basketball training risk data provided by an embodiment of the present invention;

[0031] Figure 3 It is a flowchart of the working process of the identification module in an intelligent management system for sports basketball training risk data provided by an embodiment of the present invention;

[0032] Figure 4 It is a block diagram of the principle of an intelligent management system for sports basketball training risk data provided by an embodiment of the present invention;

[0033] In the figure:

[0034] 10. Acquisition module; 20. Identification module; 30. Analysis module; 40. Feedback module. Detailed Embodiments

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0036] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0037] It should be noted that, without conflict, the features in the embodiments of the present invention can be combined with each other.

[0038] In the field of intelligent sports training, the sports training system based on intelligent recognition is becoming an important tool for optimizing athletes' training and improving their performance. In basketball training, although the traditional training methods rely on the coach's experience, they lack quantitative data support and it is difficult to accurately judge the technical details and potential risky actions of athletes during training. The high-intensity confrontation in basketball training is likely to cause athletes to get injured. Especially during high-intensity training, athletes are prone to make non-standard actions or sudden physical collisions.

[0039] Most current intelligent recognition systems rely on visual recognition technology to evaluate athletes' performance through pose estimation and action recognition. However, in basketball training, due to athletes frequently changing positions quickly and interacting with other athletes, problems such as body occlusion are caused, resulting in the existing technology being unable to accurately capture athletes' actions in some scenarios. In addition, the existing systems also lack the ability to classify and predict risky actions and cannot effectively identify the actions that may cause injuries during training.

[0040] Therefore, designing a risk action data management system based on intelligent recognition, which can provide comprehensive support from aspects such as data collection, action analysis, risk identification to feedback management, not only helps to improve athletes' training performance, but also can effectively ensure athletes' safety. Combining Figures 1 to 4 As shown, to solve the above problems, the embodiments of the present invention provide an intelligent management system for sports basketball training risk data. Through real-time monitoring and deep learning analysis, the system can accurately identify potential risky actions of athletes during training, provide early warnings in a timely manner, reduce the risk of athletes getting injured, and can achieve safety guarantee and training optimization in basketball training under the framework of intelligent monitoring and risk assessment, promoting the development of intelligent sports technology.

[0041] Herein refer to Figure 4, an embodiment of the present invention proposes a risk data intelligent management system that captures three-dimensional space trajectories, trains deep learning models, iteratively learns and classifies risk actions, and finally retraces and analyzes image frames.

[0042] Specifically, it is first the acquisition module 10 for collecting athlete training data. The acquisition module 10 constructs the three-dimensional trajectory and pose information of the athlete in the training venue through multi-angle video acquisition. In some embodiments of the present invention, this module uses multiple video cameras to simultaneously capture the actions of the athlete in the training venue from different angles, and reconstructs the position and pose of the athlete in the three-dimensional space through multi-view two-dimensional video data to generate the three-dimensional trajectory data of the athlete.

[0043] More specifically, in some feasible embodiments of the present invention, multiple cameras are arranged in the training venue to ensure real-time video capture of the athlete from different angles (such as different positions or different heights). Among them, each camera obtains a two-dimensional image of the athlete, and these images record the body position and pose of the athlete at different time points. It should be noted here that the cameras are arranged at different positions to ensure the capture of the action information of the athlete at different angles during training, ensuring comprehensive motion trajectory and pose data, and each camera works synchronously to ensure that the time stamps of the video frames from each perspective are consistent to avoid time sequence confusion.

[0044] Feature extraction is performed on multiple two-dimensional images. Commonly used features include corner points, edges, textures, etc., which are extracted through algorithms such as SIFT, SURF, and ORB. Then, through image matching technology, the same object points in different perspectives are found, such as the joints and body parts of the athlete, and their positions in different images. The goal of the matching is to find the projections of the same object point in multiple perspectives.

[0045] Based on the principle of triangulation, according to the image information from multiple perspectives, the depth of the object point (i.e., the distance from the camera) is calculated. Among them, each pair of matched points can calculate the coordinates of the object in the three-dimensional space according to the position difference (parallax) in the images. For example, if two cameras are used, the depth information of the point can be calculated by calculating the position difference (parallax) of the same object point in the two perspectives, and the three-dimensional coordinates can be obtained.

[0046] Based on the depth information calculated from different perspectives, the three-dimensional coordinates of all matched object points are integrated to construct a three-dimensional point cloud of the scene. Using these three-dimensional points, the three-dimensional trajectory of the athlete in the venue can be further deduced.

[0047] Combined with the above, the process of generating the three-dimensional trajectory and pose information of the athlete depends on the three-dimensional reconstruction of the two-dimensional video images captured from multiple perspectives and then analyzing in combination with the athlete's movement process. Specifically, based on the point cloud data reconstructed in three dimensions, the position change of the athlete in the three-dimensional space is tracked. The three-dimensional coordinates corresponding to each frame of the image can be used to calculate the movement path of the athlete by comparing the point cloud data of adjacent frames, thereby obtaining the three-dimensional trajectory of the athlete. In some embodiments of the present invention, algorithms such as Kalman filtering can be used to smooth the position of the athlete to obtain more accurate trajectory information.

[0048] For reference here Figure 1 , three cameras are set up on the basketball training ground, and the cameras are located at different positions on the training ground. Here, it should be noted that it is necessary to ensure that the viewing angles and positions of these three cameras will cover the entire training ground so that they can simultaneously capture the movement of the athlete on the field. In Figure 1 the schematic diagram shown, the first camera is located at a corner on the side of the basketball hoop on the training ground, with a top-down view angle, mainly shooting a part of the field; the second camera is located at the front-facing angle of the basketball hoop, slightly tilted, which can supplement the area that the first camera cannot see; the third camera is located on the side of the training ground, providing a side view angle and being able to capture the side movements of the athlete.

[0049] Each camera independently captures two-dimensional video images. In order to reconstruct three-dimensional information from three shooting perspectives, after determining the spatial position and orientation of each camera, that is, after constructing a three-dimensional space coordinate system based on the training ground, first determine the coordinates of each camera in this coordinate system. In addition, it is also necessary to understand the basic parameters such as the focal length, field of view, and distortion of each camera to ensure that each point in the two-dimensional image can be correctly mapped into the three-dimensional space. In some embodiments, the points in these two-dimensional images are converted into three-dimensional coordinates by the triangulation method.

[0050] Specifically, in the two-dimensional image frames captured by each camera, identify the joint positions and body positions of the athlete as feature points, and obtain the two-dimensional coordinates of these feature point positions in their respective perspectives. Through image processing and matching algorithms, find the corresponding relationships of these feature points in different perspectives. Here, the above-mentioned SURF or SIFT algorithms can be used to help identify and match these feature points. Suppose now that the two-dimensional coordinates of the left knee in the three cameras are known as (x1, y1), (x2, y2), and (x3, y3) respectively, as well as the position and orientation of each camera. According to the camera parameters of each camera, the two-dimensional coordinates can be projected into the three-dimensional space to obtain a three-dimensional point (X i , Y i , Z i ), which is the position of the athlete's left knee in the three-dimensional space.

[0051] In summary, it can be concluded that the two-dimensional images of the multi-view video captured are converted into three-dimensional data through feature matching and triangulation. By further analyzing the reconstructed three-dimensional point cloud data, the position changes of the athlete in space are identified to construct the three-dimensional trajectory information of the athlete. Through these steps, the acquisition module 10 can accurately track the movement trajectory of the athlete and generate three-dimensional pose information, thereby providing data support for the subsequent risk action analysis and feedback module 40.

[0052] In addition, it can be understood that compared with three-dimensional sensors, the hardware cost of multi-angle two-dimensional video acquisition is lower. The number and arrangement of cameras can be flexibly adjusted according to requirements. The two-dimensional video system can be more flexibly deployed and expanded to adapt to different training scenarios and requirements. Even if the site environment changes, the cameras can be flexibly moved and adjusted, while three-dimensional devices may need to be reconfigured. Moreover, the current three-dimensional reconstruction, pose estimation, and action recognition technologies based on multi-view two-dimensional videos are relatively mature, and many off-the-shelf deep learning models can efficiently complete these tasks. For example, using a trained deep learning model (such as Open Pose, HR Net, etc.) for human pose estimation can deduce the joint positions and movement trajectories in three-dimensional space based on multiple views. More importantly, multi-angle two-dimensional video acquisition does not need to rely on athletes wearing sensors or marker points. The system can complete all tasks only through video capture and calculation, without disturbing the normal training of athletes and without restricting the freedom of movement of athletes.

[0053] The acquisition module 10 captures the athlete's actions through multi-view videos, and the extraction module analyzes these image frames in real time to identify whether a high-risk action has occurred. Regarding the extraction module, its purpose is to identify the risk actions of the athlete and conduct a detailed analysis by backtracking based on key frames.

[0054] More specifically, the extraction module performs real-time analysis on the captured motion data through a deep learning algorithm to identify whether there are high-risk actions. The system determines whether the athlete's actions conform to the preset "risk action" criteria based on the deep learning model of the training data. Regarding the training of the deep learning model, in some embodiments, first, a labeled data set containing known risk actions needs to be established. By labeling the key frames of the athlete in the video, it can be indicated which actions belong to risk actions, such as excessive knee bending, touching the ground, sudden turning, multi-person confrontation, etc. These labeled data will be used as the training set of the deep learning model. Use deep learning models (such as convolutional neural network CNN, long short-term memory network LSTM, etc.) to train the action recognition model. The input of the model is the video frame or its corresponding pose information, and the output is the classification result of the action, determining whether it is a risk action. For example, the deep learning model will learn how to identify high-risk actions through features such as joint angles, movement speeds, and accelerations. As the training data increases and the labeled actions are improved, the model will be continuously iterated to optimize the accuracy of its risk recognition.

[0055] More specifically, with reference herein Figure 2 , through the trained deep learning model, perform real-time pose analysis on the three-dimensional trajectory of the athlete to determine whether the athlete's actions belong to known risk actions. Common models include temporal convolutional network (TCN), long short-term memory network (LSTM), etc., which are used to identify actions from time series data. Among them, pose estimation is to identify the joint angles, relative positions and angles of body parts of the athlete by analyzing the results of three-dimensional reconstruction. A common technique is to analyze the athlete's pose through a skeleton model, and the positions and angles of each joint of the athlete can be deduced.

[0056] It can be understood that when the recognition module 20 processes the three-dimensional trajectory data of the athlete in real time, it will determine whether the current action belongs to the defined risk actions. Based on the deep learning model, the system will analyze the pose changes and detect risk actions based on the preset threshold. The detection process depends on the joint angles, position changes, and relative motion patterns of each frame of data. If a risk action is detected, the system will mark the current frame as a key frame and start the subsequent image backtracking mechanism.

[0057] Regarding key frames, a key frame refers to the two-dimensional image at the moment when a risk action is detected. These image frames can provide the spatial position and pose information of the athlete. Generally speaking, frames with obvious risk characteristics or abnormal actions are selected as key frames. For example, if an athlete's knee shows excessive bending during a turn, this frame of image is a key frame. When a risk action is detected, the system marks this frame as the starting point for analysis and conducts subsequent image backtracking. In some embodiments, threshold judgments, such as joint angles, speed changes, etc., can also be used to define what constitutes a key frame. Once a certain action or pose exceeds the threshold, the system immediately records this frame as a key frame and starts capturing the previous few frames of images. It should be noted here that the system uses multi-view two-dimensional video acquisition. Therefore, when the system identifies a certain risk action, the system will capture the multi-view two-dimensional image frames at that moment. For example, assuming there are multiple cameras (with different perspectives), the image frames captured by each camera are considered to be the two-dimensional projections of the athlete at this moment. These image frames provide a multi-dimensional view of the athlete's pose from different angles, facilitating subsequent analysis.

[0058] Regarding the image backtracking mechanism, specifically, once a key frame is marked, the system will then backtrack the image frames forward. These backtracked image frames also come from different perspectives, and these frames conduct supplementary analysis on the actions and joint angle changes in three-dimensional space. During the backtracking process, the system can analyze all the key poses and joint angle changes of the athlete before performing this action. Through these backtracked images, the system can help analyze how the action changes from normal to risky. For example, when an action of excessive knee bending is detected, backtracking can help the system find out the process of the increasing knee bending, so as to analyze whether the action gradually causes harm. Regarding the number of backtracked image frames, it is set according to the actual situation of training and is not specifically set in the embodiments of the present invention.

[0059] More specifically, when a certain frame is identified as a key frame, the system will backtrack the previous image frames through the time series. These backtracked frames come from multiple cameras (different perspectives), so each backtracked frame represents the two-dimensional projection of the athlete at different times from different angles. For each backtracked frame, the system synchronizes the perspectives of multiple cameras to ensure that each backtracked frame is consistent with the time point of the key frame. In this way, the backtracked image frames can accurately reflect the specific pose and action of the athlete before the risk action occurs.

[0060] The set of risk image frames is composed of the backtracked multi-view image frames and the key frames. Based on these sets of image frames, the system conducts a detailed analysis to infer the changes in the athlete's movements before and after. In some embodiments, after combining the backtracked image frames with the three-dimensional trajectory information, the system can not only obtain the two-dimensional image frames of the athlete, but also calculate the accurate position, joint angles, and postures of the athlete in the three-dimensional space through the information of these frames. In this way, the system can accurately reproduce the continuous movement process of the athlete on the time axis.

[0061] Through these steps of the above-mentioned recognition module 20, the system can identify risk actions in real time, and by backtracking and analyzing the past movement trajectories and postures, provide optimized training feedback and action correction suggestions for the athlete.

[0062] In the analysis module 30, pose analysis is performed on the set of risk image frames to extract kinematic features, including the calculation of joint angle changes, the movement trajectories of limbs, and movement speeds.

[0063] Specifically, the calculation of joint angles is obtained from the three-dimensional joint position data of the athlete, that is, from the three-dimensional trajectory data obtained by the acquisition module 10, including the three-dimensional coordinate positions of each joint of the athlete, such as the joint positions of the shoulders, knees, elbows, ankles, etc.

[0064] The calculation of joint angles is usually based on the angle between two adjacent bones. For example, the angle of the knee joint is represented by the angle between the femur and the tibia. Assuming that the adjacent bones of joint A are P1 and P2 respectively, the joint angle θ can be calculated by the following formula:

[0065]

[0066] where and are the vectors between the two bones and the joint respectively.

[0067] Using these joint angle data, it can be analyzed whether the athlete has made non-standard actions such as excessive flexion or extension. Especially when risk actions (such as excessive knee flexion, excessive internal rotation of the knee during turning, etc.) occur, the changes in joint angles can provide obvious judgment bases.

[0068] Regarding the calculation of movement speed and acceleration, using the three-dimensional trajectory data, analyze the spatial movement path of the athlete during training. For example, by calculating the starting position, moving direction, and final position of the athlete and detecting the time interval of backtracking, the movement speed and acceleration of the athlete can be calculated. The speed is inferred through the time change of the joint position, so as to evaluate whether the athlete's movement is too intense or sudden and whether it will cause injury.

[0069] In addition, in some embodiments, the movement change trajectory of the athlete can also be analyzed, and the evolution of the movement can be deeply analyzed in combination with time factors. For example, certain movements may gradually become irregular within a certain time span. Through time series analysis, the turning points of the movements can be identified to help better predict and judge risky movements.

[0070] In summary, the analysis module 30 calculates the angles between joints using the known two-dimensional coordinates and three-dimensional reconstruction data, and calculates the kinematic characteristics such as the speed and acceleration of the movement trajectory using the positions of each joint of the athlete in the three-dimensional space. Through these analyses, more accurate training feedback can be provided for the athlete to help reduce the movement risk and improve the movement performance.

[0071] Finally, there is the feedback module 40. The feedback module 40 is connected to the analysis module 30 and performs iterative learning of the deep learning model based on the detected high-risk movement data. By collecting the training data of the athlete and analyzing the movement patterns, the deep learning model continuously optimizes the accuracy and response speed of risk movement recognition.

[0072] The feedback module 40 first receives the extracted kinematic characteristics from the analysis module 30, such as joint angles, movement speed, acceleration, etc., and combines the preset kinematic criteria, such as the threshold of certain joint angles, the upper limit of movement speed, etc., to perform a risk assessment on the current movement. The assessment result will represent the risk level of the movement, usually divided into low risk, medium risk and high risk. For example, excessive knee flexion is marked as a high-risk movement, while slight flexion is marked as a low-risk movement.

[0073] The learning model will generate a prediction of the athlete's current movement based on the previous training data, including joint angles, movement trajectories, etc. Compare the assessment result with the initial prediction result of the deep learning model. If the assessment result is consistent with the prediction result, that is, the model correctly identifies a risky movement or a normal movement, then the existing state of the model does not need to be greatly adjusted. If the assessment result is inconsistent with the prediction result, that is, the model misjudges a certain risky movement or fails to identify a potential risk, then iterative learning and update of the model are required.

[0074] Regarding the iteration and update of the learning model, in some embodiments of the present invention, the backpropagation algorithm is used to pass the assessment result as a "new label" to the deep learning model. In iterative learning, the label is the true category or risk level of the movement, and the input is the features extracted from the image frame, such as joint angles, movement speed and acceleration, etc. For those movements that are mispredicted, the system will provide the correct label (such as high risk) together with the output of the model (such as low risk) to the training algorithm.

[0075] During the backpropagation process, the gradient descent method is used to adjust the weights and biases of the model. The aim is to reduce the prediction error so that the model can better fit the training data. Updating the parameters of the model is achieved through the loss function. During the training process, an online learning approach can be adopted, enabling the model to be fine-tuned and optimized immediately every time new data is received. This is to ensure that in each training session of the athlete, the deep learning model can respond and update its recognition ability as soon as possible.

[0076] In summary, the entire system collects data through multi-view video capture and 3D reconstruction technology, without the need to set sensors on the athletes, and does not interfere with their training process at all. By analyzing the characteristics such as the postures and joint angles of the athletes in real time, the system can provide customized action suggestions for each athlete, effectively improving the training quality. In addition, the system can perform iterative updates of the deep learning model based on the evaluation results of the feedback module 40. As the training process progresses, the recognition accuracy and risk judgment ability of the system will continuously improve, ensuring more accurate recognition of risky actions.

[0077] The above is only a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of patent protection of the present invention is subject to the claims. Any equivalent structural changes made by using the description and drawings of the present invention should be equally included in the protection scope of the present invention.

Claims

1. An intelligent management system for sports basketball training risk data, characterized in that, Including: A collection module, based on visual capture technology, simultaneously captures two-dimensional videos of athletes from different shooting angles in a training field, constructs three-dimensional spatial coordinates based on the training field, performs three-dimensional reconstruction on the captured two-dimensional videos, and generates three-dimensional trajectories of the athletes; An identification module, based on a trained deep learning model, performs pose analysis on the three-dimensional trajectories of the athletes, identifies risk actions pre-labeled by the learning model therefrom, and uses the two-dimensional image frames captured from different shooting angles when such risk actions occur as key frames, and backward traces a plurality of set numbers of image frames forward in time series based on the key frames. The backward-traced image frames and the key frames together constitute a risk image frame set; An analysis module, used to perform pose analysis on the risk image frame set and the three-dimensional trajectory at the corresponding moment to extract kinematic features, where the kinematic features include joint angles of the athletes, movement trajectories of limbs, movement speeds, and accelerations; A feedback module, receives the extracted kinematic features, compares them with the corresponding preset standard ranges, respectively obtains feature deviations between the joint angle detection values, movement speed detection values, acceleration detections and the corresponding preset standard ranges, and performs weighted processing on each of the feature deviations according to preset weights to evaluate the risk level of the current risk action of the athlete.

2. The intelligent management system for sports basketball training risk data according to claim 1, characterized in that The collection module performs real-time video capture of the athlete from different angles in the training field, extracts features from the captured multiple two-dimensional images, then constructs a three-dimensional coordinate space of the training field, and calculates the depth information of the athlete in the training field through feature matching and triangulation of each two-dimensional image; Based on the depth information calculated from different perspectives, integrate the three-dimensional coordinates of all matched feature points to construct a three-dimensional point cloud in the training field; Based on the three-dimensional point cloud data reconstructed in three dimensions, track the position change of the athlete in three-dimensional space; among them, the three-dimensional coordinates corresponding to each frame of two-dimensional image can be used to infer the movement path of the athlete by comparing the point cloud data of adjacent frames, so as to obtain the three-dimensional trajectory of the athlete.

3. An intelligent management system for sports basketball training risk data according to claim 1, characterized in that The identification module identifies the defined risk actions from the three-dimensional trajectory through the trained deep learning model, where the marked risk actions are defined as actions that pose a threat to the safety of the athlete, and the deep learning model is trained to identify the marked risk actions from the three-dimensional trajectory through a temporal convolutional network.

4. The intelligent management system for sports basketball training risk data according to claim 3, characterized in that, When the identification module processes the three-dimensional data of the athlete in real time, the deep learning model extracts kinematic features from the three-dimensional trajectory of the athlete, and performs detection and judgment of risk actions after comparing with a preset feature recognition threshold.

5. The intelligent management system for sports basketball training risk data according to claim 4, characterized in that, When the identification module detects a risk action, define the multi-perspective two-dimensional image frames captured at that moment as the key frames; among them, the key frames are two-dimensional images captured based on multiple perspectives, and each of the two-dimensional images provides visual information of different directions of the athlete; Backward trace a set number of image frames forward in time based on the key frames; among them, the backward-traced image frames are two-dimensional images captured from different perspectives at the same moment; Construct the set of risk image frames from the key frames and a set number of image frames traced back in time series.

6. The intelligent management system for sports basketball training risk data according to claim 1, characterized in that, The recognition module extracts the joint positions of the athlete in each frame from the set of risk image frames, and calculates the joint angles based on the extracted joint positions; Analyze the dynamic angle changes of the joints during the occurrence of a risk action based on the calculated joint angles using time series.

7. The intelligent management system for sports basketball training risk data according to claim 6, characterized in that, The recognition module calculates the movement speed and acceleration of the athlete based on the three-dimensional trajectory data at the corresponding moments of the set of risk image frames; specifically, calculate the movement speed and acceleration based on the time variation of the positions of each joint of the athlete in three-dimensional space.

8. The intelligent management system for sports basketball training risk data according to claim 1, wherein The feedback module receives the kinematic features extracted from the analysis module, including the joint angles, movement speed, and acceleration at the moment when the athlete performs a risk action recognized by the deep learning model, and combines preset kinematic criteria, including thresholds for different joint angles and the upper limit speed of movement, to calculate the risk score for the current action. If the risk score is greater than the upper limit value of the set range, the risk action is evaluated as a high risk; if the risk score is less than the lower limit value of the preset range, the risk action is evaluated as a low risk.

9. The intelligent management system for sports basketball training risk data according to claim 1, wherein, The deep learning model generates a prediction result for the current risk action of the athlete based on previous training data, compares the received risk score with the prediction result. If there is a deviation beyond the preset range, adjust the weights of the network using the backpropagation algorithm according to the kinematic features extracted by evaluating the analysis module, and update the parameters of the model.

Images