Basketball sports training method based on virtual reality

Abstract

The invention provides a basketball sports training method based on virtual reality, and belongs to the technical field of intelligent analysis, and the method comprises the steps: obtaining personal basic data of an athlete and capture data of typical actions, and fusing the data into a virtual character; a virtual reality training scene is constructed based on an actual scene of basketball sports and competition requirements, virtual opponents and teammates are added in the virtual reality training scene, and different competitive levels and tactical styles are given through an artificial intelligence algorithm. A personalized training scheme is formulated by combining the action defects positioned by the excavated wavelet features of the historical actions of the athletes and typical error trajectory simulation training of the virtual reality training scene; the action specification condition is automatically marked to guide and intervene the training process of the athlete, the training effect of the athlete is comprehensively evaluated according to a preset evaluation index system after training is finished, and a personalized training scheme is optimized. The training efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of virtual reality, and in particular to a basketball sports training method based on virtual reality. Background Art

[0002] In the digital transformation of sports training, virtual training technology provides an innovative path for sports teaching and training. However, existing solutions still have significant shortcomings in terms of data depth, scenario dynamics, solution accuracy, and closed-loop feedback:

[0003] Existing systems focus primarily on the surface geometric features of motion trajectories (e.g., joint positions and motion paths), failing to fully explore the temporal and frequency details of historical movements and ignoring the dynamic relationship between physiological data and movements. This single-dimensional data collection results in movement defect detection remaining at the level of posture deviation, failing to address core issues such as abnormal force timing and low coordination efficiency. This makes it difficult to support the precise design of training programs. In virtual training scenarios, the competitive level and tactical style of virtual opponents / teammates are mostly statically preset (e.g., fixed defensive intensity and single tactical scripts), failing to utilize the dynamic capabilities of real-world game trajectory libraries (e.g., real-world positioning and defensive strategy data for NBA / CBA players). Furthermore, scenario construction only replicates physical space (e.g., basketball court dimensions) without simulating actual combat elements. This results in significant differences in behavioral logic and decision-making complexity between training scenarios and real-world games, making it difficult to transfer training results to actual combat. Furthermore, the development of personalized training programs still relies on qualitative analysis dominated by coaching experience. Any of these factors contributes to low training efficiency.

[0004] Therefore, the present invention proposes a basketball sports training method based on virtual reality. Summary of the Invention

[0005] The present invention provides a basketball sports training method based on virtual reality to solve the above-mentioned technical problems.

[0006] The present invention provides a basketball sports training method based on virtual reality, comprising:

[0007] Step 1: Obtain the athlete's basic personal data and capture data of typical movements, and integrate them into the virtual character;

[0008] Step 2: Build a virtual reality training scenario based on the actual scenarios and competition requirements of basketball, add virtual opponents and teammates to the virtual reality training scenario, and use artificial intelligence algorithms to assign different competitive levels and tactical styles. Based on the collected real-world trajectory library, the movement trajectories are randomly assigned to the corresponding virtual opponents and teammates.

[0009] Step 3: Develop a personalized training plan based on the athlete's training goals, personal basic data, and current training level, combined with the athlete's historical movement wavelet feature location and typical error trajectory simulation training in the virtual reality training scenario;

[0010] Step 4: During the training process, the athlete's motion data and physiological data are captured in real time, and the movement standards are automatically marked to guide and intervene in the athlete's training process. After the training, the athlete's training effect is comprehensively evaluated according to the preset evaluation index system to optimize the personalized training plan.

[0011] Preferably, after obtaining the capture data of the athlete, the method further includes:

[0012] dividing the action video of the athlete into frames based on a number of repetitions of each typical action;

[0013] According to the action type of each typical action, match the salient features from the type-salient comparison table;

[0014] Arrange the limb contour trajectory based on the limb key points in each repeated video frame corresponding to the typical action in sequence according to the first and last coordinates to obtain the initial complete trajectory;

[0015] Extracting the whole-body motion profile of each video frame for the corresponding number of repetitions of the typical action, analyzing the whole-body motion profile according to the salient characteristics, locking the initial key trigger frame and the end key trigger frame for the corresponding typical action, and obtaining the key trajectory segment;

[0016] Determine the splicing state of the splicing position of adjacent video frames. When the splicing position is in a key trajectory segment, extract local geometric features of corresponding limb key points in n0 adjacent video frames closest to the splicing position, and place the local geometric features in chronological order. Combined with the splicing state, estimate the first geometric feature of the corresponding splicing position. Combined with the salient characteristics, perform an interpolation processing method for the corresponding splicing position in the key trajectory segment to optimize and update the connection between the two splicing points of the corresponding splicing position.

[0017] When the splicing position is not in a key trajectory segment, when the splicing state is point overlap, the intersection point of the corresponding splicing position is kept unchanged and redundant non-overlapping line segments are removed; when the splicing state is point non-intersection, the two points of the splicing position of adjacent video frames are connected by a straight line;

[0018] Obtaining a required complete trajectory corresponding to a limb key point based on an updated result of the splicing position in the initial complete trajectory;

[0019] Perform coordinate mapping on all required complete trajectories of each limb key point under the same typical action to determine the commonality of the corresponding limb key points;

[0020] At the same time, a dynamic analysis is performed on each limb movement under the same typical action, and the motion commonality of each limb key point is determined, wherein the point commonality is related to the amplitude characteristic distribution of the corresponding limb key point, and the motion commonality is related to the motion speed distribution of the corresponding limb key point;

[0021] Inputting the point commonality and movement commonality into a balance analysis model to obtain a balance coefficient corresponding to a key point of the limb;

[0022] If the balance coefficient is greater than the preset coefficient, the captured data under the corresponding typical action is kept unchanged;

[0023] Otherwise, the captured data under the corresponding typical action is regarded as the first data, and the first data is subjected to wavelet analysis to determine the filtering threshold, and the noise data of the captured data is filtered out to obtain the second data;

[0024] The retained data and the second data are converted into an action standard to obtain an action sequence, and the action sequence is merged with the basic sequence obtained by converting the personal basic data into a standard sequence, and is integrated into the virtual character.

[0025] Preferably, performing wavelet analysis on the first data to determine the filtering threshold comprises:

[0026] Based on the CNN wavelet-based learning module, basketball action clips are pre-classified, the optimal wavelet decomposition scale L is automatically matched, and the key action dimensions are enhanced through the attention mechanism of the decomposed wavelet coefficients;

[0027] Extract the action periodicity of the key action dimension through Fourier transform and determine the action period weight factor Wherein, E0 represents the main frequency energy proportion of the key action dimension; Emax represents the energy proportion of the ideal stable period;

[0028] Obtain the angle sequence of the limb key points of each repetitive action video under each typical action in turn, determine the average change rate of each limb key point, and obtain the standard deviation σi of each limb key point and the mean absolute value covariance between the limb key points age Get the action complexity Cf=∝1· Where ∝1 and ∝2 are weights respectively, and ∝1+∝2=1;

[0029] Calculate the filtering threshold Yy according to the action cycle weight factor ω1 and the action complexity Cf;

[0030]

[0031] Wherein, δ1 represents the deviation between the median and mean of the wavelet coefficients; N represents the data length corresponding to the first data.

[0032] Preferably, the balance analysis model includes: M1 parallel feature encoding subnetworks, at least two gated hybrid subnetworks, and a confidence-driven integration network;

[0033] The feature encoding subnetwork includes K layers of encoding units, and each layer of encoding units is connected to a GELU activation function. The input of the first L1 layer encoding unit is the common point features of the limb key points, including: spatial position distribution and amplitude discreteness. The input of the second K-L1 layer encoding unit is the common motion features of the limb movement, including speed temporal changes and distribution entropy. The number of neurons in the first L1 layer encoding unit is configured as 64, and the number of neurons in the second K-L1 layer encoding unit is configured as 128.

[0034] The gated hybrid sub-network is constructed based on the input layer, the attention hidden layer, and the gated output layer. The outputs of the M1 feature encoding sub-networks are dynamically weighted through the attention mechanism. The attention hidden layer adjusts the weight coefficient in real time according to the correlation between features.

[0035] The confidence-driven integrated network is used to calculate the variance of the output of each gated hybrid sub-network to estimate the prediction confidence, and dynamically weighted sum the output of the gated hybrid sub-network according to the confidence to generate the balance coefficient of the corresponding limb key point.

[0036] Preferably, step 3 includes:

[0037] A twin model of movement defects and error trajectories is constructed, and the wavelet time-frequency characteristics of the athlete's historical movements are correlated with the spatial tactical characteristics of typical error trajectories in virtual reality training scenarios in time and space dimensions to generate a causal relationship map between movement defects and error inducements;

[0038] Based on the training goals, personal basic data and current training level, and in combination with the cause-effect relationship map, an initial training plan is generated.

[0039] Preferably, after generating the initial training plan, the method further includes:

[0040] Extract the abnormal frequency components of the angles of the limb key points from the wavelet features of historical movements, and build a set of mechanical constraint rules for movement correction based on the human joint kinematic thresholds.

[0041] The initial training plan is modified based on the mechanical constraint rule set to obtain a personalized training plan.

[0042] Preferably, step 4 includes:

[0043] Extracting the motion specification features of the motion data based on a spatiotemporal graph convolutional network, and analyzing the dynamic evolution pattern of the physiological data based on a variational autoencoder;

[0044] Based on a graph neural network, the causal link between movement standard defects and abnormal physiological responses is learned, and the movement standard characteristics and dynamic evolution patterns are analyzed to construct the athlete's movement-physiology spatiotemporal causal graph;

[0045] Based on the pre-trained time series prediction model, the action-physiology spatiotemporal causal graph is predicted and analyzed, and the pre-intervention strategy is output in combination with the automatically labeled action specifications;

[0046] The causal effect value of the pre-intervention strategy and the training effect is quantified to determine the significant status, and the personalized training program is optimized.

[0047] Preferably, the automatically marked action specifications include:

[0048] A 3D pose estimation network is used to analyze the motion data, output normative scores of limb key points, and locate key propagation nodes of motion defects through an attention mechanism;

[0049] determining a force efficiency coefficient of the athlete based on the physiological data;

[0050] The standard score, key propagation nodes, force efficiency coefficient and fatigue correlation are automatically labeled to obtain the movement standardization situation.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] Through the full-link technology of high-precision motion capture - virtual reality scene construction - intelligent defect analysis - dynamic intervention optimization, the personalization and precision of basketball training are achieved. The real game experience is reproduced through virtual characters and scenes, and AI and signal processing technology are used to locate movement defects and dynamically optimize training plans to improve training efficiency.

[0053] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0054] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0056] Figure 1 This is a flowchart of a basketball training method based on virtual reality in an embodiment of the present invention;

[0057] Figure 2 This is a structural diagram of the arrangement of the first and last coordinates in sequence in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0059] The present invention provides a basketball sports training method based on virtual reality, such as Figure 1 As shown, including:

[0060] Step 1: Obtain the athlete's basic personal data and capture data of typical movements, and integrate them into the virtual character;

[0061] Step 2: Build a virtual reality training scenario based on the actual scenarios and competition requirements of basketball, add virtual opponents and teammates to the virtual reality training scenario, and use artificial intelligence algorithms to assign different competitive levels and tactical styles. Based on the collected real-world trajectory library, the movement trajectories are randomly assigned to the corresponding virtual opponents and teammates.

[0062] Step 3: Develop a personalized training plan based on the athlete's training goals, personal basic data, and current training level, combined with the athlete's historical movement wavelet feature location and typical error trajectory simulation training in the virtual reality training scenario;

[0063] Step 4: During the training process, the athlete's motion data and physiological data are captured in real time, and the movement standards are automatically marked to guide and intervene in the athlete's training process. After the training, the athlete's training effect is comprehensively evaluated according to the preset evaluation index system to optimize the personalized training plan.

[0064] In this embodiment, the personal basic data is basic information reflecting the athlete's physiological characteristics and sports background, such as height (1.98m), weight (90kg), wingspan (2.05m), dominant hand (right hand), years of sports experience (5 years), and injury history (after left knee ligament repair surgery). Specifically, physiological parameters are collected through physical measurement equipment (height and weight meter, 3D scanner), and background information is recorded in combination with the athlete questionnaire.

[0065] The captured data of typical actions are high-precision kinematic data when athletes complete core basketball actions (such as shooting, dribbling, and passing). For example, the three-dimensional coordinate sequence of the wrist joint when shooting a three-pointer (sampling frequency 120Hz) and the angular velocity change curve of the ankle when dribbling and changing direction. Specifically, a 16-lens optical motion capture system (such as Vicon) is used to synchronize inertial sensors (IMU) to record limb movement details.

[0066] In this embodiment, the virtual character is a digital avatar constructed based on the athlete's data and is used for training simulation in a virtual reality scene. For example, a 3D model of the athlete's body shape and movement habits is reproduced in VR, and his shooting gestures and movement steps can be accurately reproduced. Specifically, 3D modeling is performed through Blender, combined with MotionBuilder to bind action bones, and drive the movement of the virtual character.

[0067] In this embodiment, the virtual reality training scene is a digital space that simulates a real basketball game environment, with physical properties and interactive functions. For example, it replicates an NBA standard basketball court (28m×15m) on a 1:1 scale, including details such as lighting, audience sound effects, and a floor friction coefficient (0.8). Specifically, the scene is built using the Unity engine, and the NVIDIA PhysX physics engine is integrated to simulate real collisions and forces.

[0068] Virtual opponents and teammates are digital characters controlled by AI, used to simulate the confrontation and cooperation in the game. For example, virtual teammates can execute pick-and-roll tactics, and virtual opponents can adjust defensive strategies according to the athlete's habits (such as predicting their right-hand breakthrough). Specifically: the AI ​​decision-making model is trained based on the reinforcement learning algorithm (PPO), and the character action logic is controlled through the behavior tree.

[0069] Artificial intelligence algorithms are computing models used to give virtual characters the ability to make autonomous decisions. For example, the Deep Q Network (DQN) allows virtual opponents to learn when to choose defensive backup, and clustering algorithms are used to divide teammates' cooperation preferences. Specifically, the model is trained using the PyTorch framework and deployed on edge computing devices (such as NVIDIA Jetson) to achieve low-latency response.

[0070] Competitive level and tactical style refer to the ability level and tactical execution characteristics of the virtual character. For example, the competitive level is divided into 5 levels (amateur → professional), corresponding to different defensive reaction speeds (amateur 0.8 seconds vs. professional 0.3 seconds); tactical styles include run and gun (mainly fast break) and half-court zone defense (zone defense).

[0071] In this embodiment, the real trajectory library is a database that records the movement paths of players in real games. For example, it includes the average running trajectory of players per game (such as Curry's average moving distance of 9.2 km per game) and the distribution of the screener's cutting route after the pick-and-roll. Specifically, professional game data is extracted through video analysis software (such as SportVU) and stored as a time-space coordinate sequence (x, y, t).

[0072] Training goals are the specific results that athletes hope to achieve through training, such as increasing three-point shooting percentage from 32% to 40% within 3 months or reducing the shaking of the center of gravity when sliding on defense.

[0073] The current training level refers to the athlete's current skill mastery and physical condition. For example, after testing, the current three-point shooting percentage is 28%, and the defensive lateral speed is 2.8m / s (lower than the average of 3.2m / s for the same position).

[0074] The wavelet features of historical movements are the frequency and energy features extracted after wavelet transform of historical movement data, which are used to analyze movement stability. For example, the wavelet decomposition results of shooting movements show that the elbow joint has abnormal high-frequency jitter of 15Hz during the force-generating phase (normal is 5-8Hz).

[0075] In this embodiment, movement defects are parts of the athlete's movements that do not conform to biomechanical efficiency or standards. For example, when shooting, the elbow valgus angle exceeds 30° (the standard should be ≤15°), resulting in dispersed force; when dribbling, the waist does not sink and the center of gravity is too high.

[0076] In this embodiment, the typical error trajectory is a common incorrect movement path simulated in the virtual reality scene, such as a lateral movement trajectory that fails to make timely defensive moves when defending (deviation of 1.5m from the optimal path), and a passing route that ignores the defensive pressure of teammates when passing.

[0077] In this embodiment, the personalized training program is a training plan customized according to the characteristics of the athlete, including elements such as content, intensity, and frequency. For example, 10 sets × 10 elbow correction shots per day (combined with resistance band training), and 20 minutes of defensive backup simulation in a VR scene every day (the difficulty increases with performance).

[0078] Real-time capture is a technical process of synchronously collecting data during training. For example, during training, 120 frames of motion data (joint angles, positions) and 200 times of physiological data (heart rate, electromyography) are collected per second.

[0079] Motion data and physiological data are two core types of data that reflect the state of movement, for example, motion data (wrist angle 120° when shooting, ball speed 8m / s); physiological data (heart rate 160 beats / minute, biceps electromyography amplitude 2.5mV).

[0080] Automatically marking the standardization of movements is an automatic evaluation result of the compliance and efficiency of the movements. For example, it is marked that the shooting elbow angle exceeds the standard (35°), the force efficiency coefficient is 0.7 (out of 1.0), and the fatigue correlation is 0.6 (indicating that the movement may be deformed due to fatigue).

[0081] In this embodiment, guidance and intervention are dynamic adjustments to the training process based on real-time data. For example, real-time voice prompts are given to adduct the elbows 5°, and training is automatically paused when the heart rate exceeds 180 beats / minute, with breathing adjustment guidance popping up.

[0082] The preset evaluation index system is a multi-dimensional standard for measuring training effectiveness, such as movement standardization (weight 40%), hit rate (30%), physiological recovery rate (20%), and tactical execution accuracy (10%).

[0083] Comprehensive evaluation and program optimization is the analysis of the effect and iteration of the program after training. For example, the evaluation showed that the elbow correction training was effective (standardization increased by 25%), but the endurance was insufficient (the movement was deformed after the 8th set). The optimization program increased the halftime break time (from 60 seconds to 90 seconds).

[0084] The beneficial effects of the above technical solution are: through the full-link technology of high-precision motion capture-virtual reality scene construction-intelligent defect analysis-dynamic intervention optimization, the personalization and precision of basketball training can be achieved. It not only reproduces the real game experience through virtual characters and scenes, but also uses AI and signal processing technology to locate movement defects and dynamically optimize training plans to improve training efficiency.

[0085] The present invention provides a basketball sports training method based on virtual reality, which, after acquiring the capture data of the athlete, further comprises:

[0086] dividing the action video of the athlete into frames based on a number of repetitions of each typical action;

[0087] According to the action type of each typical action, match the salient features from the type-salient comparison table;

[0088] Arrange the limb contour trajectory based on the limb key points in each repeated video frame corresponding to the typical action in sequence according to the first and last coordinates to obtain the initial complete trajectory;

[0089] Extracting the whole-body motion profile of each video frame for the corresponding number of repetitions of the typical action, analyzing the whole-body motion profile according to the salient characteristics, locking the initial key trigger frame and the end key trigger frame for the corresponding typical action, and obtaining the key trajectory segment;

[0090] Determine the splicing state of the splicing position of adjacent video frames. When the splicing position is in a key trajectory segment, extract local geometric features of corresponding limb key points in n0 adjacent video frames closest to the splicing position, and place the local geometric features in chronological order. Combined with the splicing state, estimate the first geometric feature of the corresponding splicing position. Combined with the salient characteristics, perform an interpolation processing method for the corresponding splicing position in the key trajectory segment to optimize and update the connection between the two splicing points of the corresponding splicing position.

[0091] When the splicing position is not in a key trajectory segment, when the splicing state is point overlap, the intersection point of the corresponding splicing position is kept unchanged and redundant non-overlapping line segments are removed; when the splicing state is point non-intersection, the two points of the splicing position of adjacent video frames are connected by a straight line;

[0092] Obtaining a required complete trajectory corresponding to a limb key point based on an updated result of the splicing position in the initial complete trajectory;

[0093] Perform coordinate mapping on all required complete trajectories of each limb key point under the same typical action to determine the commonality of the corresponding limb key points;

[0094] At the same time, a dynamic analysis is performed on each limb movement under the same typical action, and the motion commonality of each limb key point is determined, wherein the point commonality is related to the amplitude characteristic distribution of the corresponding limb key point, and the motion commonality is related to the motion speed distribution of the corresponding limb key point;

[0095] Inputting the point commonality and movement commonality into a balance analysis model to obtain a balance coefficient corresponding to a key point of the limb;

[0096] If the balance coefficient is greater than the preset coefficient, the captured data under the corresponding typical action is kept unchanged;

[0097] Otherwise, the captured data under the corresponding typical action is regarded as the first data, and the first data is subjected to wavelet analysis to determine the filtering threshold, and the noise data of the captured data is filtered out to obtain the second data;

[0098] The retained data and the second data are converted into an action standard to obtain an action sequence, and the action sequence is merged with the basic sequence obtained by converting the personal basic data into a standard sequence, and is integrated into the virtual character.

[0099] In this embodiment, the number of repetitions is the number of times the athlete is required to repeat a typical action (excluding accidental mistakes) in order to obtain stable action characteristics, and is generally set to 10.

[0100] In this embodiment, the action video is a dynamic image that records the athlete's action process, which must meet the requirements of high frame rate and clear image quality. For example, a 4K camera (frame rate 60fps) is used to shoot from the front and side dual perspectives to ensure that the limbs are not obstructed. Specifically, a Sony PXW-Z190 camera is used, fixed on a tripod, and dual cameras are triggered synchronously for shooting.

[0101] In this embodiment, frame division is to decompose the continuous video into single static images (frames) to facilitate frame-by-frame analysis of action details. For example, a 10-second shooting video (60fps) can be divided into 600 frames, with each frame interval of 0.0167 seconds.

[0102] In this embodiment, the action type is a subdivision of typical actions, reflecting the technical attributes of the action. For example, shooting can be subdivided into three-point shots and bank shots; dribbling can be subdivided into front-body changes of direction and behind-the-back dribbling. The type-prominence comparison table is a predefined mapping table of action types and key technical characteristics, jointly developed by coaches and sports biomechanics experts, as shown in Table 1:

[0103] Table 1 Type-prominence comparison table

[0104] Action Type Outstanding Features Three-point shooting Wrist extension trajectory curvature, elbow lift angle Front-body change of direction dribble Ankle turning speed and weight shift range

[0105] In this embodiment, the prominent characteristics are the core technical indicators that determine the quality of the action and directly affect the action effect (such as hit rate and ball control stability). For example, the curvature of the wrist extension trajectory in a three-point shot - too large a curvature will easily cause the ball to deviate to the left, and too small a curvature will easily cause the ball to deviate to the right.

[0106] In this embodiment, limb key points are joints that provide critical support or force in the human motion chain and are the core markers for trajectory analysis. For example, there are 15 key points for the upper limbs (shoulder, elbow, wrist, and fingers), lower limbs (hip, knee, and ankle), and trunk (waist). A limb contour trajectory is the motion path formed by the key points of a single limb (such as an arm) in consecutive frames, reflecting the limb's motion trajectory. For example, when shooting a basketball, the key point trajectory (arc-shaped) of the elbow from ball raising to release is generated by concatenating the key point coordinates of consecutive frames to generate the original trajectory segment.

[0107] Connect head to tail. For example, the trajectory of limb key point A in video frame 1 is B1, and the trajectory in video frame 2 is B2. At this time, video frame 1 is before video frame 2, and video frame 1 is adjacent to video frame 2. At this time, point B11 in trajectory B1 is the first point, point B12 is the last point, and point B21 in trajectory B2 is the first point, and point B22 is the last point. At this time, the first and last coordinates are arranged in sequence, that is, trajectory B1 is arranged before trajectory B2, and each point has its corresponding coordinates, which can be arranged according to the corresponding coordinates, such as Figure 2 As shown:

[0108] When point B12 does not intersect point B21, refer to result R1;

[0109] When only point B12 intersects point B21, refer to result R2;

[0110] When point B12 does not intersect with point B21, but trajectory B1 intersects with trajectory B2, see result R3.

[0111] The initial complete trajectory is a continuous path formed by splicing the limb contour trajectories of single frames in chronological order (frame sequence), without any optimization processing. The full-body motion contour is the edge outline of the athlete's entire body in a single frame, reflecting the overall posture and spatial positioning of the limbs. For example, when shooting, the three-frame full-body contour of bending the knees → raising the ball → shooting can be used to intuitively observe the degree of body stretch.

[0112] The initial key trigger frame is a landmark frame in the action startup phase and the starting point of the key trajectory segment. For example, the initial key trigger frame of a three-point shot is the frame where the knee is bent to the maximum angle (knee angle ≤ 90°). The end key trigger frame is a landmark frame in the action completion phase and the end point of the key trajectory segment. For example, the end key trigger frame of a three-point shot is the frame where the wrist is fully extended and the ball leaves the hand. The key trajectory segment is the trajectory segment between the initial key trigger frame and the end key trigger frame, which includes the core force process of the action.

[0113] In this embodiment, the splicing position is the connection point of the trajectory segments of two adjacent frames (the corresponding area between the end point of the previous frame and the starting point of the next frame), for example, the connection position of the wrist end point (x50, y50) of the 50th frame and the wrist starting point (x51, y51) of the 51st frame.

[0114] The splicing state is the geometric relationship of the splicing positions, which is divided into point overlap and point non-intersection.

[0115] The n0 adjacent video frames are the number of frames before and after the splicing position (empirical value n0 = 3 to 5), which are used to extract local features to optimize the splicing. For example, when the splicing position is at the 50-51 frame, the 48th, 49th, 50th, 51st, and 52nd frames (n0 = 5) are taken for analysis.

[0116] Local geometric features are the spatial relationship features of key points near the splicing position, reflecting the local movement rules of the limbs. For example, the local features of the wrist splicing point include the angle formed by the elbow-wrist-finger three points (such as 120°) and the rate of change of the distance between the wrist and the waist.

[0117] The first geometric feature is the ideal geometric feature of the predicted splicing point (consistent with motion continuity) based on the local geometric features and the splicing state. For example, if the wrist points of adjacent frames do not intersect (with an error of 5 pixels), the splicing point is predicted to be (x_pred, y_pred) based on the motion trend of the previous three frames, ensuring a smooth transition of the trajectory.

[0118] The interpolation processing method is an optimization algorithm for splicing points selected based on prominent features, which is used to correct the discontinuity of the trajectory. For example, Bezier curve interpolation (preserving curvature characteristics) is used for the key wrist trajectory segment of a three-point shot; linear interpolation is used for non-critical segments.

[0119] In this embodiment, point overlap means that the coordinates of the splicing points of adjacent frames are close (error ≤ 2 pixels), which is considered to be effective overlap. For example, in the preparation stage of shooting, the elbow points of adjacent frames overlap (error 1 pixel), and the intersection points are directly retained. Point non-intersection means that the coordinate deviation of the splicing points of adjacent frames is large (error > 2 pixels), and the trajectory is broken. For example, when sliding on defense, due to the fast movement, the ankle points of adjacent frames do not intersect (error 6 pixels).

[0120] In this embodiment, straight line connection is a simplified splicing method for non-critical trajectory segments, connecting non-intersecting points with the shortest path to reduce the amount of calculation.

[0121] The required complete trajectory is a high-precision limb trajectory that has been spliced ​​and optimized, eliminating inter-frame jitter and breakage while retaining key features. For example, the optimized wrist trajectory for three-point shooting is a smooth arc without burrs, accurately reflecting the force process.

[0122] In this embodiment, coordinate mapping is the conversion of the trajectories of repeated actions into a unified coordinate system (such as a three-dimensional coordinate system with the basket as the origin) to facilitate comparative analysis, and the point commonality is the position distribution pattern of the same key point in the repeated actions (reflecting the stability of the action). For example, the standard deviation of the x-coordinate of the wrist release point of 10 three-point shots in the unified coordinate system is 0.05m (the smaller the better).

[0123] The amplitude characteristic distribution is a quantitative indicator of the commonality of point positions, describing the degree of dispersion of key point positions (such as standard deviation, range, and distribution entropy). For example, the wrist amplitude characteristic distribution has a standard deviation of 0.05m in the x-direction and a range of 0.12m in the y-direction.

[0124] Dynamic analysis calculates physical quantities such as force and torque based on kinematic data (position and velocity) to analyze the force patterns of the movement. Motion commonality is the distribution pattern of motion parameters (velocity and acceleration) at the same key point in multiple repeated movements. For example, the average ankle turning speed in 10 front-body change-of-direction dribbles is 120° / s, with a standard deviation of 15° / s. Speed ​​distribution is a quantitative indicator of motion commonality, describing the statistical characteristics of the speed of key points (such as average speed and the time when the speed peak occurs).

[0125] The balance analysis model is a machine learning model that integrates point commonalities (spatial stability) and movement commonalities (temporal fluency). It outputs a movement stability score, and the balance coefficient is the score output by the model (0 to 10 points). The higher the value, the more stable the movement and the more consistent it is with the laws of biomechanics. For example, a professional player's three-point shooting wrist balance coefficient is 8.5 points, while a novice may have a score of 4.2 points.

[0126] The preset coefficient is the threshold for determining whether the motion data needs to be filtered (set according to the level of exercise, 7 points for professional training and 5 points for general training).

[0127] The first data is low-quality data (containing noise or unstable movement) with a balance coefficient lower than a preset value, and wavelet analysis decomposes the frequency components of the data through wavelet transform, identifies and filters out noise (such as high-frequency noise caused by sensor jitter), and the filtering threshold is the retention threshold of the wavelet coefficient, which is dynamically adjusted according to the balance coefficient (the lower the coefficient, the stricter the threshold).

[0128] Noise data is a non-effective signal that interferes with motion analysis (such as video compression artifacts and sensor drift). The second data is high-quality data after wavelet filtering, which retains effective motion features, eliminates noise, and maintains high-quality original data with a balance coefficient that meets the standard and does not require filtering.

[0129] In this embodiment, the action standard conversion is to convert the data into a format that can be recognized by the virtual character (such as a skeletal animation key frame sequence). Specifically, the BVH (BiovisionHierarchy) format is used to record the joint rotation angle (unit: degree). The action sequence is a standardized continuous action frame that can directly drive the movement of the virtual character. For example, a three-point shooting action sequence contains 200 frames, and each frame records the rotation angle of 15 joints.

[0130] The basic sequence is a benchmark action framework generated based on personal basic data (such as limb proportion animation that conforms to height). Fusion is the combination of action sequence and basic sequence, so that the virtual character not only conforms to the athlete's body shape but also reproduces its action details. Specifically: a weighted fusion algorithm is used (action sequence weight 0.8, basic sequence weight 0.2). Integration into the virtual character is to bind the fused sequence to the virtual character's skeletal system to achieve action replication.

[0131] The beneficial effects of the above technical solution are: through the full process of feature-driven trajectory enhancement - intelligent optimization of splicing points - balance coefficient screening - wavelet filtering - data fusion, the accuracy and stability of athletes' motion data are significantly improved, so that virtual characters can reproduce the technical characteristics of athletes with high fidelity, providing a precise digital foundation for subsequent virtual reality training, and improving the targetedness and effectiveness of training.

[0132] The present invention provides a basketball sports training method based on virtual reality, which performs wavelet analysis on the first data to determine a filtering threshold, including:

[0133] Based on the CNN wavelet-based learning module, basketball action clips are pre-classified, the optimal wavelet decomposition scale L is automatically matched, and the key action dimensions are enhanced through the attention mechanism of the decomposed wavelet coefficients;

[0134] Extract the action periodicity of the key action dimension through Fourier transform and determine the action period weight factor Wherein, E0 represents the main frequency energy proportion of the key action dimension; Emax represents the energy proportion of the ideal stable period;

[0135] Obtain the angle sequence of the limb key points of each repetitive action video under each typical action in turn, determine the average change rate of each limb key point, and obtain the standard deviation σi of each limb key point and the mean absolute value covariance between the limb key points age Get action complexity Where ∝1 and ∝2 are weights respectively, and ∝1+∝2=1;

[0136] Calculate the filtering threshold Yy according to the action cycle weight factor ω1 and the action complexity Cf;

[0137]

[0138] Wherein, δ1 represents the deviation between the median and mean of the wavelet coefficients; N represents the data length corresponding to the first data.

[0139] In this embodiment, traditional wavelet filtering uses a global fixed threshold method, which cannot cope with the diversity of movements. For example, ball movement data has the characteristics of "detail-sensitive and noise-complex": Detail-sensitive: The 10-15Hz force tremor of the shooting wrist is the key to the hit rate and needs to be accurately preserved; Noise-complex: Sensor jitter (20-30Hz), video compression artifacts (low-frequency noise) and other interferences coexist. If a fixed threshold is used: for stable movements (such as professional players shooting free throws), the fixed threshold may mistakenly filter out force details (such as high-frequency wrist tremors), resulting in distortion of the virtual character's movements; for disordered movements (such as novice dribbling with a change of direction), the fixed threshold may retain too much noise (such as ankle shaking), interfering with subsequent analysis. Therefore, this solution dynamically adjusts the threshold from four aspects: period dimension (ω1), complexity dimension (Cf), data dimension (log2N), and wavelet dimension (δ1, L).

[0140] In this embodiment, Wherein, X() is the frequency spectrum; f0 is the main frequency, and the value range of Emax is 90% to 95%, which is obtained by counting the main frequency energy of stable actions of professional athletes (such as NBA players' free throws).

[0141] In this embodiment, δ1 is used to measure the degree of discreteness of coefficient distribution. The larger the wavelet decomposition scale L is, the higher the frequency resolution is, the larger the denominator 2L-1 is, and the lower the threshold is overall.

[0142] In this embodiment, the CNN wavelet basis learning module is an intelligent model that integrates convolutional neural networks (CNN) and wavelet analysis to learn the mapping relationship between motion features and wavelet decomposition parameters. Specifically, a ResNet-18 network is constructed, an angle sequence (such as 15 key points × 100 frames) is input, the optimal L is output, and PyTorch is used for training. The loss function is the mean square error between the predicted scale and the manually labeled optimal scale.

[0143] In this embodiment, the basketball action clip is a single typical action sequence (such as a complete three-point shot) intercepted from a training video, for example, an 80-frame video clip (1.3 seconds in length) from ball lifting to shooting.

[0144] In this embodiment, pre-classification is to identify the action type (such as shooting, dribbling) before wavelet decomposition and adapt different processing strategies. Specifically, a Softmax classification head is added to the CNN module to output the action type (shooting, dribbling, passing, etc.).

[0145] In this embodiment, the optimal wavelet decomposition scale L is the scale that makes the key action features clearest after wavelet decomposition (affecting the time-frequency resolution). For example, L=2 (high time resolution) is used for the shooting power stage, and L=5 (high frequency resolution) is used for defensive sliding.

[0146] In this embodiment, the wavelet coefficients are time-frequency coefficients obtained after wavelet decomposition, which include the low-frequency trend (such as trajectory) and high-frequency details (such as jitter) of the action. For example, in the wavelet coefficients of the shooting action, the low-frequency coefficients reflect the overall trajectory, and the high-frequency coefficients reflect the wrist jitter.

[0147] In this embodiment, the attention mechanism is to strengthen the wavelet coefficients related to key actions (such as the high-frequency coefficients of the shooting wrist) and suppress noise. For example, the attention layer identifies the 5-10Hz coefficient corresponding to the wrist force and increases its weight to 1.5 times. Specifically, a multi-head attention layer is added (such as Transformer's Multi-HeadAttention) to calculate the correlation between the coefficient and the key action dimension.

[0148] In this embodiment, the key movement dimensions are the core body parts or parameters that determine the quality of the movement (such as the wrist extension angle for shooting).

[0149] Movement periodicity is the temporal repetition of movement (the more stable the period, the more standardized the movement).

[0150] The angle sequence of limb key points is a sequence of angle changes of a key point (such as the elbow) in consecutive frames (such as [120°, 125°, 130°, ...]).

[0151] The average rate of change is the average value of the angle change of adjacent frames (reflecting the average speed of the action), the standard deviation is the degree of dispersion of the angle change rate of a single key point, and the mean of the absolute value of the covariance is the average of the absolute values ​​of the covariance of the angle change rates of multiple key points (such as elbows, wrists, and shoulders) (reflecting the collaborative complexity).

[0152] The beneficial effect of the above technical solution is: through the full-link design of intelligent scale selection-attention enhancement-periodic quantization-complexity fusion-dynamic threshold, adaptive wavelet filtering of basketball action data is realized: it not only accurately retains key force, coordination and other action characteristics, but also effectively filters out sensor noise and video interference, providing high-fidelity data support for subsequent action analysis and virtual character construction, making training plan optimization more accurate.

[0153] The present invention provides a basketball sports training method based on virtual reality, wherein the balance analysis model comprises: M1 parallel feature encoding subnetworks, at least two gated hybrid subnetworks, and a confidence-driven integrated network;

[0154] The feature encoding subnetwork includes K layers of encoding units, and each layer of encoding units is connected to a GELU activation function. The input of the first L1 layer encoding unit is the common point features of the limb key points, including: spatial position distribution and amplitude discreteness. The input of the second K-L1 layer encoding unit is the common motion features of the limb movement, including speed temporal changes and distribution entropy. The number of neurons in the first L1 layer encoding unit is configured as 64, and the number of neurons in the second K-L1 layer encoding unit is configured as 128.

[0155] The gated hybrid sub-network is constructed based on the input layer, the attention hidden layer, and the gated output layer. The outputs of the M1 feature encoding sub-networks are dynamically weighted through the attention mechanism. The attention hidden layer adjusts the weight coefficient in real time according to the correlation between features.

[0156] The confidence-driven integrated network is used to calculate the variance of the output of each gated hybrid sub-network to estimate the prediction confidence, and dynamically weighted sum the output of the gated hybrid sub-network according to the confidence to generate the balance coefficient of the corresponding limb key point.

[0157] In this embodiment, M1 parallel feature encoding sub-networks are parallel branches that independently process the features of different limb joints (such as encoding the features of the wrist, elbow, and shoulder respectively), avoiding confusion of multi-joint information and improving feature extraction accuracy. For example, M1=3, branch 1 encodes the key point features corresponding to the wrist joint, branch 2 encodes the key point features corresponding to the elbow joint, and branch 3 encodes the key point features corresponding to the shoulder joint.

[0158] In this embodiment, the K-layer encoding unit is a stacked fully connected layer (or convolutional layer), which gradually extracts the deep representation of the features (K = 5, an empirical value that can be determined by hyperparameter tuning). Specifically, nn.Sequential defines 5 layers of nn.Linear, and the dimension of each layer is: 12→64→128→256→256.

[0159] In this embodiment, the GELU activation function is a Gaussian error linear unit (formula: GELU(x) = x·Φ(x), where Φ is the standard normal distribution CDF), which is smoother than ReLU and alleviates gradient vanishing. Specifically, after the elbow joint velocity change rate passes through GELU, the activation value is more continuous, retaining the detailed trend of acceleration → deceleration.

[0160] Common features of points (first L1 layer input):

[0161] Spatial position distribution: Spatial coordinate statistics (such as mean, range, and cluster center) of key points in multiple repeated actions. For example, the wrist joint release point of 10 shots has the following values: x mean 1.2m, y mean 2.1m, and range (x∈[1.15,1.25], y∈[2.05,2.15]).

[0162] Amplitude dispersion is the degree of dispersion of key point positions (standard deviation, coefficient of variation), which reflects the consistency of movement. For example, the standard deviation of the x-coordinate of the wrist joint is 0.05m, and the coefficient of variation CV = 0.05 / 1.2×100% ≈ 4.2% (<5% is stable).

[0163] Motion common features (after K-L1 layer input):

[0164] The temporal change of speed is the trend of the speed of the key point over time (such as the slope of the acceleration segment and the duration of the deceleration segment). For example, the ankle speed when dribbling and changing direction: 0→2s acceleration (slope 1.5m / s 2 ), 2→4s deceleration (slope -1.2m / s 2 ).

[0165] Computation: Numerical differentiation Fit the speed curve to find the trend.

[0166] Distribution entropy is the probability distribution entropy of motion parameters (such as speed). The larger the entropy, the more disordered the motion. For example, the distribution entropy of wrist speed for shooting is H=1.2 (<2 means regularity, >3 means disorder).

[0167] The gated hybrid subnetwork is a module that fuses the outputs of multiple encoding subnetworks. It dynamically assigns weights through the attention mechanism to highlight joint features that are strongly related to movement stability. For example, when shooting, the output weight of the wrist joint encoding subnetwork (0.6) is higher than that of the elbow joint (0.3) and shoulder joint (0.1). Specifically, a three-layer network is constructed (input layer → attention hidden layer → gated output layer), and the nn.MultiheadAttention mechanism is used to implement the attention mechanism.

[0168] The attention hidden layer calculates correlations between features and generates dynamic weights (e.g., the higher the correlation between wrist joint spatial distribution and hand stability, the greater the weight). For example, if the wrist joint feature has a correlation of 0.85 with the balance coefficient, the weight is increased to 0.7 (the initial uniform weight is 0.33). Specifically, multi-head attention (number of heads = 4) is used, with query, key, and value dimensions all 256, and weighted features are output. The gated output layer controls the intensity of feature flow through using a Sigmoid gate. The formula is: Output = Attention(X)⊙σ(WX+b) (⊙ is element-by-element multiplication). For example, the gating coefficient for the elbow joint feature is 0.8 (allowing 80% of features to flow through) to suppress noise interference.

[0169] In this embodiment, the confidence-driven integration network calculates the variance of the outputs of multiple hybrid sub-networks, evaluates the prediction confidence, and dynamically weights the fusion results (the smaller the variance, the higher the weight). For example, the output variance of hybrid sub-network A is 0.01 (high confidence, weight 0.6), and the output variance of sub-network B is 0.04 (low confidence, weight 0.4). The output variance is the standard deviation of the outputs of multiple hybrid sub-networks (the larger the variance, the more unstable the prediction).

[0170] Dynamic weighted summation is a confidence weighted formula and is an existing technology.

[0171] The beneficial effects of the above technical solution are: through the deep architecture of multi-joint parallel encoding, dynamic fusion of attention, and confidence-weighted integration, the spatial stability (point commonality) and movement smoothness (movement commonality) of limb key points are accurately decoupled, the feature weights of different joints are dynamically adapted, the prediction error of the balance coefficient is reduced, and the output balance coefficient provides a highly reliable basis for locating movement defects and optimizing training programs.

[0172] The present invention provides a basketball sports training method based on virtual reality, step 3, comprising:

[0173] A twin model of movement defects and error trajectories is constructed, and the wavelet time-frequency characteristics of the athlete's historical movements are correlated with the spatial tactical characteristics of typical error trajectories in virtual reality training scenarios in time and space dimensions to generate a causal relationship map between movement defects and error inducements;

[0174] Based on the training goals, personal basic data and current training level, and in combination with the cause-effect relationship map, an initial training plan is generated.

[0175] Preferably, after generating the initial training plan, the method further includes:

[0176] Extract the abnormal frequency components of the angles of the limb key points from the wavelet features of historical movements, and build a set of mechanical constraint rules for movement correction based on the human joint kinematic thresholds.

[0177] The initial training plan is modified based on the mechanical constraint rule set to obtain a personalized training plan.

[0178] In this embodiment, the twin model uses digital twin technology to construct a mapping relationship between real-life action defects and virtual error trajectories (typical erroneous motion paths in virtual scenes). For example, elbow valgus (real-life action defect) is associated with a shot 0.3m to the left (virtual error trajectory). Specifically, PyTorch is used to construct a two-branch neural network. One branch inputs the joint angle sequence of historical actions (real data), and the other inputs the spatiotemporal coordinates of the error trajectory of the virtual scene (virtual data). The model is trained through contrastive learning to minimize the characteristic distance between similar defect-trajectories.

[0179] In this embodiment, the wavelet time-frequency characteristics of historical actions are the action frequency-time distribution extracted by wavelet transform (such as the 10-15Hz high-frequency component of the wrist force when shooting, reflecting jitter); the spatial tactical characteristics of typical error trajectories refer to the spatial distribution of virtual error trajectories (such as 80% of pass errors are concentrated at the top of the arc) and tactical background (such as failed baseline breakthroughs under double-team defense).

[0180] The spatiotemporal correlation mapping combines the two: for example, a 20Hz abnormal elbow frequency 0.2 seconds before a shot (a time-frequency feature) leads to a shot deflecting 0.3 meters to the left (a spatial tactical feature). The causal relationship map presents this relationship as a graph (nodes represent the abnormal elbow frequency and the shot deflecting to the left, and edges represent the cause). Specifically, PyWavelets performs a continuous wavelet transform (CWT) to extract time-frequency features, records virtual error trajectories in Unity, and creates a spatial heat map. A graph neural network (GNN, such as PyTorchGeometric) learns the associations and generates a causal map (node ​​embeddings and edge weights represent causal strength).

[0181] In this embodiment, an initial training program is generated and modified. For example, for elbow valgus → missed shooting, 10 sets of elbow correction shooting training are planned every day.

[0182] Abnormal angular frequency components are frequencies outside the normal range in the wavelet time-frequency spectrum (e.g., a 15Hz component appears at the elbow during a basketball shot, while the normal range is 5-10Hz). The human joint kinematic threshold is the physiological range of joint motion (e.g., knee flexion angle of 120°-150°). The mechanical constraint rule set combines these two: for example, if the 15Hz component at the elbow is >30% and the angle is >175°, the training intensity is limited to 20%. Specifically, a genetic algorithm is used to optimize the initial plan (constrained by the defect-cause relationship in the causal graph). Rules are written in Prolog or logically reasoned in Python, and the plan is modified based on the mechanical constraints (e.g., reducing 10 shooting sets to 8 sets and adding flexibility training). Ultimately, a personalized training plan is developed to adapt to the athlete's physiological and motor deficiencies, such as adding resistance training and neurofeedback to address wrist tremors.

[0183] The beneficial effect of the above technical solution is: through the link between digital twin correlation defects and virtual errors, and mechanical constraint-driven solution iteration, it breaks through the experience limitations of traditional training: first accurately map the spatiotemporal cause and effect of movement defects → virtual errors (such as the quantitative correlation between elbow valgus and missed shots), and then use joint kinematic thresholds to construct constraint rules (prohibiting physiological risk actions), so that personalized solutions can improve training efficiency.

[0184] The present invention provides a basketball sports training method based on virtual reality, step 4 comprising:

[0185] Extracting the motion specification features of the motion data based on a spatiotemporal graph convolutional network, and analyzing the dynamic evolution pattern of the physiological data based on a variational autoencoder;

[0186] Based on a graph neural network, the causal link between movement standard defects and abnormal physiological responses is learned, and the movement standard characteristics and dynamic evolution patterns are analyzed to construct the athlete's movement-physiology spatiotemporal causal graph;

[0187] Based on the pre-trained time series prediction model, the action-physiology spatiotemporal causal graph is predicted and analyzed, and the pre-intervention strategy is output in combination with the automatically labeled action specifications;

[0188] The causal effect value of the pre-intervention strategy and the training effect is quantified to determine the significant status, and the personalized training program is optimized.

[0189] Preferably, the automatically marked action specifications include:

[0190] A 3D pose estimation network is used to analyze the motion data, output normative scores of limb key points, and locate key propagation nodes of motion defects through an attention mechanism;

[0191] determining a force efficiency coefficient of the athlete based on the physiological data;

[0192] The standard score, key propagation nodes, force efficiency coefficient and fatigue correlation are automatically labeled to obtain the movement standardization situation.

[0193] In this embodiment, the spatiotemporal graph convolutional network is used to extract action specification features: taking basketball dribbling as an example, the human joints are modeled as a graph structure (nodes are joints such as wrists, elbows, shoulders, and edges are joint connection relationships), and the spatiotemporal convolution layer is used to capture spatiotemporal features such as whether the ankle twisting angle is within the physiological threshold (20°-30°) when dribbling and changing direction, and output the action specification feature vector (such as a vector with a dimension of 128, encoding the compliance of joint angles and speeds).

[0194] Variational Autoencoders (VAEs) analyze the dynamic evolution of physiological data: For physiological signals such as heart rate and electromyography, a VAE model (implemented in PyTorch, including encoder and decoder) is constructed to learn the latent distribution of these signals. For example, during training, the frequency of electromyography signals decreases from 25Hz to 10Hz as fatigue increases. The VAE extracts this high-frequency to low-frequency evolution pattern and represents it as a trajectory in the latent space.

[0195] Graph neural networks (GNNs) learn causal links: They use movement defects (such as elbow valgus exceeding 15°) and physiological response abnormalities (such as a 30% drop in biceps brachii EMG energy) as graph nodes and calculate node association weights using an attention mechanism (such as the GAT model). For example, the GNN discovered that the causal strength between elbow valgus and EMG abnormalities was 0.8 (edge ​​weight), and that the association occurred 0.1 seconds after the movement (temporal attribute). Ultimately, a movement-physiology spatiotemporal causal graph was constructed (nodes contain spatiotemporal labels, and edges contain causal strengths).

[0196] The time series prediction model, based on causal graphs, takes a historical 100-frame causal graph sequence as input and predicts the next 30 seconds of motion risks (e.g., wrist shooting angle deviating by 10°) and physiological risks (e.g., heart rate exceeding 180 beats per minute). Combined with automatically annotated motion specifications (e.g., the 3D posture estimation network outputs a wrist specification score of 82, with the elbow as the key propagation node), it outputs a pre-intervention strategy: If wrist deformation is predicted, a real-time prompt is triggered to apply 15% additional wrist force; if physiological limits are exceeded, a mandatory 60-second rest period plus fascia relaxation guidance is initiated.

[0197] Causal effect sizes were quantified using a causal forest. For example, adding wrist resistance training resulted in a 12% improvement in shooting accuracy, yielding a calculated effect size of 0.12 (>0.1 is significant). This strategy was retained and strengthened. Based on this, personalized training plans were dynamically optimized. If the effect size for elbow flexibility training was low (<0.05), neuromuscular electrical stimulation training was substituted, achieving precise iteration of the strategy.

[0198] The 3D posture estimation network (such as HRNet) analyzes the motion data: it outputs a joint standard score (such as a wrist standard score of 85 points out of 100 for shooting) and locates key propagation nodes through the attention mechanism (such as the elbow is a key node in elbow valgus → wrist deformation). The force efficiency coefficient is calculated based on the electromyographic signal: when shooting, the effective electromyographic force component accounts for 75% (the ideal value is 85%), resulting in a coefficient of 0.75. The fatigue correlation quantifies the correlation between motion deformation and fatigue (such as 0.65, indicating that 65% of the motion deformation is caused by fatigue). Finally, the standard score of 85, the key node elbow, the force efficiency of 0.75, and the fatigue correlation of 0.65 are automatically labeled to fully describe the motion standardization.

[0199] The beneficial effects of the above technical solution are: using spatiotemporal graph convolution and VAE to deeply explore action-physiological associations (such as the millisecond-level causality between elbow valgus and electromyographic abnormalities), and then through time series prediction, combined with causal effect quantification to achieve scientific iteration of strategies.

[0200] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A basketball sports training method based on virtual reality, characterized in that: include: Step 1: Obtain the athlete's basic personal data and capture data of typical movements, and integrate them into the virtual character; Step 2: Build a virtual reality training scenario based on the actual scenarios and competition requirements of basketball, add virtual opponents and teammates to the virtual reality training scenario, and use artificial intelligence algorithms to assign different competitive levels and tactical styles. Based on the collected real-world trajectory library, the movement trajectories are randomly assigned to the corresponding virtual opponents and teammates. Step 3: Develop a personalized training plan based on the athlete's training goals, personal basic data, and current training level, combined with the athlete's historical movement wavelet feature location and typical error trajectory simulation training in the virtual reality training scenario; Step 4: During the training process, the athlete's motion data and physiological data are captured in real time, and the movement standards are automatically marked to guide and intervene in the athlete's training process. After the training, the athlete's training effect is comprehensively evaluated according to the preset evaluation index system to optimize the personalized training plan.

2. The basketball sports training method based on virtual reality according to claim 1, characterized in that: After obtaining the athlete's capture data, it also includes: dividing the action video of the athlete into frames based on a number of repetitions of each typical action; According to the action type of each typical action, match the salient features from the type-salient comparison table; Arrange the limb contour trajectory based on the limb key points in each repeated video frame corresponding to the typical action in sequence according to the first and last coordinates to obtain the initial complete trajectory; Extracting the whole-body motion profile of each video frame for the corresponding number of repetitions of the typical action, analyzing the whole-body motion profile according to the salient characteristics, locking the initial key trigger frame and the end key trigger frame for the corresponding typical action, and obtaining the key trajectory segment; Determine the splicing state of the splicing position of adjacent video frames. When the splicing position is in a key trajectory segment, extract local geometric features of corresponding limb key points in n0 adjacent video frames closest to the splicing position, and place the local geometric features in chronological order. Combined with the splicing state, estimate the first geometric feature of the corresponding splicing position. Combined with the salient characteristics, perform an interpolation processing method for the corresponding splicing position in the key trajectory segment to optimize and update the connection between the two splicing points of the corresponding splicing position. When the splicing position is not in a key trajectory segment, when the splicing state is point overlap, the intersection point of the corresponding splicing position is kept unchanged and redundant non-overlapping line segments are removed; when the splicing state is point non-intersection, the two points of the splicing position of adjacent video frames are connected by a straight line; Obtaining a required complete trajectory corresponding to a limb key point based on an updated result of the splicing position in the initial complete trajectory; Perform coordinate mapping on all required complete trajectories of each limb key point under the same typical action to determine the commonality of the corresponding limb key points; At the same time, a dynamic analysis is performed on each limb movement under the same typical action, and the motion commonality of each limb key point is determined, wherein the point commonality is related to the amplitude characteristic distribution of the corresponding limb key point, and the motion commonality is related to the motion speed distribution of the corresponding limb key point; Inputting the point commonality and movement commonality into a balance analysis model to obtain a balance coefficient corresponding to a key point of the limb; If the balance coefficient is greater than the preset coefficient, the captured data under the corresponding typical action is kept unchanged; Otherwise, the captured data under the corresponding typical action is regarded as the first data, and the first data is subjected to wavelet analysis to determine the filtering threshold, and the noise data of the captured data is filtered out to obtain the second data; The retained data and the second data are converted into an action standard to obtain an action sequence, and the action sequence is merged with the basic sequence obtained by converting the personal basic data into a standard sequence, and is integrated into the virtual character.

3. The basketball sports training method based on virtual reality according to claim 2, characterized in that: Performing wavelet analysis on the first data to determine a filtering threshold includes: Based on the CNN wavelet-based learning module, basketball action clips are pre-classified, the optimal wavelet decomposition scale L is automatically matched, and the key action dimensions are enhanced through the attention mechanism of the decomposed wavelet coefficients; Extract the action periodicity of the key action dimension through Fourier transform and determine the action period weight factor Wherein, E0 represents the main frequency energy proportion of the key action dimension; Emax represents the energy proportion of the ideal stable period; Obtain the angle sequence of the limb key points of each repetitive action video under each typical action in turn, determine the average change rate of each limb key point, and obtain the standard deviation σi of each limb key point and the mean absolute value covariance between the limb key points age Get action complexity Where ∝1 and ∝2 are weights respectively, and ∝1+∝2=1; Calculate the filtering threshold Yy according to the action cycle weight factor ω1 and the action complexity Cf; Wherein, δ1 represents the deviation between the median and mean of the wavelet coefficients; N represents the data length corresponding to the first data.

4. The basketball sports training method based on virtual reality according to claim 2, characterized in that: The balance analysis model includes: M1 parallel feature encoding sub-networks, at least two gated hybrid sub-networks, and a confidence-driven integration network; The feature encoding subnetwork includes K layers of encoding units, and each layer of encoding units is connected to a GELU activation function. The input of the first L1 layer encoding unit is the common point features of the limb key points, including: spatial position distribution and amplitude discreteness. The input of the second K-L1 layer encoding unit is the common motion features of the limb movement, including speed temporal changes and distribution entropy. The number of neurons in the first L1 layer encoding unit is configured as 64, and the number of neurons in the second K-L1 layer encoding unit is configured as 128. The gated hybrid sub-network is constructed based on the input layer, the attention hidden layer, and the gated output layer. The outputs of the M1 feature encoding sub-networks are dynamically weighted through the attention mechanism. The attention hidden layer adjusts the weight coefficient in real time according to the correlation between features. The confidence-driven integrated network is used to calculate the variance of the output of each gated hybrid sub-network to estimate the prediction confidence, and dynamically weighted sum the output of the gated hybrid sub-network according to the confidence to generate the balance coefficient of the corresponding limb key point.

5. The basketball sports training method based on virtual reality according to claim 1, characterized in that: Step 3 includes: A twin model of movement defects and error trajectories is constructed, and the wavelet time-frequency characteristics of the athlete's historical movements are correlated with the spatial tactical characteristics of typical error trajectories in virtual reality training scenarios in time and space dimensions to generate a causal relationship map between movement defects and error inducements; Based on the training goals, personal basic data and current training level, and in combination with the cause-effect relationship map, an initial training plan is generated.

6. The basketball sports training method based on virtual reality according to claim 5, characterized in that: After generating the initial training plan, it also includes: Extract the abnormal frequency components of the angles of the limb key points from the wavelet features of historical movements, and build a set of mechanical constraint rules for movement correction based on the human joint kinematic thresholds. The initial training plan is modified based on the mechanical constraint rule set to obtain a personalized training plan.

7. The basketball sports training method based on virtual reality according to claim 1, characterized in that: Step 4 includes: Extracting the motion specification features of the motion data based on a spatiotemporal graph convolutional network, and analyzing the dynamic evolution pattern of the physiological data based on a variational autoencoder; Based on a graph neural network, the causal link between movement standard defects and abnormal physiological responses is learned, and the movement standard characteristics and dynamic evolution patterns are analyzed to construct the athlete's movement-physiology spatiotemporal causal graph; Based on the pre-trained time series prediction model, the action-physiology spatiotemporal causal graph is predicted and analyzed, and the pre-intervention strategy is output in combination with the automatically labeled action specifications; The causal effect value of the pre-intervention strategy and the training effect is quantified to determine the significant status, and the personalized training program is optimized.

8. The basketball sports training method based on virtual reality according to claim 7, characterized in that: Automatically marked action specifications include: A 3D pose estimation network is used to analyze the motion data, output normative scores of limb key points, and locate key propagation nodes of motion defects through an attention mechanism; determining a force efficiency coefficient of the athlete based on the physiological data; The standard score, key propagation nodes, force efficiency coefficient and fatigue correlation are automatically labeled to obtain the movement standardization situation.

Images