Motion injury prevention system based on image analysis

Abstract

The present application relates to the technical field of computer vision and sports biomechanics, and particularly relates to a motion injury prevention system based on image analysis, wherein a posture estimation module reconstructs three-dimensional key point coordinates through perspective quality weighted triangulation; a feature calculation module performs confidence adaptive weighted filtering, recursively calculates joint torque through inverse dynamics, and propagates uncertainty quantization torque error boundary along the kinematic chain; a fatigue feature extraction module extracts motion fatigue features through a time convolution network embedded with a fatigue feedback gating mechanism; a risk assessment module fuses biomechanical and fatigue features through a cascaded bidirectional cross-attention mechanism to output an injury risk level probability distribution; and a baseline calibration module forms a three-dimensional complementary calibration system with Mahalanobis distance, dynamic time warping distance and periodic index bias to eliminate evaluation misjudgment caused by individual differences. The present application realizes non-contact quantitative evaluation and individual calibration of motion injury risk.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of computer vision and sports biomechanics, specifically to a sports injury prevention system based on image analysis. Background Technology

[0002] Sports injury prevention has long been a core issue of concern in the fields of sports science and sports medicine. Traditional sports injury risk assessment mainly relies on the subjective experience of sports medicine experts or contact-based acquisition schemes based on wearable sensors. The former is limited by the professional level and observation angle of the assessor, while the latter has inherent limitations such as high equipment cost, discomfort when worn, and interference with athletic performance. In recent years, with the rapid development of computer vision and deep learning technologies, non-contact sports injury prevention schemes based on image analysis have gradually gained attention. These schemes quantitatively assess athletes' injury risks by detecting key points and calculating biomechanical features from motion images captured by cameras. However, existing systems still have significant shortcomings in areas such as control of detection error propagation, adaptive extraction of fatigue state, and calibration for individual differences.

[0003] Chinese patent CN112568898A discloses a method, apparatus, and device for automatically assessing injury risk and correcting movement based on visual images of human motion. This scheme acquires human motion videos using camera equipment, extracts key point coordinates using a posture estimation network, calculates kinematic parameters such as joint angles, and compares them with a standard movement template to assess injury risk and provide correction suggestions. The core flaw of this scheme lies in treating the key point coordinates output by the posture estimation network as deterministic input for direct kinematic parameter calculation, without applying differentiated weighting to key point coordinates at different detection accuracies. In actual sports scenarios, factors such as rapid athlete movement, limb occlusion, and changes in camera perspective cause significant fluctuations in the detection accuracy of some key points, and the coordinate errors of low-accuracy key points are directly transmitted to the calculated kinematic parameters. Furthermore, this scheme only performs template-based risk assessment at the joint angle level, without establishing a joint torque calculation framework based on inverse dynamics. It cannot quantify the cumulative effect of joint load and the cascading amplification effect of detection errors propagating along the skeletal kinematic chain from distal to proximal.

[0004] Chinese patent CN118072389A discloses a method for calculating joint torque based on visual recognition. This method acquires human motion images using a camera, obtains the coordinates of key points on the human body using visual recognition technology, constructs a rigid multibody dynamics model based on a human ensemble model, and calculates the torque during joint movement using mechanical theory. The core limitation of this method is that it treats the key point coordinates output by visual recognition as deterministic input and directly uses them for dynamic modeling. The dynamic calculation chain does not include the propagation and quantification of uncertainty in the input data. The inherent random error in attitude estimation will propagate and amplify step by step along the kinetic chain during the inverse dynamics recursion process. This method only calculates and outputs instantaneous joint torque, without quantifying the cumulative effect of joint torque over a period of time, and lacks the ability to assess the risk of repetitive load injury due to fatigue accumulation.

[0005] Chinese patent CN118780762A discloses a method for early warning of sports injury risks in athletes. This method establishes a dedicated database by inputting athlete information, periodically monitors training data and physical state parameters, and builds a database of injury-prone behaviors. It then determines injury risk based on the deviation between training characteristics and standard parameters. However, this method uses fixed statistical thresholds to assess injury risk and lacks a multi-dimensional baseline calibration mechanism to address individual differences. Significant differences exist in the body structure, exercise habits, and basic athletic abilities of different athletes; the same movement can exhibit different joint angle distributions, torque characteristics, and movement rhythms across individuals. Furthermore, this method fails to quantify the differences between the current exercise state and the individual baseline state from multiple dimensions, such as amplitude deviation, trajectory morphology deviation, and movement rhythm deviation. This makes it prone to misjudgment when using fixed thresholds or group statistical thresholds for assessment.

[0006] Therefore, existing technologies in the field of image analysis-based sports injury prevention face three common technical bottlenecks: First, existing vision-based solutions either lack an inverse dynamics joint torque calculation framework and remain at a superficial kinematic level, or although they have inverse dynamics calculation capabilities, they fail to quantify the effects of uncertainty in key point detection propagating along the kinetic chain, resulting in insufficient reliability of proximal joint torque calculation results; Second, there is a lack of adaptive feature extraction mechanisms and joint load accumulation quantification methods for sports fatigue, making it difficult to effectively capture the dynamic evolution characteristics of fatigue state and the cumulative effect of repetitive loads in the time dimension; Third, assessment schemes based on fixed thresholds or population statistical thresholds cannot adapt to significant differences in movement characteristics between individuals, and the assessment results are greatly affected by the deviation of individual baseline states. Summary of the Invention

[0007] This invention discloses a sports injury prevention system based on image analysis, which aims to overcome the shortcomings of existing sports injury assessment systems in key point detection error propagation control, fatigue feature adaptive extraction, and individual difference baseline calibration, and to establish a complete non-contact sports injury prevention technology solution from image acquisition to injury risk calibration output.

[0008] An image-based sports injury prevention system includes a processor, a memory, an image acquisition module, a posture estimation module, a feature calculation module, a fatigue feature extraction module, a risk assessment module, and a baseline calibration module. The memory stores computer-readable instructions. When the processor executes these instructions, it causes the system to perform processing operations in the image acquisition module, posture estimation module, feature calculation module, fatigue feature extraction module, risk assessment module, and baseline calibration module, respectively.

[0009] The image acquisition module acquires a multi-view synchronous image sequence of the target person during movement and transmits it to the attitude estimation module; the attitude estimation module performs human key point detection and three-dimensional coordinate reconstruction on the multi-view synchronous image sequence, generates a temporal key point coordinate sequence and a confidence value corresponding to each human key point, and transmits the temporal key point coordinate sequence and the confidence value to the feature calculation module, and at the same time transmits the temporal key point coordinate sequence to the fatigue feature extraction module.

[0010] The feature calculation module performs adaptive weighted filtering on the temporal key point coordinate sequence based on the confidence value, and calculates the spatiotemporal biomechanical feature vector including joint angle, joint angular velocity, joint torque and cumulative joint load based on the filtered temporal key point coordinate sequence and human skeletal topology. The spatiotemporal biomechanical feature vector is then transmitted to the fatigue feature extraction module and the risk assessment module, respectively.

[0011] The fatigue feature extraction module takes the temporal key point coordinate sequence and the spatiotemporal biomechanical feature vector as dual inputs, and extracts the motion fatigue feature vector along the time dimension through a temporal convolutional network including channel attention mechanism and temporal attention mechanism and transmits it to the risk assessment module.

[0012] The risk assessment module uses spatiotemporal biomechanical feature vectors as query inputs and motion fatigue feature vectors as key inputs. It generates a fusion feature vector that couples biomechanics and fatigue through a cross-attention mechanism, and outputs the probability distribution corresponding to each damage risk level after performing nonlinear transformation and normalization classification processing on the fusion feature vector.

[0013] The baseline calibration module obtains a baseline biomechanical feature vector based on a baseline motion image sequence collected when the target person performs similar motion actions under a fatigue-free baseline state. It calculates the deviation measure of the current motion's spatiotemporal biomechanical feature vector relative to the baseline biomechanical feature vector, maps the deviation measure to a calibration weight coefficient, and uses the calibration weight coefficient to weight and correct the probability distribution corresponding to each damage risk level output by the risk assessment module.

[0014] Furthermore, the image acquisition module includes three or more image acquisition devices, which are respectively deployed at different spatial locations on the sports field. These devices acquire image frames simultaneously via a hardware synchronization triggering mechanism, outputting three or more image frames from different spatial locations at each moment to form a multi-view synchronized image sequence. The pose estimation module includes a 3D keypoint coordinate reconstruction module. This module performs triangulation calculations on the 2D pixel coordinates of human keypoints extracted from the image frames based on the camera intrinsic and extrinsic parameter matrices of the image acquisition devices, converting the 2D pixel coordinates into 3D spatial coordinates. It then establishes a world coordinate system with the sports field ground as the reference plane and outputs the 3D coordinate values ​​of each human keypoint in the world coordinate system at each moment. The 3D keypoint coordinate reconstruction module also includes a viewpoint quality weighted reconstruction module. This module performs a weighted geometric average calculation on the detection confidence value of each human keypoint under each visible image acquisition device, the viewpoint cosine value between the camera optical axis and the line connecting the keypoint to the camera center, and the baseline depth ratio to obtain a viewpoint quality factor. The calculation formula is:

[0015] ;

[0016] in, For the first The confidence value of the detection of key points on the human body by an image acquisition device. For the first The angle between the optical axis of the camera of the image acquisition device and the line connecting the key point to the center of the camera. For the first The stereo baseline length between an image acquisition device and its adjacent image acquisition devices. For the key points of this human body to the first The depth distance of the camera center of each image acquisition device;

[0017] The viewpoint quality weighted reconstruction module selects the two target viewpoints with the highest viewpoint quality factors from all visible image acquisition devices of the human body key point, performs triangulation calculation on the two-dimensional pixel coordinates of the two target viewpoints, and performs weighted fusion processing on the triangulation results using the viewpoint quality factors of the two target viewpoints as weighting coefficients to generate the optimized three-dimensional coordinate values ​​of the human body key point.

[0018] Furthermore, the pose estimation module also includes a feature extraction backbone network, a feature pyramid network, and a keypoint regression head. The feature extraction backbone network performs convolution operations on each frame of the multi-view synchronized image sequence and extracts multi-scale feature maps. The feature pyramid network performs top-down feature fusion processing on the multi-scale feature maps and outputs semantically enhanced feature maps. The keypoint regression head outputs the two-dimensional coordinates and corresponding confidence values ​​of each visible human keypoint based on the semantically enhanced feature maps. The pose estimation module compares the confidence values ​​with a confidence threshold determined based on the statistical analysis of keypoint detection accuracy in the training set and retains valid keypoints with confidence values ​​higher than the confidence threshold. It performs inter-frame matching and temporal smoothing processing on valid keypoints in consecutive frames to generate an initial temporal keypoint coordinate sequence. The pose estimation module also includes a biomechanical constraint verification module, which receives the initial temporal keypoint coordinate sequence. The algorithm calculates the joint angle change rate of each joint between adjacent frames based on the keypoint coordinates of adjacent frames. It then compares the joint angle change rate of each joint with the upper bound of the corresponding angle change rate constraint in the human joint physiological motion parameter set. Simultaneously, it compares the joint angle of each frame with the corresponding angle range boundary in the human joint physiological motion parameter set. When the joint angle change rate of a joint in a frame exceeds the upper bound of the angle change rate constraint or the joint angle exceeds the angle range boundary, the joint angle of that joint in that frame is projected onto the corresponding constraint boundary value. Simultaneously, the 3D coordinates of the corresponding keypoint in that frame are corrected, and the confidence value of the corresponding keypoint in that frame is reduced by a preset attenuation ratio. The algorithm outputs the initial temporal keypoint coordinate sequence and the updated confidence value after constraint verification. These initial temporal keypoint coordinate sequence and updated confidence value are then output as the temporal keypoint coordinate sequence and confidence value, respectively.

[0019] Furthermore, the feature calculation module includes an adaptive weight filtering module, a joint angle calculation module, an angular velocity calculation module, and a torque calculation module. The adaptive weight filtering module constructs an adaptive weight matrix based on the confidence value of each human keypoint output by the keypoint regression head. This adaptive weight matrix is ​​then used to perform weighted filtering on the 3D coordinates of each human keypoint in the temporal keypoint coordinate sequence, and the filtered temporal keypoint coordinate sequence is output. The values ​​of the diagonal elements in the adaptive weight matrix are linearly proportional to the confidence value. The calculation method is as follows:

[0020] ;

[0021] in For the first Confidence values ​​of key points on an individual's body The scaling factor is used to map the confidence score to the weight range, and is set as the ratio of the upper bound of the weight range to the maximum confidence score.

[0022] The joint angle calculation module calculates the joint angle based on the direction vectors of adjacent bone segments in three-dimensional space in the filtered temporal keypoint coordinate sequence, and the angular velocity calculation module calculates the joint angular velocity based on the ratio of the joint angle difference between adjacent frames to the corresponding time interval.

[0023] The joint angular velocity The calculation formula is:

[0024] ;

[0025] in, The discrete frame number of the current frame. Frame number Corresponding joint angles Frame number The corresponding joint angle in the previous frame, The sampling time interval between adjacent frames;

[0026] The torque calculation module calculates joint torques using a bottom-up inverse dynamics recursive method based on human segmental inertial parameters and kinematic data derived from filtered temporal key point coordinate sequences. The human segmental inertial parameters include segmental mass, segmental center of mass position, and segmental moment of inertia. The inverse dynamics recursive method, based on Newton's laws of motion, recursively calculates torques segment by segment from the distal limb towards the proximal end. For the ankle joint, the joint torque... The calculation formula is:

[0027] ;

[0028] in, For foot segment quality, It is the acceleration due to gravity. Let be the linear acceleration of the foot segment's center of mass in the sagittal plane along the vertical direction. Let be the horizontal coordinate of the foot segment's centroid in the sagittal plane. The horizontal coordinate of the ankle joint center in the sagittal plane. The horizontal lever arm of the foot segment's center of mass relative to the ankle joint center is given. The vertical acceleration of the foot segment's center of mass is obtained by performing a second-order central difference calculation on the vertical coordinates of the foot segment's center of mass in consecutive frames. For the knee and hip joints, the corresponding joint torques are calculated sequentially along the limb chain towards the proximal end based on the topology of the human skeleton.

[0029] Furthermore, the feature calculation module also includes an uncertainty propagation module, a joint load accumulation module, and a feature aggregation module. The uncertainty propagation module calculates the coordinate variance estimate based on the confidence value of each human key point, and propagates the coordinate variance estimate segment by segment along the kinematic chain defined by the human skeletal topology from the distal segment to the proximal segment. For each joint angle, the joint angle variance is calculated using the error propagation formula based on the coordinate variance estimate of the key points corresponding to the two adjacent skeletal segments constituting the joint. The calculation formula is:

[0030] ;

[0031] in, For the first part that constitutes this joint The coordinate components of each key point in three-dimensional space This is the estimated coordinate variance for that coordinate component, and the estimated coordinate variance is inversely proportional to the square of the confidence level. The partial derivative of the joint angle with respect to the coordinate components of its corresponding key point is obtained by differentiating the analytical calculation expression of the joint angle with respect to the coordinate components. For each joint moment, based on the inverse dynamic recursive relationship of the moment calculation module, the partial derivatives of the joint moment with respect to the joint angle, the partial derivatives of the joint moment with respect to the segmental inertia parameter, and the partial derivatives of the joint moment with respect to the distal joint moment are substituted into the moment error propagation formula to calculate the uncertainty boundary of the joint moment. The moment error propagation formula is as follows:

[0032] ;

[0033] in, For the first Variance of joint torque of each joint Traversal affects the first All joint angles of a joint moment For the first Each joint angle For the first The variance of each joint angle, Traversing the first The inertial parameters of all segments of the joint and its associated segments. For the first Segment inertial parameters, For the first The variance of the inertial parameters of each segment, The traversal is related to the inverse dynamical recurrence relation with the first The distal joints associated with each joint For the first Joint torque of each distal joint, For the first Variance of joint moment of each distal joint For the first The joint torque is related to the first The partial derivatives of the torques of each distal joint; the joint load accumulation module performs cumulative integration calculations along the time dimension on the joint torques of each frame within the same sliding time window to obtain the cumulative load value of each joint. The calculation formula is:

[0034] ;

[0035] in, For the first time within the sliding time window The first frame Joint torque of each joint The total number of frames within the sliding time window. The sampling time interval between adjacent frames is defined as follows: The feature aggregation module combines the joint angle, joint angular velocity, joint torque, joint torque uncertainty boundary, and cumulative joint load value corresponding to consecutive frames within the same sliding time window into a spatiotemporal biomechanical feature vector.

[0036] Furthermore, the fatigue feature extraction module includes a temporal convolutional network structure, which comprises stacked one-dimensional causal convolutional layers and residual connections, a channel attention mechanism layer, a fatigue feedback gating mechanism layer, a temporal attention mechanism layer, a global average pooling layer, and a fully connected mapping layer. The fatigue feature extraction module concatenates the temporal keypoint coordinate sequence with the spatiotemporal biomechanical feature vector along the feature channel dimension to generate a dual-path fusion feature matrix. The one-dimensional causal convolutional layer performs causal convolution operations on the dual-path fusion feature matrix along the time dimension to extract temporal dependent features including changes in motion rhythm and the cumulative trend of joint load. The residual connection adds the input and output features of the one-dimensional causal convolutional layer element-wise. The channel attention mechanism layer processes the output of the residual connection... In the feature matrix, a channel description vector is obtained by calculating the global average pooling value for each feature channel. This channel description vector is then input into a two-layer fully connected network (including a first and a second fully connected layer) to generate a channel attention weight vector. The feature matrix is ​​then weighted channel-wise using this channel attention weight vector. The fatigue feedback gating mechanism layer performs global average pooling along the time dimension on the feature matrix after channel attention weighting to generate a fatigue state description vector. This fatigue state description vector is then input into a fully connected layer activated by the Sigmoid activation function to generate a fatigue gating weight vector. The feature matrix after channel attention weighting is then subjected to element-wise multiplication using this fatigue gating weight vector to obtain a gated modulation feature matrix. The calculation formula for the gated modulation is as follows:

[0037] ;

[0038] in, For the gated modulation feature matrix, The feature matrix after channel attention weighting is... This is an element-wise multiplication operation. It is the Sigmoid activation function. and These are the weight matrix and bias vector of the fully connected layer, respectively. The process involves a global average pooling operation along the time dimension. The temporal attention mechanism layer performs an average pooling operation on the gated modulation feature matrix along the feature channel dimension to obtain a time description vector. This time description vector is then input into a two-layer fully connected network, including a third and a fourth fully connected layer, to generate a temporal attention weight vector. This temporal attention weight vector is used to perform weighted processing on each time step of the gated modulation feature matrix. The global average pooling layer performs a global average pooling operation on the feature matrix after temporal attention weighting to obtain a fixed-dimensional temporal feature vector. The fully connected mapping layer performs a nonlinear mapping on the temporal feature vector to output the motion fatigue feature vector.

[0039] Furthermore, the risk assessment module includes a cascaded bidirectional cross-attention mechanism layer and a multi-layer feedforward neural network layer. The multi-layer feedforward neural network layer includes residual connections and layer normalization. The cascaded bidirectional cross-attention mechanism layer includes a first-stage cross-attention layer and a second-stage cross-attention layer. The first-stage cross-attention layer uses the spatiotemporal biomechanical feature vector as the query vector and the motion fatigue feature vector as the key and value vectors, and performs the first-stage cross-attention calculation to generate a globally fused feature vector according to the following formula:

[0040] ;

[0041] in, For global fusion feature vectors, The query vector is constructed from spatiotemporal biomechanical feature vectors. The key vector is constructed from the eigenvectors of motion fatigue. This is a value vector constructed from the eigenvectors of motion fatigue. To determine the dimensions of the query vector and the key vector, This represents the matrix transpose operation; the second-stage cross-attention layer uses the motion fatigue feature vector after residual connection processing as the query vector, and the spatiotemporal biomechanical feature vector as the key vector and value vector, to perform the second-stage cross-attention calculation to generate a local fusion feature vector. The formula for the second-stage cross-attention calculation is as follows:

[0042] ;

[0043] in, For local fusion feature vectors, The query vector is constructed from the motion fatigue feature vectors after residual join processing. The key vector is constructed from spatiotemporal biomechanical eigenvectors. The value vector is constructed from spatiotemporal biomechanical feature vectors; the cascaded bidirectional cross-attention mechanism layer concatenates the global fusion feature vector and the local fusion feature vector along the feature dimension to generate a bidirectional fusion feature vector; the multi-layer feedforward neural network layer performs nonlinear feature transformation on the bidirectional fusion feature vector and outputs the probability distribution corresponding to each damage risk level after Softmax normalization.

[0044] Furthermore, the baseline calibration module obtains a baseline biomechanical feature vector based on a sequence of baseline motion images collected when the target person performs similar movements under fatigue-free baseline conditions. It then calculates the mean vector and covariance matrix of the baseline biomechanical feature vector. During the target person's current movement, it calculates the Mahalanobis distance between the spatiotemporal biomechanical feature vector and the mean vector as a measure of point-state deviation. The calculation formula is:

[0045] ;

[0046] in, This represents the current spatiotemporal biomechanical feature vector. The mean vector of the baseline biomechanical eigenvectors. The covariance matrix of the baseline biomechanical eigenvectors. It is the inverse of the covariance matrix. This represents a matrix transpose operation; the baseline calibration module also constructs a current biomechanical trajectory vector based on the spatiotemporal biomechanical feature vectors corresponding to consecutive frames within the sliding time window, extracts a baseline biomechanical trajectory vector from the baseline motion image sequence, and calculates the dynamic time warping distance between the current biomechanical trajectory vector and the baseline biomechanical trajectory vector as a trajectory-level deviation metric. The calculation method involves searching for the optimal alignment path on the cost matrix using a dynamic programming algorithm. The cost matrix contains the [missing information - likely a specific path or element]. Line 1 The element of the column is the th element in the current biomechanical trajectory vector. The spatiotemporal biomechanical eigenvector at time t and the baseline biomechanical trajectory vector at time t. The Euclidean distance between the baseline biomechanical eigenvectors at each time step;

[0047] The baseline calibration module also calculates a periodic exponential deviation metric based on the spatiotemporal biomechanical feature vector within the sliding time window. It calculates the periodic exponent for both the current trajectory and the baseline trajectory. The periodic exponent is calculated by taking the Euclidean norm of the feature vector at each moment in the time series of the spatiotemporal biomechanical feature vector within the sliding time window to obtain a scalar time series. A normalized autocorrelation function is then calculated on the scalar time series. Finally, each effective peak value (excluding zero delay) in the normalized autocorrelation function is summed after exponential decay according to the delay order corresponding to the peak value. The effective peak value is a local maximum point where the normalized autocorrelation function value exceeds a peak threshold. The periodic exponent... The calculation formula is:

[0048] ;

[0049] in, For the delay order, This represents the total number of valid peak values ​​detected within the sliding time window. Delay order The normalized autocorrelation function value at each point is included in the summation calculation only if the value is greater than the peak threshold. The normalized autocorrelation function value is calculated by subtracting the time mean from the scalar time series, calculating the autocorrelation function, and then dividing by the autocorrelation value at zero delay. To control the decay coefficient of the peak weight decay rate at long distances; the periodicity exponent of the current motion trajectory is... Periodicity index of the reference motion trajectory The absolute value of the difference is used as a measure of periodic deviation. Mahalanobis distance Dynamic time warping distance and periodic deviation measurement The comprehensive deviation metric is obtained by weighting and summing the weight ratios determined through cross-search of the validation set. The comprehensive deviation metric is then mapped to a calibration weight coefficient, which is used to weight and correct the probability distribution corresponding to each damage risk level output by the risk assessment module.

[0050] The advantages of this invention compared to the prior art are:

[0051] 1. This invention uses adaptive weighted filtering and kinematic chain uncertainty propagation mechanism to convert confidence information into coordinate variance estimates and propagate them segment by segment along the skeletal kinematic chain from the distal end to the proximal end. The error propagation formula quantifies the cascading effect of detection errors in inverse dynamics calculation, so that the joint torque output carries uncertainty boundary information, overcoming the problem of inaccurate proximal joint torque calculation caused by low-precision key points.

[0052] 2. This invention extracts motion fatigue features through a temporal convolutional network with an embedded fatigue feedback gating mechanism. It constructs a fatigue state description vector using the global statistics of channel-weighted features and performs element-wise modulation on the features through Sigmoid gating, forming a feedback loop from fatigue state to feature selection. This allows the temporal attention layer to capture the temporal fatigue evolution law in the feature space modulated by fatigue state.

[0053] 3. This invention integrates biomechanical and fatigue features from both global and local perspectives through a cascaded bidirectional cross-attention mechanism. The first stage captures the fatigue component in global biomechanical anomalies, and the second stage captures fatigue-induced local biomechanical anomaly patterns. The outputs of the two stages are spliced ​​together to generate a probability distribution of damage risk level, providing multi-perspective fusion feature basis for damage risk localization.

[0054] 4. This invention constructs a three-dimensional complementary calibration system by using point-state Mahalanobis distance, trajectory dynamic time warping distance, and rhythm periodic exponential deviation. It quantifies individual differences from three independent dimensions: amplitude deviation, trajectory shape deviation, and motion rhythm deviation. This eliminates misjudgments caused by individual differences in fixed threshold assessment schemes and makes the assessment results adapt to the target person's own baseline motion state. Attached Figure Description

[0055] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0056] In the attached diagram:

[0057] Figure 1 This is a block diagram of the overall structure of the sports injury prevention system based on image analysis of the present invention.

[0058] Figure 2 This is a data processing flowchart of the image analysis-based sports injury prevention system of the present invention.

[0059] Figure 3 This is a flowchart of the internal processing flow of the feature calculation module of the image analysis-based sports injury prevention system of the present invention.

[0060] Figure 4 This is a network structure diagram of the fatigue feature extraction module of the image analysis-based sports injury prevention system of the present invention.

[0061] Figure 5 This is a flowchart of the three-dimensional deviation measurement calculation process for the baseline calibration module of the image analysis-based sports injury prevention system of the present invention. Detailed Implementation

[0062] The following detailed description of the embodiments is intended to exemplify the principles of this application, but should not be used to limit the scope of this application. That is, the image analysis-based sports injury prevention system of this application is not limited to the described embodiments.

[0063] The present invention will be further described below with reference to embodiments.

[0064] Example 1

[0065] This embodiment uses an indoor basketball training court as the application scenario. The court dimensions are 28 m long and 15 m wide. The target person is a basketball player, 180 cm tall and weighing 80 kg. The exercise involves performing jump shots and layups to assess the risk of anterior cruciate ligament (ACL) injury and patellar tendon strain. This embodiment defines 19 key points on the human body, including the left and right hip joints, left and right knee joints, left and right ankle joints, left and right toes, left and right heels, left and right shoulder joints, left and right elbow joints, left and right wrist joints, as well as the top of the head, neck, and center of the trunk. The lower limb inverse dynamics model divides the human lower limb into four rigid body segments: foot, lower leg, thigh, and pelvis.

[0066] In specific embodiments, such as Figure 1 As shown, the system includes a processor, memory, an image acquisition module, a pose estimation module, a feature calculation module, a fatigue feature extraction module, a risk assessment module, and a baseline calibration module. The processor uses an NVIDIA RTX 4090 graphics processor, and the memory includes 64 GB of DDR5 RAM and a 1 TB solid-state drive. The memory stores computer-readable instructions written based on the PyTorch deep learning framework. When the processor executes these computer-readable instructions, the system performs the processing operations of the image acquisition module, pose estimation module, feature calculation module, fatigue feature extraction module, risk assessment module, and baseline calibration module, respectively. The data flow between the modules is as follows: Figure 2 As shown, the output of the image acquisition module is connected to the input of the attitude estimation module. The output of the attitude estimation module is connected to the input of the feature calculation module and the fatigue feature extraction module. The output of the feature calculation module is connected to the input of the fatigue feature extraction module and the risk assessment module. The output of the fatigue feature extraction module is connected to the input of the risk assessment module. The baseline calibration module uses calibration weight coefficients to weight and correct the probability distribution output by the risk assessment module, and then outputs the calibrated damage risk assessment result.

[0067] Furthermore, such as Figure 1 and Figure 2 As shown, the image acquisition module deploys four industrial cameras around the sports field as image acquisition devices. The four cameras are located at four opposite corners of the field, with an installation height of 3.5 m and their optical axes pointing towards the center of the field. The four cameras are connected via hardware synchronization trigger lines and acquire one frame of image simultaneously under the drive of an external trigger signal. The sampling frequency is set to 120 Hz, and the resolution of each frame is [resolution missing]. Pixels. At each moment, each of the four cameras outputs one image frame. These four frames constitute a multi-view synchronized image sequence. Before acquisition, the cameras undergo intrinsic and extrinsic parameter calibration. The intrinsic parameter matrix includes focal length and principal point coordinates, while the extrinsic parameter matrix includes rotation matrix and translation vector. The calibration process employs the Zhang Zhengyou calibration method and uses a checkerboard calibration board to acquire calibration images at multiple preset locations within the field. The intrinsic and extrinsic parameter matrices for each camera are then calculated. The image acquisition module transmits the continuous multi-view synchronized image sequence to the attitude estimation module.

[0068] Furthermore, such as Figure 2 As shown, the attitude estimation module includes a 3D key point coordinate reconstruction module. The 3D key point coordinate reconstruction module performs triangulation calculations on the two-dimensional pixel coordinates of human key points extracted from each frame image based on the intrinsic and extrinsic parameter matrices of the four cameras. It converts the two-dimensional pixel coordinates into three-dimensional spatial coordinates and establishes a world coordinate system with the sports field ground as the reference plane. It outputs the three-dimensional coordinate values ​​of each human key point in the world coordinate system at each moment.

[0069] The 3D keypoint coordinate reconstruction module also includes a viewpoint quality-weighted reconstruction module, which calculates the detection confidence value of each human keypoint under each visible image acquisition device. The angle between the camera's optical axis and the line connecting this key point to the camera's center. The absolute value of the cosine and the baseline depth ratio The view quality factor is obtained by performing a weighted geometric mean operation. The calculation formula is:

[0070] ;

[0071] in For the first The confidence value of the detection of key points on the human body by an image acquisition device. For the first The angle between the optical axis of the camera of the image acquisition device and the line connecting the key point to the center of the camera. For the first The stereo baseline length between an image acquisition device and its adjacent image acquisition devices. For the key points of this human body to the first The depth distance of the camera center of the image acquisition device.

[0072] Taking the key point of the left knee joint as an example, this key point is visible in all four image acquisition devices. The detection confidence values ​​of the four image acquisition devices are 0.95, 0.88, 0.42, and 0.76, respectively. The confidence value of the third image acquisition device is lower because the athlete's right leg is partially obscured. After calculating the viewpoint quality factor, the viewpoint quality factors of the first and second image acquisition devices are the highest. The viewpoint quality weighted reconstruction module selects these two image acquisition devices as the target viewpoints, performs triangulation calculations on the two-dimensional pixel coordinates of the two target viewpoints, and performs weighted fusion processing on the triangulation results using the viewpoint quality factors of the two as weighting coefficients to generate optimized three-dimensional coordinate values ​​of the key point of the left knee joint. Through this processing, the triangulation results of low-confidence viewpoints are excluded, thereby ensuring the accuracy of three-dimensional reconstruction.

[0073] Furthermore, such as Figure 2 As shown, the pose estimation module also includes a feature extraction backbone network, a feature pyramid network, and a keypoint regression head. The feature extraction backbone network uses the high-resolution representation learning network HRNet-W48 to perform convolution operations on each frame of the multi-view synchronized image sequence and extract multi-scale feature maps. The feature pyramid network performs top-down feature fusion processing on the multi-scale feature maps and outputs semantically enhanced feature maps. The keypoint regression head outputs the two-dimensional coordinate values ​​and corresponding confidence values ​​of each visible human keypoint based on the semantically enhanced feature maps. The confidence values ​​are compared with a confidence threshold determined based on the keypoint detection accuracy statistics in the training set. In this embodiment, the confidence threshold is set to 0.3, and valid keypoints with confidence values ​​higher than 0.3 are retained. Inter-frame matching and temporal smoothing processing are performed on valid keypoints in consecutive frames to generate an initial temporal keypoint coordinate sequence.

[0074] The attitude estimation module also includes a biomechanical constraint verification module. This module receives the initial temporal keypoint coordinate sequence and calculates the joint angle change rate of each joint between adjacent frames based on the keypoint coordinates of adjacent frames. The human joint physiological motion parameter set in this embodiment refers to the human joint physiological motion range data described in Winter's *Biomechanics and Motor Control of Human Movement*. Taking the flexion and extension motion of the knee joint in the sagittal plane as an example, the angle range boundary of this joint is set as follows: to The upper bound of the angle change rate constraint is set to per frame. This value corresponds to a sampling frequency of 120 Hz per second. The maximum physiological angular velocity.

[0075] When the knee joint angle in a certain frame exceeds to The range boundary or the rate of change of the knee joint angle between adjacent frames exceeds the range boundary of each frame. When the upper bound of the constraint is reached, the biomechanical constraint verification module projects the knee joint angle of the frame onto the corresponding constraint boundary value, simultaneously corrects the three-dimensional coordinates of the corresponding key point of the frame, and lowers the confidence value of the key point of the joint in the frame by a preset attenuation ratio. In this embodiment, the attenuation ratio is set to 0.5, that is, the confidence value is halved, so that the weight of the key point in the frame in the downstream adaptive weight filtering module is reduced accordingly. The initial temporal key point coordinate sequence after constraint verification and the updated confidence value are used as the final output of the attitude estimation module. The temporal key point coordinate sequence is simultaneously transmitted to the feature calculation module and the fatigue feature extraction module, and the confidence value is transmitted to the feature calculation module.

[0076] Furthermore, such as Figure 3 As shown, the feature calculation module includes an adaptive weight filtering module, a joint angle calculation module, an angular velocity calculation module, a torque calculation module, an uncertainty propagation module, a joint load accumulation module, and a feature aggregation module.

[0077] The adaptive weight filtering module constructs an adaptive weight matrix based on the confidence values ​​of each human keypoint output by the keypoint regression head. The adaptive weight matrix is ​​a diagonal matrix, with diagonal elements... The calculation method is as follows ,in For the first Confidence values ​​of key points on an individual's body This is a scaling factor that maps confidence values ​​to the weight range. In this embodiment, the weight range is set to... Furthermore, the maximum confidence score is normalized to 1, therefore The value is 1. When the confidence value of a key point is 0.9, its corresponding weight value is 0.9, and the 3D coordinates of the key point retain most of the original information in the weighted filtering process; when the confidence value of a key point is only 0.35, its corresponding weight value is 0.35, and the 3D coordinates of the key point are significantly suppressed in the weighted filtering process, while the smooth information of adjacent frames dominates.

[0078] The joint angle calculation module calculates the joint angle based on the direction vectors of adjacent skeletal segments in three-dimensional space within the filtered temporal keypoint coordinate sequence. Taking the knee flexion-extension angle as an example, this angle is determined by the angle between the direction vectors of the thigh and lower leg segments in the sagittal plane. The angular velocity calculation module calculates the joint angular velocity based on the ratio of the joint angle difference between adjacent frames to the corresponding time interval. The calculation formula is:

[0079] ;

[0080] in, The discrete frame number of the current frame. Frame number Corresponding joint angles Frame number The corresponding joint angle in the previous frame, The sampling time interval between adjacent frames and in this embodiment s. When the knee angle in frame 30 is The knee joint angle in frame 29 is When, substitute into the formula to calculate and obtain per second .

[0081] Furthermore, such as Figure 3 As shown, the torque calculation module calculates joint torques using a bottom-up inverse dynamic recursive method based on the segmental inertial parameters of the human body and the kinematic data derived from the filtered temporal key point coordinate sequence. The inertial parameters of each segment are calculated using standard anthropometry data recorded in Winter's *Biomechanics and Motor Control of Human Movement*, based on a target person's weight of 80 kg, and the foot segment mass... The weight of the lower leg segment is 1.16 kg, the weight of the lower leg segment is 3.72 kg, and the weight of the upper leg segment is 8.0 kg.

[0082] The inverse dynamics recursive method, based on Newton's laws of motion, recursively extrapolates from the distal segment of the limb (i.e., the foot segment) towards the proximal end. For ankle joint torque... The calculation formula is:

[0083] ;

[0084] in For foot segment quality, Values ​​for gravitational acceleration , Let be the linear acceleration of the foot segment's center of mass in the sagittal plane along the vertical direction. Let be the horizontal coordinate of the foot segment's centroid in the sagittal plane. The horizontal coordinate of the ankle joint center in the sagittal plane. The horizontal lever arm represents the foot's center of mass relative to the ankle joint center. The vertical acceleration of the foot's center of mass is obtained by performing a second-order central difference calculation on the vertical coordinates of the foot's center of mass in consecutive frames. The above formula is a simplified calculation expression for the ankle joint torque in the sagittal plane, only including the torque contribution generated by the vertical inertial force of the foot's center of mass and gravity through the horizontal lever arm. The complete sagittal inverse dynamic equation should also include the torque contribution generated by the horizontal inertial force of the foot's center of mass through the vertical lever arm, as well as the torque contribution generated by the product of the foot's rotational inertia and angular acceleration. This embodiment only retains the dominant term for simplified calculation.

[0085] Taking a single frame of the athlete's landing moment as an example, at this time for and The value is 0.06 m. After substituting into the formula, the ankle joint torque is calculated. Approximately N·m. For the knee and hip joints, the corresponding joint torques are calculated sequentially from the proximal end along the kinetic chain based on the topology of the human skeleton. During the recursive calculation, the forces and torques transmitted from the distal joints must be included in the torque calculation of each proximal joint.

[0086] Furthermore, such as Figure 3 As shown, the uncertainty propagation module calculates the coordinate variance estimate based on the confidence value of each key point on the human body, and the coordinate variance estimate is inversely proportional to the square of the confidence value. In this embodiment, the coordinate variance estimate is calculated as follows: ,in The variance baseline value is determined based on the statistical analysis of keypoint detection errors in the training set. For the first The confidence score of each key point. A higher confidence score indicates a smaller coordinate variance, suggesting that the 3D coordinates of the key point are relatively reliable. Conversely, a lower confidence score indicates a larger coordinate variance, suggesting that the 3D coordinates of the key point have greater uncertainty.

[0087] For each joint angle, the joint angle variance is calculated using the error propagation formula based on the estimated variance of the coordinates of the key points corresponding to the two adjacent skeletal segments constituting that joint. The calculation formula is:

[0088] ;

[0089] in For the first part that constitutes this joint The coordinate components of each key point in three-dimensional space This is the estimated value of the coordinate variance corresponding to this coordinate component. The partial derivative of the joint angle with respect to the coordinate components of its corresponding key points is obtained by differentiating the analytical expression of the joint angle with respect to the coordinate components. Taking the knee joint angle as an example, the key points constituting this joint include the center of the hip joint, the center of the knee joint, and the center of the ankle joint. The partial derivatives of the coordinate components of each key point with respect to the knee joint angle are obtained by differentiating the inverse cosine functions of the thigh direction vector and the lower leg direction vector.

[0090] For each joint moment, based on the inverse dynamic recursion of the moment calculation module, the partial derivatives of the joint moment with respect to the joint angle, the partial derivatives of the joint moment with respect to the segmental inertia parameter, and the partial derivatives of the joint moment with respect to the distal joint moment are substituted into the moment error propagation formula to calculate the joint moment variance as a measure of the uncertainty boundary of the joint moment. The moment error propagation formula is as follows:

[0091] ;

[0092] in, For the first Variance of joint torque of each joint For the first Joint torque of each joint Traversal affects the first All joint angles of a joint moment For the first Each joint angle For the first The variance of each joint angle, For the first The joint torque is related to the first Partial derivatives of each joint angle, Traversing the first The inertial parameters of all segments of the joint and its associated segments. For the first Segment inertial parameters, For the first The variance of the inertial parameters of each segment, For the first The joint torque is related to the first Partial derivatives of the inertial parameters of each segment, The traversal is related to the inverse dynamical recurrence relation with the first The distal joints associated with each joint For the first Joint torque of each distal joint, For the first Variance of joint moment of each distal joint For the first The joint torque is related to the first The partial derivatives of the moment at each distal joint. Since the joint angular velocity and segment centroid acceleration are both obtained by numerical differentiation from the coordinates of the key points, their errors have been implicitly propagated to the first term above through the position coordinate variance, so the acceleration error term is not listed separately.

[0093] The third term of this formula embodies the core characteristic of kinetic chain uncertainty propagation. Taking the variance of knee joint moment as an example, its calculation requires incorporating the partial derivative of the variance of ankle joint moment. The contribution from propagation, specifically the contribution from the knee joint moment variance, needs to be included in the hip joint moment variance, which in turn includes the contribution from the ankle joint moment variance. This segment-by-segment propagation from distal to proximal causes the moment uncertainty boundary of the proximal joint to accumulate and increase progressively, objectively reflecting the physical law of error accumulation and amplification during the inverse dynamics recursion process.

[0094] Furthermore, such as Figure 3 As shown, the joint load accumulation module performs cumulative integration calculations along the time dimension on the joint torque of each frame within the same sliding time window to obtain the cumulative load value of each joint. The calculation formula is:

[0095] ;

[0096] in, For the first time within the sliding time window The first frame Joint torque of each joint The total number of frames within the sliding time window. This is the sampling time interval between adjacent frames. In this embodiment, the sliding time window length is set to 60 frames, corresponding to a time span of 0.5 seconds. Taking the knee joint as an example, if the average knee joint torque within 60 frames is... N·m, then substitute into the formula to calculate the cumulative load value of the knee joint. for N·m·s.

[0097] The feature aggregation module combines the joint angles, joint angular velocities, joint torques, joint torque uncertainty boundaries, and cumulative joint load values ​​corresponding to consecutive frames within the same sliding time window into a spatiotemporal biomechanical feature vector. In this embodiment, the spatiotemporal biomechanical feature vector corresponding to each sliding time window includes the angle sequence, angular velocity sequence, torque sequence, torque variance sequence, and cumulative load values ​​of the three joints of the lower limb. After being spliced ​​together, they form a feature vector of fixed dimensions.

[0098] Furthermore, such as Figure 4As shown, the fatigue feature extraction module includes a temporal convolutional network structure, which includes stacked one-dimensional causal convolutional layers and residual connections, a channel attention mechanism layer, a fatigue feedback gating mechanism layer, a temporal attention mechanism layer, a global average pooling layer, and a fully connected mapping layer.

[0099] The fatigue feature extraction module receives two inputs: one is the temporal keypoint coordinate sequence output by the attitude estimation module, and the other is the spatiotemporal biomechanical feature vector output by the feature calculation module. The fatigue feature extraction module expands the joint angle sequence, angular velocity sequence, torque sequence, and torque variance sequence from the spatiotemporal biomechanical feature vector into a temporal feature matrix of the same length as the temporal keypoint coordinate sequence, step by step. The cumulative load value of each joint is copied to each time step and concatenated with the expanded temporal feature matrix, then concatenated with the temporal keypoint coordinate sequence along the feature channel dimension to generate a dual-path fused feature matrix. A one-dimensional causal convolutional layer performs causal convolution operations along the time dimension on the dual-path fused feature matrix to extract temporal dependent features containing changes in motion rhythm and trends in joint load accumulation. In this embodiment, the temporal convolutional network includes three stacked one-dimensional causal convolutional layers with a kernel size of 3 and 128 channels. Residual connections element-wise add the input and output features of each one-dimensional causal convolutional layer to alleviate the gradient decay problem in deep networks.

[0100] The channel attention mechanism layer calculates the global average pooling value for each feature channel in the feature matrix output by the residual connection to obtain a channel description vector. This channel description vector is then input into a two-layer fully connected network (including a first and a second fully connected layer) to generate a channel attention weight vector. This channel attention weight vector is then used to perform a channel-wise weighted processing on the feature matrix. The first fully connected layer compresses the channel dimension back to the original dimension. After ReLU activation, the second fully connected layer is restored to its original channel dimension and then activated by Sigmoid, with the output value ranging from [value range missing]. The channel attention weight vector.

[0101] The fatigue feedback gating mechanism layer performs global average pooling along the time dimension on the feature matrix after channel attention weighting to generate a fatigue state description vector. This fatigue state description vector is then input into a fully connected layer activated by the Sigmoid activation function to generate a fatigue gating weight vector. This fatigue gating weight vector is then used to perform element-wise multiplication of the feature matrix after channel attention weighting to obtain a gated modulation feature matrix. The calculation formula for the gated modulation is as follows:

[0102] ;

[0103] in, For the gated modulation feature matrix, The feature matrix after channel attention weighting is... This is an element-wise multiplication operation. It is the Sigmoid activation function. and These are the weight matrix and bias vector of the fully connected layer, respectively. This is a global average pooling operation along the time dimension. The gated modulation process transforms the channel-weighted global feature statistics into values ​​ranging from... The gating weights are applied and element-wise scaling is performed on the feature matrix. When the gating weight for a feature channel is close to 0, the feature representation of that channel is significantly weakened; when the gating weight is close to 1, the feature representation of that channel is fully preserved. The fatigue state description vector gathers global temporal statistics of the channel-weighted features and can reflect the overall fatigue level of the current motion stage, enabling the gating modulation processing to adaptively adjust the feature representation intensity according to the fatigue level.

[0104] The temporal attention mechanism layer performs average pooling along the feature channel dimension of the gated modulation feature matrix to obtain a temporal description vector. This temporal description vector is then input into a two-layer fully connected network (including a third and fourth fully connected layer) to generate a temporal attention weight vector. This weight vector is then used to weight each time step of the gated modulation feature matrix. The value of each element in the temporal attention weight vector reflects the importance of the corresponding time step in fatigue feature extraction, enabling the network to adaptively focus on key time steps that significantly contribute to changes in fatigue state during the motion cycle.

[0105] A global average pooling layer performs global average pooling on the feature matrix after temporal attention weighting to obtain a fixed-dimensional temporal feature vector. A fully connected mapping layer performs nonlinear mapping on the temporal feature vector to output a motion fatigue feature vector. In this embodiment, the dimension of the motion fatigue feature vector is set to 64 dimensions.

[0106] Furthermore, such as Figure 1 and Figure 2 As shown, the risk assessment module includes a cascaded bidirectional cross-attention mechanism layer and a multi-layer feedforward neural network layer. The multi-layer feedforward neural network layer includes residual connections and layer normalization. The cascaded bidirectional cross-attention mechanism layer includes a first-stage cross-attention layer and a second-stage cross-attention layer.

[0107] The first-stage cross-attention layer uses the spatiotemporal biomechanical feature vector as the query vector and the motion fatigue feature vector as the key and value vectors. It calculates and generates a global fusion feature vector according to the following formula:

[0108] ;

[0109] in, For global fusion feature vectors, The query vector is constructed from spatiotemporal biomechanical feature vectors. The key vector is constructed from the eigenvectors of motion fatigue. This is a value vector constructed from the eigenvectors of motion fatigue. To query the dimensions of the vector and key vector, This represents the matrix transpose operation. The first stage, cross-attention calculation, establishes a global mapping between biomechanical features and fatigue features, and globally fuses feature vectors. It characterizes the fatigue-related abnormal components in the current biomechanical state.

[0110] The second-stage cross-attention layer fuses the motion fatigue feature vector with the globally fused feature vector output from the first stage. After performing element-wise addition of residual joins, the motion fatigue feature vector processed by the residual joins is used as the query vector, and the spatiotemporal biomechanical feature vector is used as the key and value vectors. A second-stage cross-attention calculation is then performed to generate a locally fused feature vector. The formula for the second-stage cross-attention calculation is as follows:

[0111] ;

[0112] in, For local fusion feature vectors, The query vector is constructed from the motion fatigue feature vectors after residual join processing. The key vector is constructed from spatiotemporal biomechanical eigenvectors. This is a value vector constructed from spatiotemporal biomechanical feature vectors. The second stage, cross-attention calculation, retrieves biomechanical feature components closely related to fatigue state from a fatigue perspective, and locally fuses the feature vectors. It characterized patterns of biomechanical abnormalities induced by fatigue, such as decreased knee joint control and altered landing cushioning strategies caused by muscle fatigue.

[0113] The cascaded bidirectional cross-attention mechanism layer globally fuses feature vectors. and local fusion feature vector A bidirectional fused feature vector is generated by concatenating features along the feature dimensions. A multi-layer feedforward neural network performs a nonlinear feature transformation on the bidirectional fused feature vector and outputs the probability distribution corresponding to each damage risk level after Softmax normalization. In this embodiment, the damage risk levels are divided into five levels: no risk, low risk, medium risk, high risk, and extremely high risk. The output probability distribution is a five-dimensional vector, with each dimension corresponding to the probability value of a risk level.

[0114] Furthermore, such as Figure 5 As shown, the baseline calibration module first acquires a baseline motion image sequence of the target person performing similar movements under a fatigue-free baseline state before the target person's formal exercise. In this embodiment, the baseline motion image sequence is acquired by having the athlete perform 10 standard jump throws while fully warmed up and without fatigue. The system statistically analyzes the spatiotemporal biomechanical feature vectors of the 10 movements to calculate the mean vector of the baseline biomechanical feature vector. Covariance Matrix When the number of repetitions of the baseline action is less than the dimension of the spatiotemporal biomechanical eigenvector, the covariance matrix... Since the covariance matrix is ​​not of rank, to ensure the computability of the inverse covariance matrix in Mahalanobis distance calculation, this embodiment applies Tikhonov regularization to the covariance matrix. Alternative Perform the inverse operation, where It is an identity matrix and For regularization parameters, in this embodiment The value is 0.01.

[0115] During the target person's current movement, the baseline calibration module calculates the Mahalanobis distance between the spatiotemporal biomechanical feature vector and the mean vector as a measure of point-state deviation. The calculation formula is:

[0116] ;

[0117] in, This represents the current spatiotemporal biomechanical feature vector. The mean vector of the baseline biomechanical eigenvectors. The covariance matrix of the baseline biomechanical eigenvectors. It is the inverse of the covariance matrix. This represents the matrix transpose operation. Mahalanobis distance normalizes the deviations of each dimension by using the inverse of the covariance matrix, thereby eliminating the influence of differences in the dimensions and value ranges between different feature dimensions. When the number of repetitions of the baseline action is less than the number of dimensions of the spatiotemporal biomechanical feature vector, the covariance matrix is ​​not full rank. In this case, the formula actually uses the inverse of the covariance matrix after Tikhonov regularization for calculation.

[0118] The baseline calibration module also constructs the current biomechanical trajectory vector based on the spatiotemporal biomechanical feature vectors corresponding to consecutive frames within the sliding time window. It extracts the baseline biomechanical trajectory vector from the baseline motion image sequence and calculates the dynamic time warping distance between the current biomechanical trajectory vector and the baseline biomechanical trajectory vector as a trajectory-level deviation metric. (Dynamic time warping distance) The calculation method involves searching for the optimal alignment path on the cost matrix using a dynamic programming algorithm. Line 1 The element of the column is the th element in the current biomechanical trajectory vector. The spatiotemporal biomechanical eigenvector at time t and the baseline biomechanical trajectory vector at time t. The Euclidean distance between the reference biomechanical feature vectors at each moment, and the dynamic time warping algorithm allow the current trajectory to be non-linearly aligned with the reference trajectory on the time axis, thus making it suitable for situations where the motion is stretched or deformed on the time scale.

[0119] The baseline calibration module also calculates a periodic exponential deviation metric based on the spatiotemporal biomechanical feature vectors within the sliding time window, calculating the periodic exponent for both the current trajectory and the baseline trajectory. The periodic exponent is calculated by taking the Euclidean norm of the feature vector at each moment in the time series of the spatiotemporal biomechanical feature vectors within the sliding time window to obtain a scalar time series. A normalized autocorrelation function is then calculated on the scalar time series. Finally, the effective peaks in the normalized autocorrelation function (excluding zero delay) are summed exponentially, weighted according to the delay order corresponding to each peak, to obtain the periodic exponent. The calculation formula is:

[0120] ;

[0121] in, For the delay order, This represents the total number of valid peak values ​​detected within the sliding time window. Delay order The normalized autocorrelation function value at a given point is included in the summation calculation only if the value is greater than the peak threshold. In this embodiment, the peak threshold is set to 0.2. The normalized autocorrelation function value is calculated by taking the scalar time series... through After centering, the autocorrelation function is calculated and divided by the autocorrelation value at zero delay. To control the attenuation coefficient of the long-distance peak weight attenuation rate, and in this embodiment... The value is 0.05.

[0122] Cyclical index Characterizing the rhythmic regularity of movement, when athletes perform repetitive movements under baseline conditions, their joint angles and torques exhibit stable periodic changes, and the normalized autocorrelation function shows a significant peak at integer multiples of the movement cycle delay, indicating a periodicity index. The value of is relatively large. When athletes enter a state of fatigue, their movement rhythm tends to be irregular, and the peak decay of the normalized autocorrelation function accelerates, making the periodicity index... Decreasing the value of will reduce the periodicity index of the current motion trajectory. Periodicity index of the reference motion trajectory Substitution Calculate the periodic deviation measure.

[0123] Mahalanobis distance Dynamic time warping distance and periodic deviation measurement The comprehensive bias metric is obtained by weighted summation of the weight ratios determined through cross-validation on the validation set. In this embodiment, the weight ratios of the three dimensions are determined by leave-one-out cross-validation on the benchmark dataset. The comprehensive bias metric is then mapped using the Sigmoid function to a value range of [value range missing]. calibration weight coefficient The probability distribution corresponding to each damage risk level output by the risk assessment module is then weighted and corrected using calibration weighting coefficients. Specifically, the weighting correction is calculated as follows: ,in, This is the original output probability distribution of the risk assessment module. To ensure uniform distribution, For normalization. The weighting coefficients are calibrated when the overall deviation metric is large. A smaller overall bias makes the output probability distribution tend to be more uniform, indicating a lower confidence level in the current risk assessment results and avoiding false alarms caused by deviations from individual baseline states; when the overall bias metric is small, the calibration weighting coefficients are adjusted accordingly. It approaches 1, thus preserving the original output of the risk assessment module.

[0124] It should be noted that the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0125] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A sports injury prevention system based on image analysis, characterized in that: The system includes a processor, a memory, an image acquisition module, a posture estimation module, a feature calculation module, a fatigue feature extraction module, a risk assessment module, and a baseline calibration module. The memory stores computer-readable instructions. When the processor executes these computer-readable instructions, it causes the system to perform processing operations on the image acquisition module, posture estimation module, feature calculation module, fatigue feature extraction module, risk assessment module, and baseline calibration module, respectively. The image acquisition module acquires a multi-view synchronous image sequence of the target person during movement and transmits it to the attitude estimation module; the attitude estimation module performs human key point detection and three-dimensional coordinate reconstruction on the multi-view synchronous image sequence, generates a temporal key point coordinate sequence and a confidence value corresponding to each human key point, and transmits the temporal key point coordinate sequence and the confidence value to the feature calculation module, and at the same time transmits the temporal key point coordinate sequence to the fatigue feature extraction module. The feature calculation module performs adaptive weighted filtering on the temporal key point coordinate sequence based on the confidence value, and calculates the spatiotemporal biomechanical feature vector including joint angle, joint angular velocity, joint torque and cumulative joint load based on the filtered temporal key point coordinate sequence and human skeletal topology. The spatiotemporal biomechanical feature vector is then transmitted to the fatigue feature extraction module and the risk assessment module, respectively. The fatigue feature extraction module takes the temporal key point coordinate sequence and the spatiotemporal biomechanical feature vector as dual inputs, and extracts the motion fatigue feature vector along the time dimension through a temporal convolutional network including channel attention mechanism and temporal attention mechanism and transmits it to the risk assessment module. The risk assessment module uses spatiotemporal biomechanical feature vectors as query inputs and motion fatigue feature vectors as key inputs. It generates a fusion feature vector that couples biomechanics and fatigue through a cross-attention mechanism, and outputs the probability distribution corresponding to each damage risk level after performing nonlinear transformation and normalization classification processing on the fusion feature vector. The baseline calibration module obtains a baseline biomechanical feature vector based on a baseline motion image sequence collected when the target person performs similar motion actions under a fatigue-free baseline state. It calculates the deviation measure of the current motion's spatiotemporal biomechanical feature vector relative to the baseline biomechanical feature vector, maps the deviation measure to a calibration weight coefficient, and uses the calibration weight coefficient to weight and correct the probability distribution corresponding to each damage risk level output by the risk assessment module.

2. The image analysis-based sports injury prevention system according to claim 1, characterized in that, The image acquisition module includes three or more image acquisition devices, which are respectively deployed at different spatial locations on the sports field. These devices acquire image frames simultaneously via a hardware synchronization triggering mechanism, outputting three or more image frames from different spatial locations at each moment to form a multi-view synchronized image sequence. The pose estimation module includes a 3D keypoint coordinate reconstruction module. This module performs triangulation calculations on the 2D pixel coordinates of human keypoints extracted from the image frames based on the camera intrinsic and extrinsic parameter matrices of the image acquisition devices, converting the 2D pixel coordinates into 3D spatial coordinates. It then establishes a world coordinate system with the sports field ground as the reference plane and outputs the 3D coordinate values ​​of each human keypoint in the world coordinate system at each moment. The 3D keypoint coordinate reconstruction module also includes a view quality weighted reconstruction module. This module performs a weighted geometric average calculation on the detection confidence value of each human keypoint under each visible image acquisition device, the view cosine value between the camera optical axis and the line connecting the keypoint to the camera center, and the baseline depth ratio to obtain a view quality factor. The calculation formula is: ； in, For the first The confidence value of the detection of key points on the human body by an image acquisition device. For the first The angle between the optical axis of the camera of the image acquisition device and the line connecting the key point to the center of the camera. For the first The stereo baseline length between an image acquisition device and its adjacent image acquisition devices. For the key points of this human body to the first The depth distance of the camera center of each image acquisition device; The viewpoint quality weighted reconstruction module selects the two target viewpoints with the highest viewpoint quality factors from all visible image acquisition devices of the human body key point, performs triangulation calculation on the two-dimensional pixel coordinates of the two target viewpoints, and performs weighted fusion processing on the triangulation results using the viewpoint quality factors of the two target viewpoints as weighting coefficients to generate the optimized three-dimensional coordinate values ​​of the human body key point.

3. The image analysis-based sports injury prevention system according to claim 2, characterized in that, The pose estimation module further includes a feature extraction backbone network, a feature pyramid network, and a keypoint regression head. The feature extraction backbone network performs convolution operations on each frame of the multi-view synchronized image sequence and extracts multi-scale feature maps. The feature pyramid network performs top-down feature fusion processing on the multi-scale feature maps and outputs semantically enhanced feature maps. The keypoint regression head outputs the two-dimensional coordinate values ​​and corresponding confidence values ​​of each visible human keypoint based on the semantically enhanced feature maps. The pose estimation module compares the confidence values ​​with a confidence threshold determined based on the statistical accuracy of keypoint detection in the training set and retains valid keypoints with confidence values ​​higher than the confidence threshold. It performs inter-frame matching and temporal smoothing processing on valid keypoints in consecutive frames to generate an initial temporal keypoint coordinate sequence. The pose estimation module also includes a biomechanical constraint verification module, which receives the initial temporal keypoint coordinate sequence. The process calculates the joint angle change rate of each joint between adjacent frames based on the keypoint coordinates of adjacent frames. It then compares the joint angle change rate of each joint with the upper bound of the corresponding angle change rate constraint in the human joint physiological motion parameter set. Simultaneously, it compares the joint angle of each frame with the corresponding angle range boundary in the human joint physiological motion parameter set. When the joint angle change rate of a joint in a frame exceeds the upper bound of the angle change rate constraint or the joint angle exceeds the angle range boundary, the joint angle of that joint in that frame is projected onto the corresponding constraint boundary value. Simultaneously, the 3D coordinates of the corresponding keypoint in that frame are corrected, and the confidence value of the corresponding keypoint in that frame is lowered according to a preset attenuation ratio. The initial temporal keypoint coordinate sequence and the updated confidence value after constraint verification are output. These are then used as the temporal keypoint coordinate sequence and confidence value output.

4. The image analysis-based sports injury prevention system according to claim 3, characterized in that, The feature calculation module includes an adaptive weighted filtering module, a joint angle calculation module, an angular velocity calculation module, and a torque calculation module. The adaptive weighted filtering module constructs an adaptive weight matrix based on the confidence value of each human keypoint output by the keypoint regression head. This adaptive weight matrix is ​​then used to perform weighted filtering on the 3D coordinates of each human keypoint in the temporal keypoint coordinate sequence, and the filtered temporal keypoint coordinate sequence is output. The values ​​of the diagonal elements in the adaptive weight matrix are linearly proportional to the confidence value. The calculation method is as follows: ; in For the first Confidence values ​​of key points on an individual's body. The scaling factor is used to map the confidence score to the weight range, and is set as the ratio of the upper bound of the weight range to the maximum confidence score. The joint angle calculation module calculates the joint angle based on the direction vectors of adjacent bone segments in three-dimensional space in the filtered temporal keypoint coordinate sequence, and the angular velocity calculation module calculates the joint angular velocity based on the ratio of the joint angle difference between adjacent frames to the corresponding time interval. The joint angular velocity The calculation formula is: ； in, The discrete frame number of the current frame. Frame number Corresponding joint angles Frame number The corresponding joint angle in the previous frame, The sampling time interval between adjacent frames; The torque calculation module calculates joint torques using a bottom-up inverse dynamics recursive method based on human segmental inertial parameters and kinematic data derived from filtered temporal key point coordinate sequences. The human segmental inertial parameters include segmental mass, segmental center of mass position, and segmental moment of inertia. The inverse dynamics recursive method, based on Newton's laws of motion, recursively calculates torques segment by segment from the distal limb towards the proximal end. For the ankle joint, the joint torque... The calculation formula is: ； in, For foot segment quality, It is the acceleration due to gravity. Let be the linear acceleration of the foot segment's center of mass in the sagittal plane along the vertical direction. Let be the horizontal coordinate of the foot segment's centroid in the sagittal plane. The horizontal coordinate of the ankle joint center in the sagittal plane. The horizontal lever arm of the foot segment's center of mass relative to the ankle joint center is given. The vertical acceleration of the foot segment's center of mass is obtained by performing a second-order central difference calculation on the vertical coordinates of the foot segment's center of mass in consecutive frames. For the knee and hip joints, the corresponding joint torques are calculated sequentially along the limb chain towards the proximal end based on the topology of the human skeleton.

5. The image analysis-based sports injury prevention system according to claim 4, characterized in that, The feature calculation module further includes an uncertainty propagation module, a joint load accumulation module, and a feature aggregation module. The uncertainty propagation module calculates the coordinate variance estimate based on the confidence value of each key point on the human body, and propagates the coordinate variance estimate segment by segment along the kinematic chain defined by the human skeletal topology from distal to proximal segments. For each joint angle, based on the coordinate variance estimates of the key points corresponding to the two adjacent skeletal segments constituting that joint, the joint angle variance is calculated using the error propagation formula. The calculation formula is: ； in, For the first part that constitutes this joint The coordinate components of each key point in three-dimensional space This is the estimated coordinate variance for that coordinate component, and the estimated coordinate variance is inversely proportional to the square of the confidence level. The partial derivative of the joint angle with respect to the coordinate components of its corresponding key point is obtained by differentiating the analytical calculation expression of the joint angle with respect to the coordinate components. For each joint moment, based on the inverse dynamic recursive relationship of the moment calculation module, the partial derivatives of the joint moment with respect to the joint angle, the partial derivatives of the joint moment with respect to the segmental inertia parameter, and the partial derivatives of the joint moment with respect to the distal joint moment are substituted into the moment error propagation formula to calculate the uncertainty boundary of the joint moment. The moment error propagation formula is as follows: in, For the first Variance of joint torque of each joint Traversal influences the first All joint angles of a joint moment For the first Each joint angle For the first The variance of each joint angle, Traversing the first The inertial parameters of all segments of the joint and its associated segments. For the first Segment inertial parameters, For the first The variance of the inertial parameters of each segment, The traversal is related to the inverse dynamical recurrence relation with the first The distal joints associated with each joint For the first Joint torque of each distal joint, For the first Variance of joint moment of each distal joint For the first The joint torque is related to the first The partial derivatives of the torques of each distal joint; the joint load accumulation module performs cumulative integration calculations along the time dimension on the joint torques of each frame within the same sliding time window to obtain the cumulative load value of each joint. The calculation formula is: ； in, For the first time within the sliding time window The first frame Joint torque of each joint The total number of frames within the sliding time window. The sampling time interval between adjacent frames is defined as follows: The feature aggregation module combines the joint angle, joint angular velocity, joint torque, joint torque uncertainty boundary, and cumulative joint load value corresponding to consecutive frames within the same sliding time window into a spatiotemporal biomechanical feature vector.

6. The image analysis-based sports injury prevention system according to claim 5, characterized in that, The fatigue feature extraction module includes a temporal convolutional network structure, which comprises stacked one-dimensional causal convolutional layers and residual connections, a channel attention mechanism layer, a fatigue feedback gating mechanism layer, a temporal attention mechanism layer, a global average pooling layer, and a fully connected mapping layer. The fatigue feature extraction module concatenates the temporal keypoint coordinate sequence with the spatiotemporal biomechanical feature vector along the feature channel dimension to generate a dual-path fusion feature matrix. The one-dimensional causal convolutional layer performs causal convolution operations on the dual-path fusion feature matrix along the time dimension to extract temporal dependent features including changes in motion rhythm and the cumulative trend of joint load. The residual connection adds the input and output features of the one-dimensional causal convolutional layer element-wise. The channel attention mechanism layer processes the output of the residual connection... In the feature matrix, a channel description vector is obtained by calculating the global average pooling value for each feature channel. This channel description vector is then input into a two-layer fully connected network (including a first and a second fully connected layer) to generate a channel attention weight vector. The feature matrix is ​​then weighted channel-wise using this channel attention weight vector. The fatigue feedback gating mechanism layer performs global average pooling along the time dimension on the feature matrix after channel attention weighting to generate a fatigue state description vector. This fatigue state description vector is then input into a fully connected layer activated by the Sigmoid activation function to generate a fatigue gating weight vector. The feature matrix after channel attention weighting is then subjected to element-wise multiplication using this fatigue gating weight vector to obtain a gated modulation feature matrix. The calculation formula for the gated modulation is as follows: ； in, For the gated modulation feature matrix, The feature matrix after channel attention weighting is... This is an element-wise multiplication operation. It is the Sigmoid activation function. and These are the weight matrix and bias vector of the fully connected layer, respectively. The process involves a global average pooling operation along the time dimension. The temporal attention mechanism layer performs an average pooling operation on the gated modulation feature matrix along the feature channel dimension to obtain a time description vector. This time description vector is then input into a two-layer fully connected network, including a third and a fourth fully connected layer, to generate a temporal attention weight vector. This temporal attention weight vector is used to perform weighted processing on each time step of the gated modulation feature matrix. The global average pooling layer performs a global average pooling operation on the feature matrix after temporal attention weighting to obtain a fixed-dimensional temporal feature vector. The fully connected mapping layer performs a nonlinear mapping on the temporal feature vector to output the motion fatigue feature vector.

7. The image analysis-based sports injury prevention system according to claim 5, characterized in that, The risk assessment module includes a cascaded bidirectional cross-attention mechanism layer and a multi-layer feedforward neural network layer. The multi-layer feedforward neural network layer includes residual connections and layer normalization. The cascaded bidirectional cross-attention mechanism layer includes a first-stage cross-attention layer and a second-stage cross-attention layer. The first-stage cross-attention layer uses the spatiotemporal biomechanical feature vector as the query vector and the motion fatigue feature vector as the key vector and value vector, and performs the first-stage cross-attention calculation to generate a global fusion feature vector according to the following formula: ； in, For global fusion feature vectors, The query vector is constructed from spatiotemporal biomechanical feature vectors. The key vector is constructed from the eigenvectors of motion fatigue. This is a value vector constructed from the eigenvectors of motion fatigue. To determine the dimensions of the query vector and the key vector, This represents the matrix transpose operation; the second-stage cross-attention layer uses the motion fatigue feature vector after residual connection processing as the query vector, and the spatiotemporal biomechanical feature vector as the key vector and value vector, to perform the second-stage cross-attention calculation to generate a local fusion feature vector. The formula for the second-stage cross-attention calculation is as follows: ； in, For local fusion feature vectors, The query vector is constructed from the motion fatigue feature vectors after residual join processing. The key vector is constructed from spatiotemporal biomechanical eigenvectors. The value vector is constructed from spatiotemporal biomechanical feature vectors; the cascaded bidirectional cross-attention mechanism layer concatenates the global fusion feature vector and the local fusion feature vector along the feature dimension to generate a bidirectional fusion feature vector; the multi-layer feedforward neural network layer performs nonlinear feature transformation on the bidirectional fusion feature vector and outputs the probability distribution corresponding to each damage risk level after Softmax normalization.

8. The image analysis-based sports injury prevention system according to claim 5, characterized in that, The baseline calibration module obtains a baseline biomechanical feature vector based on a sequence of baseline motion images collected from a target person performing similar movements under fatigue-free baseline conditions. It calculates the mean vector and covariance matrix of the baseline biomechanical feature vector, and calculates the Mahalanobis distance between the spatiotemporal biomechanical feature vector and the mean vector during the target person's current movement as a measure of point-state deviation. The calculation formula is: ； in, This represents the current spatiotemporal biomechanical feature vector. The mean vector of the baseline biomechanical eigenvectors. The covariance matrix of the baseline biomechanical eigenvectors. It is the inverse of the covariance matrix. This represents a matrix transpose operation; the baseline calibration module also constructs a current biomechanical trajectory vector based on the spatiotemporal biomechanical feature vectors corresponding to consecutive frames within the sliding time window, extracts a baseline biomechanical trajectory vector from the baseline motion image sequence, and calculates the dynamic time warping distance between the current biomechanical trajectory vector and the baseline biomechanical trajectory vector as a trajectory-level deviation metric. The calculation method involves searching for the optimal alignment path on the cost matrix using a dynamic programming algorithm. The cost matrix contains the [missing information - likely a specific path or element]. Line 1 The element of the column is the th element in the current biomechanical trajectory vector. The spatiotemporal biomechanical eigenvector at time t and the baseline biomechanical trajectory vector at time t. The Euclidean distance between the baseline biomechanical eigenvectors at each time step; The baseline calibration module also calculates a periodic exponential deviation metric based on the spatiotemporal biomechanical feature vector within the sliding time window. It calculates the periodic exponent for both the current trajectory and the baseline trajectory. The periodic exponent is calculated by taking the Euclidean norm of the feature vector at each moment in the time series of the spatiotemporal biomechanical feature vector within the sliding time window to obtain a scalar time series. A normalized autocorrelation function is then calculated on the scalar time series. Finally, each effective peak value (excluding zero delay) in the normalized autocorrelation function is summed after exponential decay weighting according to the delay order corresponding to the peak value. The effective peak value is a local maximum point where the normalized autocorrelation function value exceeds a peak threshold. The periodic exponent... The calculation formula is: ； in, For the delay order, This represents the total number of valid peak values ​​detected within the sliding time window. Delay order The normalized autocorrelation function value at each point is included in the summation calculation only if the value is greater than the peak threshold. The normalized autocorrelation function value is calculated by subtracting the time mean from the scalar time series, calculating the autocorrelation function, and then dividing by the autocorrelation value at zero delay. To control the decay coefficient of the peak weight decay rate at long distances; the periodicity exponent of the current motion trajectory is... Periodicity index of the reference motion trajectory The absolute value of the difference is used as a measure of periodic deviation. Mahalanobis distance Dynamic time warping distance and periodic deviation measurement The comprehensive deviation metric is obtained by weighting and summing the weight ratios determined through cross-search of the validation set. The comprehensive deviation metric is then mapped to a calibration weight coefficient, which is used to weight and correct the probability distribution corresponding to each damage risk level output by the risk assessment module.

Images