An artificial intelligence-based motion analysis system

Abstract

The application relates to the technical field of artificial intelligence, and discloses a motion analysis system based on artificial intelligence, which covers a multi-source data acquisition module, a space-time preprocessing module, a feature fusion module, a motion mode recognition module, an anomaly detection module and a feedback generation module. The multi-source data acquisition module synchronously acquires data at a multi-frequency sampling rate by means of an inertial measurement unit, an optical sensor and a depth camera array, generates an original motion data set and stores the original motion data set to a distributed edge node; the space-time preprocessing module calibrates and denoises by using a multi-scale dynamic time warping algorithm and an adaptive Kalman filter; the feature fusion module extracts and fuses key space-time features; the motion mode recognition module and the anomaly detection module accurately distinguish the motion mode and locate the anomaly according to the fused feature tensor; and the feedback generation module generates an individualized executable instruction set according to the above results, thereby providing strong support for tapping the potential of athletes and breaking through bottlenecks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a sports analysis system based on artificial intelligence. BACKGROUND

[0002] In the field of sports competition, the competition of events is becoming more and more fierce, and the athletes' demand for the refinement of their own training has reached an unprecedented height. For example, in the short track and field project, every step of the sprinter in the 100-meter sprint, such as the force of the foot, the amplitude and frequency of the arm swing, and the angle of the body tilt, all subtly and accurately affect the final competition results. The traditional method of relying on the on-site naked eye observation of the coach and the rough analysis of the video after the competition cannot accurately capture these fleeting key sports details. On the one hand, the coach cannot simultaneously pay attention to the motion performance of multiple key parts of the athlete's body at the same time; on the other hand, simple video playback lacks the ability of deep data mining and cannot quantitatively analyze the advantages and disadvantages of various motion parameters. Moreover, different athletes have unique physical conditions and sports habits, and the general training guidance mode cannot meet the needs of personalized improvement. This urgently needs a system that can collect athlete's sports data in all directions with high precision, intelligently analyze sports patterns, and then provide exclusive optimization suggestions to help athletes tap their potential and break through the performance bottleneck. SUMMARY

[0003] The present application aims to provide a sports analysis system based on artificial intelligence to solve the problems raised in the background.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a sports analysis system based on artificial intelligence, the system comprising:

[0005] comprising a multi-source data acquisition module, a space-time preprocessing module, a feature fusion module, a motion pattern recognition module, an anomaly detection module and a feedback generation module;

[0006] The multi-source data acquisition module is used to acquire the original multi-modal data required for sports analysis, specifically by deploying an inertial measurement unit, an optical sensor and a depth camera array to perform synchronous data acquisition, generate an original sports data set, and transmit the original sports data set to the space-time preprocessing module;

[0007] The synchronous data acquisition specifically refers to collecting motion sequence data at different sampling rates, including multi-band data at 30 frames / second, 60 frames / second and 120 frames / second, and generating space-time synchronous data through dynamic timestamp alignment and storing it in a distributed edge node;

[0008] The spatio-temporal preprocessing module is configured to perform spatio-temporal calibration and noise reduction processing on the original data, specifically, aligning heterogeneous sensor data by a multi-scale dynamic time warping algorithm, and eliminating motion noise by an adaptive Kalman filter to generate optimized motion data, which is transmitted to the feature fusion module.

[0009] The feature fusion module is configured to extract and fuse spatio-temporal features, specifically, extracting bone joint trajectory features, motion speed distribution features and posture angle change features by a three-dimensional residual network optimized by a spatio-temporal convolution-based attention mechanism, to generate a fusion feature tensor, which is input to the motion pattern recognition module and the anomaly detection module.

[0010] The motion pattern recognition module is configured to classify and recognize motion patterns, specifically, using a multi-task graph convolution network joint generative adversarial training model to output motion class probability distribution and key action label based on the fusion feature tensor.

[0011] The anomaly detection module is configured to detect abnormal behavior in the motion process, specifically, using a hybrid model based on a variational autoencoder and an isolation forest algorithm to perform reconstruction error analysis and distribution shift detection on the fusion feature tensor to generate an anomaly score and positioning result.

[0012] The feedback generation module is configured to generate motion optimization suggestions and abnormal processing strategies, specifically, generating an individualized feedback scheme by a decision tree model driven by reinforcement learning based on the motion pattern recognition result and the anomaly detection data, and encoding it into an executable instruction set.

[0013] Preferably, the original multi-modal data includes bone joint coordinate data, accelerometer data, gyroscope angular velocity data, depth image sequence and RGB video stream data; the spatio-temporal synchronization data further contains sensor calibration parameters and ambient light compensation information.

[0014] Preferably, the multi-scale dynamic time warping algorithm specifically includes a time series slicing operator, a local similarity measurement matrix construction operator and a global path optimization operator.

[0015] The time series slicing operator is configured to divide heterogeneous sensor data into segments by windowing, and calculate the similarity between segments by dynamic programming.

[0016] The local similarity measurement matrix construction operator specifically generates a multi-modal alignment cost matrix by combining dynamic time warping distance with cosine similarity.

[0017] The global path optimization operator specifically solves the minimum cumulative cost path based on a branch and bound algorithm to achieve frame-level alignment of cross-sensor data.

[0018] Preferably, the three-dimensional residual network comprises a parallel spatio-temporal convolution branch, wherein the first branch extracts long-term temporal dependence features using dilated convolution, the second branch optimizes local motion feature weights through a channel attention mechanism, and the third branch captures multi-scale spatial context information using a dilated spatial pyramid pooling.

[0019] Preferably, the multi-task graph convolution network is jointly trained by an adversarial generation model, specifically including a skeleton topology graph construction module, a dynamic adjacency matrix learning module, an adversarial discriminator sub-network, and a multi-head classifier.

[0020] The skeleton topology graph construction module is configured to construct an initial joint connection graph according to human anatomy prior knowledge, and dynamically adjust edge weights through learnable parameters.

[0021] The dynamic adjacency matrix learning module is configured to iteratively update the relationship between nodes using a gated graph attention mechanism to generate a spatio-temporal enhanced graph representation.

[0022] The adversarial discriminator sub-network is configured to constrain the manifold consistency of feature distribution and real motion data.

[0023] The multi-head classifier is configured to output motion category probability, action phase label, and energy consumption estimation value in parallel.

[0024] Preferably, the workflow of the anomaly detection module includes:

[0025] The feature distribution of normal motion patterns is modeled by a variational autoencoder, and the reconstruction error of the input data is calculated.

[0026] The high-dimensional feature space is recursively partitioned using an isolation forest algorithm to detect abnormal subspaces deviating from the mainstream distribution.

[0027] The reconstruction error probability and subspace anomaly degree are fused to generate a comprehensive anomaly score and associated joint index.

[0028] Preferably, the reinforcement learning driven decision tree model specifically includes a state encoder, a policy search tree, and a reward function optimizer.

[0029] The state encoder is configured to map the motion pattern recognition result and anomaly detection data into a low-dimensional state vector.

[0030] The policy search tree is configured to explore the action space through Monte Carlo tree search to generate a candidate feedback path.

[0031] The reward function optimizer dynamically adjusts the policy weight according to the motion efficiency improvement degree, anomaly correction speed, and physiological load.

[0032] Preferably, the personalized feedback scheme includes motion posture correction parameters, training intensity adjustment suggestions, abnormal motion avoidance strategies and real-time biomechanical load warnings.

[0033] Preferably, the executable instruction set is encoded by the edge computing node into at least one of the following forms: a tactile feedback pulse sequence, an augmented reality visualization mark, a mechanical exoskeleton driving signal and a voice prompt instruction stream.

[0034] Compared with the prior art, the present application has the following beneficial effects:

[0035] The multi-source data acquisition module uses an inertial measurement unit, an optical sensor and a depth camera array to perform synchronous data acquisition, can capture all kinds of information of the athlete during the movement process in all directions, and is no longer limited to the one-sidedness of traditional coach naked-eye observation. Whether it is the force of each step of a 100-meter sprinter, the details of arm swinging, or body posture changes, they can be accurately recorded in the form of original motion data sets, without missing any key moment, providing a solid foundation for subsequent accurate analysis.

[0036] The movement sequence data is collected at different sampling rates (multi-band data at 30 frames / second, 60 frames / second and 120 frames / second), and spatiotemporal synchronization data is generated by dynamic timestamp alignment, stored in a distributed edge node, ensuring the integrity and accuracy of the data, meeting the needs of fine analysis of high-speed sports projects, and effectively solving the problem that traditional methods are difficult to capture fleeting key movement details.

[0037] The spatiotemporal preprocessing module uses a multi-scale dynamic time warping algorithm to align heterogeneous sensor data, so that information from different types of sensors can be accurately matched in the spatiotemporal dimension, avoiding data confusion and deviation. At the same time, adaptive Kalman filtering is used to eliminate motion noise, making the optimized motion data more pure, eliminating interference for subsequent feature extraction and analysis, greatly improving the usability and reliability of the data, and overcoming the disadvantages of simple video playback lacking depth data mining ability.

[0038] The feature fusion module uses a three-dimensional residual network optimized by a spatiotemporal convolution-based attention mechanism to extract skeletal joint trajectory features, motion speed distribution features and posture angle change features, respectively, and generate a fusion feature tensor. This deep fusion method can organically integrate multiple key dimension information of the athlete during the movement process, mine the deep movement pattern correlation hidden behind the data, and accurately grasp the movement state of the athlete, providing rich and targeted basis for personalized training guidance.

[0039] The motion mode recognition module outputs the motion category probability distribution and the key action label by using the multi-task graph convolution network joint adversarial generation training model according to the fusion feature tensor. This means that the system can not only accurately determine the motion mode category of the athlete, but also accurately locate the key action. Whether it is the start-up, acceleration, sprinting and other stages in short-distance running, or specific action links in other complex sports projects, they can be clearly identified, providing clear improvement direction for coaches and athletes, and changing the current situation that the general training guidance mode cannot meet the individual improvement needs.

[0040] The anomaly detection module generates the anomaly score and positioning result by analyzing the reconstruction error and distribution deviation of the fusion feature tensor based on the hybrid model of the variational autoencoder and the isolation forest algorithm. During the training or competition of the athlete, abnormal behaviors such as action deformation, force abnormality and rhythm disorder can be found in time. Whether the problem is caused by fatigue, injury or technical action failure, it can be detected at the first time to avoid the continuous deterioration of the abnormal situation and affect the athlete's performance. The feedback generation module generates an individual feedback scheme through the decision tree model driven by reinforcement learning according to the motion mode recognition result and the anomaly detection data, and encodes it into an executable instruction set. This enables each athlete to obtain exclusive training optimization suggestions according to their unique physical conditions and exercise habits, such as providing precise adjustment schemes for short-distance runners with insufficient arm swing amplitude and uneven foot lifting force, helping athletes to tap their potential and break through the performance bottleneck. BRIEF DESCRIPTION OF DRAWINGS

[0041] Fig. 1 The working principle diagram of the motion analysis system described in the present application;

[0042] Fig. 2 The flowchart for multi-scale alignment of heterogeneous sensor data;

[0043] Fig. 3 The flowchart for motion mode recognition. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] Please refer to Figs. 1-3 The present application provides a technical solution: a motion analysis system based on artificial intelligence, which comprises:

[0046] Multi-source data acquisition module: To obtain the original multi-modal data required for motion analysis, the system deploys inertial measurement units, optical sensors, and depth camera arrays. These devices collect motion sequence data at different sampling rates, covering multiple frequency bands of 30 frames per second, 60 frames per second, and 120 frames per second. During acquisition, time-synchronous data is generated through dynamic timestamp alignment technology and stored in distributed edge nodes. The original multi-modal data includes skeletal joint coordinate data, accelerometer data, gyroscope angular velocity data, depth image sequences, and RGB video stream data. Time-synchronous data further includes sensor calibration parameters and ambient light compensation information to ensure data accuracy and integrity.

[0047] Temporal and spatial preprocessing module: This module is responsible for temporal and spatial calibration and noise reduction of the collected raw data. A multi-scale dynamic time warping algorithm is used to align heterogeneous sensor data. This algorithm includes a time series slicing operator, a local similarity metric matrix construction operator, and a global path optimization operator. At the same time, an adaptive Kalman filter algorithm is used to eliminate motion noise, generating optimized motion data and transmitting it to the feature fusion module.

[0048] Feature fusion module: This module uses a three-dimensional residual network optimized by a spatiotemporal convolution-based attention mechanism to extract skeletal joint trajectory features, motion velocity distribution features, and posture angle change features, generating a fusion feature tensor. The three-dimensional residual network includes parallel spatiotemporal convolution branches. The first branch uses dilated convolution to extract long-term temporal dependency features, the second branch optimizes local motion feature weights through channel attention mechanism, and the third branch uses a dilated spatial pyramid pooling to capture multi-scale spatial context information. The fusion feature tensor is then input into the motion pattern recognition module and the anomaly detection module.

[0049] Motion pattern recognition module: Based on the fusion feature tensor generated by the feature fusion module, this module uses a multi-task graph convolution network joint adversarial generation training model to classify and recognize motion patterns. This model includes a skeletal topology graph construction module, a dynamic adjacency matrix learning module, an adversarial discriminator subnetwork, and a multi-head classifier. Finally, it outputs motion class probability distribution and key action labels.

[0050] Anomaly detection module: Based on a hybrid model of variational autoencoder and isolation forest algorithm, this module analyzes the reconstruction error and distribution deviation of the fusion feature tensor. The specific workflow is as follows: first, the variational autoencoder models the feature distribution of normal motion patterns, calculates the reconstruction error of the input data; then, the isolation forest algorithm is used to recursively partition the high-dimensional feature space to detect abnormal subspaces deviating from the mainstream distribution; finally, the reconstruction error probability and subspace anomaly degree are fused to generate a comprehensive anomaly score and associated joint index.

[0051] The feedback generation module generates a personalized feedback scheme based on the motion pattern recognition result and the abnormality detection data through a reinforcement learning driven decision tree model. This model includes a state encoder, a policy search tree, and a reward function optimizer. The personalized feedback scheme includes motion posture correction parameters, training intensity adjustment suggestions, abnormal motion avoidance strategies, and real-time biomechanical load warnings. The feedback generation module encodes the personalized feedback scheme into an executable instruction set, which can be encoded by an edge computing node into at least one of a tactile feedback pulse sequence, an augmented reality visualization label, a mechanical exoskeleton driving signal, and a voice prompt instruction stream.

[0052] The application will be further described in conjunction with Examples 1 to 5:

[0053] Example 1:

[0054] This example details the specific working process of the multi-source data acquisition module, as well as the characteristics of the original multi-modal data and the spatio-temporal synchronous data, providing a foundation for subsequent data analysis and processing, ensuring that the collected data is accurate, comprehensive, and meets system processing requirements.

[0055] In actual application scenarios, taking athlete training monitoring as an example, the inertial measurement units of the multi-source data acquisition module are installed on key parts of the athlete's body, such as wrists, ankles, knees, and waists. These inertial measurement units are built-in accelerometers and gyroscopes. The accelerometer is used to collect acceleration data of each part of the athlete during exercise, its principle is based on Newton's second law F = ma, by measuring the inertial force F to calculate the acceleration a, the unit is m / s 2 . The gyroscope is used to measure angular velocity data, its measurement principle is based on the law of conservation of angular momentum, the output angular velocity data unit is rad / s.

[0056] Optical sensors are arranged around the training field, which capture the athlete's skeletal joint coordinate data using optical imaging principles. The depth camera array collects depth image sequences and RGB video stream data from different angles. The depth image sequence can accurately obtain the distance information between each part of the athlete's body and the camera, and the RGB video stream data provides rich color and texture information for more comprehensive analysis of the athlete's movement posture.

[0057] During data acquisition, due to differences in the operating frequencies and performance of different sensors, motion sequence data needs to be acquired at different sampling rates. Taking multi-band data acquisition at 30 frames per second, 60 frames per second, and 120 frames per second as an example, low sampling rate (30 frames per second) data is suitable for capturing the general trend of motion, while high sampling rate (120 frames per second) data can accurately record the details of rapid motion changes. Dynamic timestamp alignment technology adds precise time stamps to each data sample, achieving spatiotemporal synchronization of data at different sampling rates. During this process, the synchronized data also records sensor calibration parameters, which are used to calibrate sensor measurement errors and ensure data accuracy. Simultaneously, ambient lighting compensation information is also recorded to eliminate the impact of ambient lighting changes on data acquisition, such as ensuring the stability of depth image and RGB video stream data quality when the brightness of lights in the training venue changes.

[0058] Example 2:

[0059] This embodiment focuses on the multi-scale dynamic time warping algorithm in the spatiotemporal preprocessing module, and details the working principle and operation process of each operator to achieve effective alignment of heterogeneous sensor data, laying the foundation for subsequent accurate feature extraction and analysis.

[0060] The multi-scale dynamic time warping algorithm in the spatiotemporal preprocessing module is crucial for aligning heterogeneous sensor data. When processing heterogeneous data from inertial measurement units, optical sensors, and depth camera arrays, a time-series slicing operator is first applied. Assume the heterogeneous sensor data sequence is X = {x1, x2, ..., x...}. m} and Y = {y1, y2, ..., y n The time-series slicing operator divides the data into segments according to a window. Let the window size be w, then the segment divided from the data sequence X is Xslice. i ={x i ,x i+1 ,…,x i+w-1 Similarly, from Y, we can divide Y into segments (i = 1, 2, ..., m - w + 1). j The similarity between segments is calculated using dynamic programming. The core idea of ​​dynamic programming is to decompose a complex problem into multiple subproblems and store the solutions to these subproblems to avoid redundant computation. Calculate the similarity between segment X. i and Y j Similarity d(X) between i ,Y j The formula for ) is:

[0061]

[0062] Here, ‖·‖ represents a certain distance metric, such as Euclidean distance.

[0063] Next, the local similarity measure matrix construction operator works. The operator generates a multi-modal alignment cost matrix by combining the dynamic time warping distance with the cosine similarity. The dynamic time warping distance (DTW distance) is used to measure the similarity of two time series in time. The calculation process is to find an optimal time warping path, so that the sum of the distances of the two time series on this path is the smallest. Let two time series A = {a1, a2, …, a p} and B = {b1, b2, …, b q}, the calculation of the DTW distance DTW(A, B) is based on the recursive formula:

[0064]

[0065] where A p-1 = {a1, a2, …, a p-1}, B q-1 is equivalent. At the same time, combined with the cosine similarity cosine(A, B), its formula is:

[0066]

[0067] The two are integrated to generate a multi-modal alignment cost matrix C, C ij represents the alignment cost between the segments X i and Y j .

[0068] Finally, the global path optimization operator solves the minimum cumulative cost path based on the branch and bound algorithm. The branch and bound algorithm is an algorithm for searching the optimal solution on the solution space tree. It divides the solution space (branch) and uses the bound function to cut off the subtree that cannot contain the optimal solution (bound), so as to improve the search efficiency. In this algorithm, by solving the minimum cumulative cost path, the frame-level alignment of cross-sensor data is realized, so that the heterogeneous sensor data can be accurately matched in time and space, providing a reliable data basis for subsequent data analysis.

[0069] Embodiment 3:

[0070] This embodiment details the structure and working principle of the three-dimensional residual network in the feature fusion module, and shows how it effectively extracts the skeleton joint trajectory feature, motion speed distribution feature and attitude angle change feature, providing high-quality fusion feature tensor for motion pattern recognition and anomaly detection.

[0071] The core of the feature fusion module is a three-dimensional residual network optimized by a spatio-temporal convolution-based attention mechanism. The network contains parallel spatio-temporal convolution branches, each branch undertaking different feature extraction tasks.

[0072] The first branch extracts long-time sequence dependent features using dilated convolution. The dilated convolution increases the dilation rate parameter based on the ordinary convolution. Assuming that the size of the ordinary convolution kernel is k x k, the convolution kernel of the dilated convolution is expanded in space with a dilation rate d. In the time dimension, for the input time sequence data T = {t1, t2, …, t n}, the dilated convolution operation can be represented as:

[0073]

[0074] where w j is the convolution kernel weight, y i is the convolution output result. By setting a suitable dilation rate, the first branch can capture the motion dependent relationship in a longer time range, for example, when analyzing the movement pattern of a long-distance runner, the step rhythm change features within a few seconds can be captured.

[0075] The second branch optimizes the local motion feature weight through the channel attention mechanism. The channel attention mechanism analyzes the channel dimension of the feature map and adaptively adjusts the weight of each channel. For the input feature map (C is the number of channels, H is the height, and W is the width), first, the feature map is globally averaged and globally maximized in the spatial dimension to obtain two one-dimensional vectors z avg and z max , the calculation is as follows:

[0076]

[0077] Then, the two vectors are respectively passed through a shared multi-layer perceptron (MLP), and the outputs are added to obtain the attention weight vector a. Then, a is weighted with the original feature map F in the channel dimension to optimize the local motion feature weight and highlight the key motion features, such as when analyzing the shooting action of a basketball player, more attention is paid to the movement features of the arm joints.

[0078] The third branch captures multi-scale spatial context information using dilated spatial pyramid pooling. Dilated spatial pyramid pooling samples spatial information at multiple scales through convolution operations with different dilation rates. Assuming that the dilation rates are r1, r2, and r3, for the input feature map F, the feature maps after convolution with different dilation rates are F1, F2, and F3. Then, these feature maps are spliced and convolved to obtain the fused feature map, so that the multi-scale spatial context information from local to global can be captured, for example, when analyzing the action of a dancer, both the subtle action changes of the local body and the overall dance posture and spatial position relationship can be grasped. Finally, the features extracted by the three branches are fused to generate a fused feature tensor, providing rich feature information for subsequent motion pattern recognition and anomaly detection.

[0079] Embodiment 4:

[0080] This embodiment analyzes in detail each component of the multi-task graph convolution network joint adversarial generation training model in the motion pattern recognition module and its working principle, and shows how to accurately classify and recognize motion patterns, output motion category probability distribution and key action label.

[0081] The multi-task graph convolution network joint adversarial generation training model of the motion pattern recognition module is composed of a skeleton topology graph construction module, a dynamic adjacency matrix learning module, an adversarial discriminator subnetwork, and a multi-head classifier.

[0082] The skeleton topology graph construction module constructs an initial joint connection graph according to prior knowledge of human anatomy. There are specific connection relationships between human skeletal joints, such as the shoulder joint connecting the upper arm bone and the clavicle, the elbow joint connecting the upper arm bone and the forearm bone, etc. The initial joint connection graph constructed according to these relationships can be represented as G=(V,E), where V represents the set of joints, and E represents the set of connection edges between joints. The edge weight is dynamically adjusted through learnable parameters, and the edge e ij connecting joints v i and v j , the edge weight w ij can be adjusted through the following formula:

[0083]

[0084] where w is the initial edge weight, and Δw ij is the weight adjustment obtained through learning, so that the connection weight between joints can be adaptively optimized according to different motion data, better reflecting the mutual relationship of joints in the motion process.

[0085] The dynamic adjacency matrix learning module iteratively updates the relationship between nodes using the gated graph attention mechanism. The gated graph attention mechanism determines the influence degree of each node on other nodes by calculating the attention weight of the node. For node v i , the calculation of its attention weight α ij is based on the following formula:

[0086]

[0087] where h i and h j are the feature representations of nodes v i and v j , W1 is a weight matrix, and N i is the number of nodes connected to node v iLeakyReLU is an activation function. Through the gating mechanism, the attention weights are filtered and adjusted, generating a spatio-temporally enhanced graph representation that more accurately captures the dynamic relationship changes of key nodes during the movement process.

[0088] The adversarial discriminator subnetwork is used to constrain the manifold consistency of the feature distribution with the real movement data. During the training process, the generator (the main part of the multi-task graph convolutional network) tries to generate samples as close as possible to the feature distribution of the real movement data, while the discriminator strives to distinguish between the generated samples and the real samples. Through adversarial training, the feature distribution generated by the generator is more similar to the manifold structure of the real movement data, improving the accuracy of movement pattern recognition.

[0089] The multi-head classifier outputs the movement category probability, action phase label, and energy consumption estimate in parallel. For the input fusion feature tensor, different branches of the multi-head classifier perform classification and regression operations. Taking the movement category classification as an example, assuming there are K movement categories, through the linear transformation of the classifier and the Softmax function, the probability distribution P(y=k|x) of each movement category is output:

[0090]

[0091] where W k and b k are the weights and biases of the classifier, and x is the input feature tensor. In this way, comprehensive analysis and recognition of movement patterns are achieved, providing users with rich movement information.

[0092] Example 5:

[0093] This embodiment describes in detail the workflow of the anomaly detection module and the running mechanism of the reinforcement learning-driven decision tree model in the feedback generation module, showing how the system detects movement anomalies and generates personalized feedback plans to ensure the safety and effectiveness of movement.

[0094] The anomaly detection module works based on a hybrid model of variational autoencoder and isolation forest algorithm. First, the variational autoencoder models the feature distribution of normal movement patterns. The variational autoencoder consists of an encoder and a decoder. The encoder maps the input fusion feature tensor x to the latent space z, and the mapping relationship can be represented as z = Encoder(x). It is usually implemented through a neural network, such as a multi-layer perceptron. The decoder reconstructs the output by minimizing the reconstruction error, such as the mean square error (MSE) loss function:

[0095]

[0096] where N is the number of samples, xi and are the original and reconstructed features of the i-th sample, respectively. During the training process, the variational autoencoder learns the feature distribution of normal motion patterns, and will produce a large reconstruction error when inputting abnormal data.

[0097] Then, the high-dimensional feature space is recursively partitioned using the Isolation Forest algorithm. The Isolation Forest algorithm is based on the assumption that points far away from other data points in the data space are more likely to be abnormal points. The algorithm builds multiple isolation trees and, for each data point, calculates its path length in the tree. The shorter the path length, the more isolated the point is, and the more likely it is to be an abnormal point. Let the path length of data point x in isolation tree T be h T (x), the calculation of its anomaly score S(x) is based on the following formula:

[0098]

[0099] where E(h T (x)) is the average path length of data point x in multiple isolation trees, and c(n) is a correction coefficient related to the number of samples n.

[0100] Finally, the reconstruction error probability and subspace anomaly degree are fused to generate a comprehensive anomaly score and the associated joint node index. Assuming that the reconstruction error probability is P recon (x), and the subspace anomaly degree is A(x), the comprehensive anomaly score R(x) can be calculated by weighted summation:

[0101] R(x)=w1P recon (x)+w2A(x)

[0102] where w1 and w2 are weight coefficients that can be determined by experiments or other optimization methods according to actual conditions. Through this comprehensive anomaly score, it can be accurately determined whether the motion data is abnormal and the joint node where the anomaly occurred can be located.

[0103] The feedback generation module generates personalized feedback schemes through a reinforcement learning-driven decision tree model. The model includes a state encoder, a policy search tree, and a reward function optimizer.

[0104] The state encoder maps the motion pattern recognition results and anomaly detection data into a low-dimensional state vector. Assuming that the motion pattern recognition results are M and the anomaly detection data are A, the state encoder converts them into a low-dimensional state vector s through neural networks or other methods:

[0105] s=Encoder state (M,A)

[0106] where Encoder stateThe mapping function of the state encoder. In this way, complex motion data and abnormal information are converted into a form suitable for processing by the decision tree model.

[0107] The policy search tree explores the action space through Monte Carlo tree search to generate candidate feedback paths. Monte Carlo tree search is a heuristic search algorithm based on random simulation, which finds the optimal strategy by constantly simulating the game or decision-making process. In the feedback generation module, the motion scene is regarded as a decision-making process, and each decision point corresponds to different feedback actions. For example, when detecting abnormal motion posture, decision points may include adjusting the angle of a certain joint, changing the motion speed, etc. Monte Carlo tree search calculates the return of each action sequence by simulating different action sequences multiple times, thereby finding the optimal candidate feedback path.

[0108] The reward function optimizer dynamically adjusts the policy weights according to the motion efficiency improvement, abnormal correction speed, and physiological load constraints. Let the motion efficiency improvement be E improve , the abnormal correction speed be V correct , and the physiological load be L physio , the reward function R total can be expressed as:

[0109] R total = w3E improve + w4V correct - w5L physio

[0110] where w3, w4, and w5 are weight coefficients. The reward function optimizer dynamically adjusts these weight coefficients according to actual motion data and targets, so that the generated feedback scheme can effectively improve motion efficiency, quickly correct abnormal actions, and avoid excessive physiological load on athletes.

[0111] The personalized feedback scheme includes motion posture correction parameters, training intensity adjustment suggestions, abnormal action avoidance strategies, and real-time biomechanical load warnings. For example, when detecting abnormal action of knee inward buckling during running, the motion posture correction parameters may include specific values for adjusting the knee angle; the training intensity adjustment suggestion may suggest increasing or decreasing running speed or distance according to the current motion intensity and the athlete's physical condition; the abnormal action avoidance strategy may provide alternative correct action demonstrations or remind the athlete to avoid certain incorrect actions; the real-time biomechanical load warning monitors the force on each part of the athlete's body and issues a warning message when the load exceeds the safety threshold, such as a voice prompt instruction stream "Your knee is under too much stress, please adjust your running posture."

[0112] These personalized feedback schemes are finally encoded into executable instruction sets, such as haptic feedback pulse sequences, augmented reality visualization annotations, mechanical exoskeleton driving signals, and voice prompt instruction streams, by the edge computing nodes. Through the haptic feedback devices installed on the athletes' bodies, the athletes are reminded to adjust their actions with different frequency and intensity of pulses; through the augmented reality devices worn by the athletes, the athletes are shown the adjustment prompts and abnormal action annotations of the movement postures; for the scenarios using mechanical exoskeletons to assist movement, the mechanical exoskeletons are driven according to the feedback schemes to adjust the athletes' movement postures; through the voice prompt instruction streams, the athletes are directly informed of the actions to be taken in simple and clear voice information. Through these executable instruction sets, the athletes can obtain feedback and adjust their movement behaviors in time, thereby improving the movement effect and safety.

[0113] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0114] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based motion analysis system, characterized by, The system comprises a multi-source data acquisition module, a space-time preprocessing module, a feature fusion module, a motion pattern recognition module, an anomaly detection module and a feedback generation module. The multi-source data acquisition module is configured to acquire original multi-modal data required for motion analysis, specifically by deploying an inertial measurement unit, an optical sensor and a depth camera array to perform synchronous data acquisition, generate an original motion data set, and transmit the original motion data set to the space-time preprocessing module. The synchronous data acquisition is specifically performed by collecting motion sequence data at different sampling rates, including multi-band data at 30 frames per second, 60 frames per second and 120 frames per second, and generating space-time synchronous data by dynamic timestamp alignment, and storing the space-time synchronous data to a distributed edge node. The space-time preprocessing module is configured to perform space-time calibration and noise reduction processing on the original data, specifically by aligning heterogeneous sensor data through a multi-scale dynamic time warping algorithm, and eliminating motion noise by using an adaptive Kalman filter to generate optimized motion data, which is transmitted to the feature fusion module. The feature fusion module is configured to extract and fuse space-time features, specifically by using a three-dimensional residual network optimized by a space-time convolution-based attention mechanism to extract bone joint trajectory features, motion speed distribution features and posture angle change features, respectively, to generate a fusion feature tensor, which is input to the motion pattern recognition module and the anomaly detection module. The motion pattern recognition module is configured to classify and identify motion patterns, specifically by using a multi-task graph convolution network joint generative adversarial training model to output motion category probability distribution and key action labels based on the fusion feature tensor. The anomaly detection module is configured to detect abnormal behavior in the motion process, specifically by using a hybrid model based on a variational autoencoder and an isolation forest algorithm to perform reconstruction error analysis and distribution shift detection on the fusion feature tensor to generate an abnormal score and positioning result. The feedback generation module is configured to generate motion optimization suggestions and abnormal processing strategies, specifically by using a reinforcement learning driven decision tree model to generate an individualized feedback scheme based on the motion pattern recognition result and the anomaly detection data, and encoding the feedback scheme into an executable instruction set. 2.The artificial intelligence-based motion analysis system of claim 1, wherein, The original multi-modal data includes bone joint coordinate data, accelerometer data, gyroscope angular velocity data, depth image sequences and RGB video stream data; the space-time synchronous data further includes sensor calibration parameters and ambient light compensation information. 3.The artificial intelligence-based motion analysis system of claim 2, wherein, The multi-scale dynamic time warping algorithm specifically includes a time series slicing operator, a local similarity measurement matrix construction operator and a global path optimization operator. The time series slicing operator is configured to divide heterogeneous sensor data into segments by windowing, and calculate the similarity between segments by dynamic programming. The local similarity measurement matrix construction operator is specifically configured to generate a multi-modal alignment cost matrix by using a dynamic time warping distance combined with a cosine similarity. The global path optimization operator is specifically configured to solve the minimum cumulative cost path based on a branch and bound algorithm to achieve frame-level alignment of cross-sensor data.

4. The artificial intelligence-based motion analysis system of claim 3, wherein, The three-dimensional residual network includes parallel space-time convolution branches, wherein a first branch extracts long-time sequence dependent features by using dilated convolution, a second branch optimizes local motion feature weights by using a channel attention mechanism, and a third branch captures multi-scale spatial context information by using a hollow spatial pyramid pooling. 5.The artificial intelligence-based motion analysis system of claim 4, wherein, The multi-task graph convolution network jointly trains a generative adversarial model, and specifically includes a skeleton topology graph construction module, a dynamic adjacency matrix learning module, an adversarial discriminant sub-network, and a multi-head classifier. The skeleton topology graph construction module is configured to construct an initial joint connection graph according to human anatomy prior knowledge, and dynamically adjust edge weights through learnable parameters. The dynamic adjacency matrix learning module specifically iteratively updates the relationship between nodes by using a gated graph attention mechanism to generate a spatio-temporal enhanced graph representation. The adversarial discriminant sub-network is configured to constrain the manifold consistency of feature distribution and real motion data. The multi-head classifier specifically outputs motion category probability, action stage label, and energy consumption estimation value in parallel. 6.The artificial intelligence-based motion analysis system of claim 5, wherein, The workflow of the anomaly detection module includes: A variational autoencoder is used to model the feature distribution of normal motion patterns, and the reconstruction error of the input data is calculated. An isolation forest algorithm is used to recursively partition the high-dimensional feature space to detect abnormal subspaces deviating from the mainstream distribution. The reconstruction error probability and subspace anomaly degree are fused to generate a comprehensive anomaly score and associated joint index.

7. The artificial intelligence-based motion analysis system of claim 6, wherein, The reinforcement learning driven decision tree model specifically includes a state encoder, a policy search tree, and a reward function optimizer. The state encoder is configured to map the motion pattern recognition result and the anomaly detection data into a low-dimensional state vector. The policy search tree specifically explores the action space by using Monte Carlo tree search to generate a candidate feedback path. The reward function optimizer dynamically adjusts the policy weight according to the motion efficiency improvement degree, the anomaly correction speed, and the physiological load.

8. The artificial intelligence-based motion analysis system of claim 7, wherein, The personalized feedback scheme includes motion posture correction parameters, training intensity adjustment suggestions, abnormal motion avoidance strategies, and real-time biomechanical load warnings. 9.The artificial intelligence-based motion analysis system of claim 8, wherein, The executable instruction set is encoded by the edge computing node into at least one of the following forms: a tactile feedback pulse sequence, an augmented reality visualization label, a mechanical exoskeleton driving signal, and a voice prompt instruction stream.

Images