Motion capture-based force training force detection method and system

Abstract

This invention discloses a method and system for detecting force exertion during strength training based on motion capture, relating to the field of motion force exertion detection technology. The method includes the following steps: determining multiple spatiotemporal connection paths; extracting key feature vectors from a pre-constructed motion knowledge graph based on each spatiotemporal connection path, performing feature concatenation, activating a force exertion determination mechanism; under conditions triggering abnormal force exertion, building a force exertion analysis model, generating the force exertion state of strength training, marking the force conflict location and time interval, and retrieving multi-source motion data from the same time interval to trigger a secondary force exertion determination mechanism; by constructing multiple types of force exertion combinations, analyzing the linear correlation with joint stress intensity, and correcting the force conflict location; finally, generating a strength training detection report. During the motion capture process, through a two-level force exertion determination mechanism, based on different strength training scenarios and training needs, and differentiated weighting rules, accurate identification and rapid response to force conflicts are achieved.

Description

Technical Field

[0001] This invention relates to the field of motion force detection technology, specifically to a method and system for detecting force during strength training based on motion capture. Background Technology

[0002] With the rapid development of big data, motion capture-based strength training detection systems have become an important tool to assist in scientific fitness and rehabilitation training. They are mainly based on motion capture systems and multi-source motion sensing devices. By collecting data such as joint angles and equipment forces during training, they usually compare real-time movements with standard movement templates to determine the compliance of the movements. In the process of strength training, incorrect force exertion, such as muscle compensation, will significantly reduce the training effect.

[0003] Existing motion capture strength training detection technologies have the following limitations: Current force detection technologies focus on the geometric shape of the movement, using skeletal key points as independent coordinates and judging movement anomalies based on static contours. This approach can only determine whether the movement is in place, lacking in-depth analysis of force detection, making it prone to misjudgments and omissions. It cannot accurately identify the source of force anomalies or distinguish whether a force exertion state exists. Furthermore, while some technologies utilize electromyography (EMG) signals to analyze muscle activation states and assist in judging force anomalies, they typically only provide a vague time period or a broad body area as error information after detecting an anomaly, reducing the reliability of detection in complex training. Due to the latency of visual detection, it cannot capture the instant of force conflict, resulting in vague error locations that fail to meet users' actual needs. Especially in scenarios with multiple users monitoring simultaneously, the generated training detection reports lack targeted correction guidance, significantly reducing adaptability and robustness. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a force training force detection method and system based on motion capture. By determining multiple spatiotemporal connection paths, an initial screening of invalid movements is constructed. Under the premise of ensuring valid movements, a fine-tuned two-level force determination mechanism is executed to complete the determination of the force state. Based on the training requirements of strength training, the linear correlation with joint force intensity is analyzed to correct the position of force conflict. The optimal balance between recognition accuracy and determination efficiency is achieved in all scenarios, solving the problems mentioned in the background technology.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, this application provides a force training force detection method based on motion capture, the method comprising: during the motion capture process, determining multiple spatiotemporal connection paths between the starting posture node and the target posture node of the object to be detected; wherein the starting posture node and the target posture node are respectively contained in different motion frames; Based on the spatiotemporal connection paths, key feature vectors are extracted from the pre-constructed motion knowledge graph, feature splicing is performed, the force determination mechanism is activated, and under the condition of triggering abnormal force exertion, a force analysis model is built to generate the force exertion state of strength training, and the force conflict location and time interval are marked in the motion knowledge graph. Based on the location and time interval of the force conflict, multi-source motion data in the same time interval are retrieved to trigger a secondary force determination mechanism. By constructing multiple types of force combinations, the linear correlation with the joint force intensity is analyzed, and the location of the force conflict is corrected. Based on the corrected force conflict location and corresponding force state, combined with the motion knowledge graph, a strength training test report is generated.

[0006] Furthermore, multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected are determined, including: Obtain the coordinate sequence of the skeletal key points of the object to be detected in consecutive action frames, perform single-frame topology construction, and generate a single-frame pose graph by using the set of all skeletal key points in a single frame as pose nodes and adjacent skeletal key points with association as endpoints to construct topological connection edges in a single frame. Using the action frame timing as the axis, the single-frame pose graphs corresponding to all action frames are temporally associated through the chain operation of the same bone key points between adjacent frames. The temporal motion association edges between frames are constructed with the pose nodes of two adjacent frames as endpoints to obtain a dynamic topology graph. Then, pose node locking is performed to determine the starting pose node and the target pose node. A bidirectional breadth-first search algorithm is adopted. In the dynamic topology graph, the starting attitude node is used as the search starting point and the target attitude node is used as the search ending point. The algorithm traverses the associated edges in both forward and reverse time sequences to obtain multiple spatiotemporal connection paths.

[0007] Furthermore, the pre-constructed motion knowledge graph includes By configuring multiple sensing units, real-time information on device operation, body parameters of the target object, and training needs during motion capture is acquired. By defining the roles and functions of each sensing unit, NLP technology is used to extract the relationships between the sensing units and establish a motion knowledge graph connecting them. Each sensing unit is linked and mapped to the corresponding node in the motion knowledge graph. Based on at least one of the extracted device operation information, body parameters, and training needs, a basic layer in the motion knowledge graph is constructed. The basic layer is then matched and associated with real-time multi-source motion data through spatiotemporal connection paths. The multi-source motion data includes at least the coordinate sequence of skeletal key points, angular acceleration of limb segments, and electromyographic signal intensity.

[0008] Furthermore, feature splicing includes: Using each spatiotemporal connection path as an index, node matching and localization are performed in the motion knowledge graph to lock the starting posture node, target posture node, and all action frames between the starting posture node and the target posture node corresponding to each spatiotemporal connection path, so as to determine the first feature vector, the second feature vector, and the third feature vector. The first spliced ​​feature is obtained by feature concatenation. Starting from the second action frame, take the pose nodes corresponding to N intermediate action frames in sequence, and extract multiple fourth feature vectors from the knowledge motion graph and concatenate them in a temporal manner to obtain the second concatenated feature; where N is a positive integer; The first and second splicing features are sequentially aligned and fully spliced ​​to obtain the feature splicing sequence.

[0009] Furthermore, triggering abnormal force exertion includes: For the feature splicing sequence of each spatiotemporal connection path, an attention learning mechanism is introduced to obtain the path feature representation of all action frame pairs on each spatiotemporal connection path; the current training requirement is retrieved and the target training action is identified. For the same target training action, the corresponding path feature representation under M spatiotemporal connection paths is extracted, and PCA dimensionality reduction is performed to obtain M two-dimensional feature points, which are then aggregated into a feature point set; where M is a positive integer. Calculate the Davidson-Bolding index of the feature point set and perform a first screening. If the Davidson-Bolding index is greater than the preset quantile threshold, it is determined to be an invalid action; otherwise, it is determined to be a valid action and a second screening is performed. Extract the attention weight set corresponding to each action frame for each path feature representation, construct a one-dimensional change curve of attention weight distribution in temporal order, and mark it as the weight distribution curve. Extract the main peak of the weight distribution curve and perform consistency verification, including: if the peak temporal positions of at least 2 / M spatiotemporal connection paths are different, it is determined to be abnormal force exertion.

[0010] Furthermore, a force analysis model is built, including: Identify the time period that triggers abnormal force exertion, divide the time period into multiple time intervals, and each time interval corresponds to multiple action frames; collect historically labeled action frame samples, and the samples contain three types of force exertion state labels: normal force exertion, compensation warning, and risk damage. Using the key feature vector of the sample as input and the corresponding force exertion state label as output, the maximum likelihood estimation method is used to train the model parameters to complete the pre-construction of the force exertion analysis model. Multiple time intervals are used as target variables and imported into a pre-defined logistic regression model to calculate the classification probability of the force application state label corresponding to each time interval. The frequency percentage of each force application state corresponding to the time interval is also counted within the current time period. The classification probability and frequency percentage are multiplied and summed to generate the force application state coefficient. The posture node with the force application state coefficient greater than or equal to the standard risk state threshold within the corresponding time interval is located in the motion knowledge graph to determine the occurrence of force application conflict.

[0011] Furthermore, the secondary force determination mechanism includes: Collect the time interval of the force conflict location, extract multiple statistical features based on multi-source motion data, including at least time domain features and frequency domain features, construct a feature candidate set, and extract the data corresponding to b statistical features from the feature candidate set through permutation and combination to generate multiple types of force combination C(n, b); where n is the total number of feature candidate sets, and n and b are both positive integers; A correlation analysis model is configured for each type of force combination. The linear correlation between posture nodes, body parameters, and joint force intensity is retrieved and analyzed, and a differentiated weighted rule base is constructed. If the current force combination is linearly correlated with the joint force intensity, the weighted rule base is called to synthesize the joint force intensity curve in real time, perform first-order difference and extract the time corresponding to the maximum peak value, calculate the time offset with the start time of the time series interval, translate and align the current time series interval, and remap the posture nodes to correct the position of force conflict. If the current force combination is not linearly correlated with the joint force intensity, no processing is performed. The correlation analysis model uses the Pearson similarity coefficient.

[0012] Furthermore, linear correlation includes strong linear correlation and weak linear correlation.

[0013] Secondly, this application provides a strength training force detection system based on motion capture, the system comprising: The determination module determines multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected during the motion capture process; wherein the starting pose node and the target pose node are contained in different motion frames. The first judgment module extracts key feature vectors from the pre-constructed motion knowledge graph based on each spatiotemporal connection path, performs feature splicing, activates the force determination mechanism, builds a force analysis model under the condition of triggering abnormal force exertion, generates the force exertion state of strength training, and marks the force conflict position and time interval in the motion knowledge graph. The second determination module retrieves multi-source motion data from the same time interval based on the location and time interval of the force conflict, triggers a secondary force determination mechanism, constructs multiple types of force combinations, analyzes the linear correlation with the joint force intensity, and corrects the location of the force conflict. The report generation module generates a strength training assessment report based on the corrected force conflict locations and corresponding force states, combined with a motion knowledge graph.

[0014] (III) Beneficial Effects This invention provides a method and system for detecting force exertion during strength training based on motion capture, which has the following beneficial effects: 1. This invention constructs a motion knowledge graph by deploying sensing unit device working information, body parameters, and training requirements. By determining the initial posture node and the target posture node, multiple spatiotemporal connection paths are obtained, discrete motion states are connected in series, and the basic layer is matched and associated with multi-source motion data, which facilitates subsequent multiple force determinations and provides a complete motion capture foundation for force detection. 2. This invention improves judgment efficiency to a certain extent by initiating a force application determination mechanism and setting up an initial screening process. This process uses PCA dimensionality reduction and Davidson-Bolding index calculation to quickly eliminate completely erroneous action frames, avoiding redundant calculations of invalid actions. A secondary screening process performs peak analysis on attention weights and executes consistency checks to quickly identify misalignments in the force application timing during strength training, further refining the judgment efficiency. Simultaneously, the judgment results can be updated to the motion knowledge graph, improving the adaptability of the judgment mechanism and enhancing the robustness of the model. 3. This invention establishes a force analysis model that considers three types of force states: normal training, compensation warning, and damage risk. This model is pre-built and validated, and through training and analysis, achieves precise location of force conflict points. A secondary force determination mechanism, combined with spatiotemporal connection paths, reconstructs the force trajectory and focuses on the time interval where the force conflict point is located for targeted analysis. By comprehensively considering time-domain and frequency-domain characteristics and performing differentiated adaptation based on different force combination types, the linear correlation of the analysis is improved, and the corresponding calculation weights are adjusted. This results in better correction effects, better adaptability, and faster response speed in correcting force conflict points. Attached Figure Description

[0015] Figure 1This is a flowchart illustrating a strength training force detection method according to an exemplary embodiment; Figure 2 This is a schematic diagram of a strength training force detection system according to an exemplary embodiment. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] The core of this invention lies in the fact that, during the motion capture process, by constructing an initial screening of invalid movements, a finely corrected two-level force determination mechanism is implemented under the premise of ensuring valid movements. Combined with the spatiotemporal connection path, the force state is determined, and based on the training requirements of strength training, the linear correlation with the joint force intensity is analyzed to correct the position of force conflict. This achieves the optimal balance between recognition accuracy and determination efficiency in all scenarios.

[0018] Example 1: This invention provides a method for detecting force exertion during strength training based on motion capture. Figure 1 This is a flowchart illustrating a strength training force detection method according to an exemplary embodiment; please refer to [link / reference]. Figure 1 The method includes the following steps: S1: During motion capture, determine multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected; wherein the starting pose node and the target pose node are contained in different motion frames; Determine multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected, including: During the motion capture process, an optical capture sensor can be configured to fix infrared reflective markers on multiple parts of the subject's body. An L-shaped calibration rod is then used to globally calibrate the motion capture camera, establishing a unified world coordinate system: with the center of the motion capture area as the origin O, the X-axis pointing horizontally to the right, the Y-axis pointing vertically upward, and the Z-axis pointing horizontally forward, conforming to the rules of the right-hand Cartesian coordinate system. The reflective markers are then pasted onto key skeletal sites of the subject, with each pasted site corresponding to a subsequent key skeletal point. Using the acquisition sequence of the motion frames as the axis, the three-dimensional coordinates of the same key skeletal point in consecutive motion frames are arranged in ascending order of frame number, generating a coordinate sequence for that key skeletal point. These key skeletal points include, but are not limited to, lumbar spine key points, bilateral hip key points, bilateral knee key points, bilateral ankle key points, bilateral shoulder key points, bilateral elbow key points, and bilateral wrist key points. Based on the obtained coordinate sequence of the skeletal keypoints of the object to be detected in continuous action frames, single-frame topology construction is performed. By using the set of all skeletal keypoints in a single frame as pose nodes and the adjacent skeletal keypoints with relationships as endpoints, topological connection edges in a single frame are constructed to generate a single-frame pose graph, which is represented as G(V, E); where V represents the pose node of a single frame and E represents the set of edges in a single frame. Using the action frame time sequence as the axis, the single-frame pose graphs corresponding to all action frames are temporally associated by concatenating the same skeletal keypoints between adjacent frames. For example, the left hip keypoint in frame i and the left hip keypoint in frame i+1 are skeletal keypoints in the same anatomical position. Then, the inter-frame point correspondence is established. Using the pose nodes corresponding to two adjacent frames as two endpoints, the associated edges of the inter-frame temporal motion are constructed to represent the pose changes and motion process between two adjacent frames. The dynamic topology graph is obtained in sequence and used to construct the connection path in subsequent time steps. By performing pose node locking, the pose node of the action frame corresponding to the start of the action is locked and marked as the starting pose node, and the pose node of the action frame corresponding to the end of the action is locked and marked as the target pose node. A bidirectional breadth-first search algorithm is employed. In the dynamic topology graph, the starting attitude node is used as the search starting point, and the target attitude node is used as the search ending point. Through bidirectional search, specifically, the starting attitude node is used as the initial enqueuing node, and the search direction is the forward temporal direction with increasing frame number to construct the first visited set. The target attitude node is used as the initial enqueuing node, and the search direction is the reverse temporal direction with decreasing frame number to construct the second visited set. A round of traversal operation is performed by alternating between the forward and reverse queues. In each round of traversal, the head node of the queue is removed, and all adjacent attitude nodes that conform to the temporal direction are traversed. Unvisited nodes are added to the queue and the corresponding visited set. After each round of bidirectional traversal, it is checked whether there is an intersection node between the first visited set and the second visited set. If there is an intersection node, it is determined that the forward search and the reverse search have met, and a complete and valid path has been found. Through path backtracking, multiple spatiotemporal connection paths from the starting attitude node to the target attitude node are obtained.

[0019] By determining the action frames, the starting posture node and the target posture node are identified, and multiple spatiotemporal connection paths are obtained for subsequent force determination, providing a complete motion capture foundation for force detection.

[0020] S2: Based on the spatiotemporal connection paths, extract key feature vectors from the pre-constructed motion knowledge graph, perform feature splicing, activate the force determination mechanism, build a force analysis model under the condition of triggering abnormal force exertion, generate the force exertion state of strength training, and mark the force conflict position and time interval in the motion knowledge graph; among which, the key feature vectors include the first feature vector, the second feature vector, the third feature vector, and the fourth feature vector; The steps for acquiring the pre-built motion knowledge graph include: acquiring real-time device operating information, body parameters of the target object, and training requirements during the motion capture process by configuring multiple sensing units; wherein, device operating information includes, but is not limited to, acquisition frequency, CPU utilization, and task queue length; body parameters of the target object include, but are not limited to, height, weight, limb segment mass, limb segment center of gravity position, limb segment rotational inertia, and personalized joint range of motion baseline; training requirements include, but are not limited to, target training movements, target training load, target number of training sets, and target muscle groups; simultaneously, by defining the roles and functions of multiple sensing units, NLP technology is used to extract the relationships between each sensing unit to establish a motion knowledge graph connecting multiple sensing units; and the various sensing units are then... Links are mapped to corresponding nodes in the motion knowledge graph. Based on at least one of the extracted equipment operating information, body parameters, and training requirements, a basic layer in the motion knowledge graph is constructed. This basic layer is then matched and associated with real-time multi-source motion data through spatiotemporal connection paths. The real-time multi-source motion data includes, but is not limited to, the skeletal key point coordinate sequence collected by optical capture sensors, angular acceleration data collected by inertial measurement sensors, electromyographic signal intensity collected by electromyography sensors, force data of the equipment collected by force sensing nodes, and preset safety risk thresholds. It should be noted that when the subject performs the target training action, all sensing units synchronously collect data according to a predefined sampling frequency, and each set of collected data is stamped with a unified timestamp of PTP synchronization. The feature stitching process includes: for each spatiotemporal connection path, using the spatiotemporal connection path as an index, performing node matching and localization in the motion knowledge graph, mapping and aligning all action frames in the spatiotemporal connection path to the corresponding nodes in the basic layer of the motion knowledge graph, in order to determine the device working information, body parameters and training requirements associated with the path nodes at each time point; at the same time, locking the starting posture node, the target posture node and all action frames corresponding to the starting posture node and the target posture node for each spatiotemporal connection path, with each posture node uniquely corresponding to one action frame; Extract the first feature vector associated with the starting pose node in the motion knowledge graph, extract the second feature vector associated with the target pose node in the motion knowledge graph, and based on all action frames, extract the third feature vector associated with the corresponding pose node at the peak moment in the motion knowledge graph. Concatenate the first, second, and third feature vectors using channel features to obtain the first feature vector of the spatiotemporal connection path. Starting from the second action frame, sequentially take the pose nodes corresponding to N intermediate action frames, and temporally concatenate them with the fourth feature vector extracted from the knowledge motion graph to obtain the second concatenated feature. Here, N is a positive integer greater than 0. Temporally align and fully concatenate the first and second concatenated features to obtain the feature concatenation sequence, and mark each element in the feature concatenation sequence as a frame-level feature vector, which uniquely corresponds to one action frame.

[0021] By deploying sensor unit device working information, body parameters, and training needs, a motion knowledge graph is constructed. Combined with spatiotemporal connection paths, the basic layer is matched and associated with multi-source motion data. In this process, through the association relationships extracted by NLP technology, all sensor units, data items, etc. are associated with unified semantic triples, providing a data foundation for attention learning analysis.

[0022] Triggering abnormal force includes: performing spatiotemporal encoding operations based on the feature splicing sequence of each spatiotemporal connection path, that is, performing a linear transformation with uniform dimension and temporal position encoding on all frame-level feature vectors in the feature splicing sequence in sequence, preserving the sequential information of action frames, sorting all frame-level feature vectors according to temporal order, and dividing the feature splicing sequence into a fixed-dimensional sequence of frame-level feature representation vectors; wherein, each vector in the sequence of frame-level feature representation vectors uniquely corresponds to a single action frame in the feature splicing sequence, and the association relationship of that action frame bound in the motion knowledge graph is preserved; From the sequence of frame-level feature representation vectors, extract the frame-level feature vectors corresponding to the starting pose node and the target pose node, and perform global average pooling aggregation to obtain the aggregated representation vector; By introducing an attention learning mechanism, a linear transformation is performed on the aggregated representation vector to generate a query vector for the attention learning mechanism. All frame-level feature representation vectors in the frame-level feature representation sequence are identified, and key and value linear transformations are performed respectively to generate corresponding key and value vectors. All key vectors are arranged in temporal order to form a key vector sequence, and all value vectors are arranged in temporal order to form a value vector sequence. Through temporal attention fusion, the dot product matching degree between the query vector and each key vector in the key vector sequence is calculated. After normalization, the attention weights corresponding to each action frame are obtained. The attention weights corresponding to each action frame are weighted and summed with the corresponding value vector to obtain initial path encoding information, which is used to strengthen key feature weights and weaken non-key feature weights. The initial path encoding information is then residually concatenated with the frame-level feature representation vector sequence and normalized to eliminate temporal encoding bias, resulting in the path encoding information for the spatiotemporal connection path. The aggregated representation vector and path encoding information are concatenated along the channel dimension to obtain the path feature representation of the action frame pair on the spatiotemporal connection path. Simultaneously, the attention weights corresponding to all frames under each spatiotemporal connection path are statistically analyzed and arranged in chronological order of action frames to obtain the attention weight distribution under that spatiotemporal connection path. By retrieving the current training requirements and identifying the target training action, for the same target training action, the path feature representations corresponding to M spatiotemporal connection paths are extracted, and PCA dimensionality reduction is performed to obtain M two-dimensional feature points, which are then aggregated into a feature point set; where M is a positive integer greater than 0. The Davidson-Burdin index of the feature point set is calculated, and a first screening is performed. If the Davidson-Burdin index is greater than a preset quantile threshold, it is marked as an action frame error, indicating that the overall posture of the current training action completely deviates from the target training action template, and is judged as an invalid action, triggering a stop action response. It should be noted that the smaller the Davidson-Burdin index value, the more compact the feature point set and the higher the action consistency; the larger the Davidson-Burdin index value, the more dispersed the feature point set and the more chaotic the action frame, and the preset quantile threshold is usually set to 0.85. Conversely, if the Davidson-Burdin index is less than or equal to the preset quantile threshold, it is judged as a valid action, and a second screening is performed. The attention weight set corresponding to each action frame is extracted from the feature representation of each path, and a one-dimensional change curve of the attention weight is constructed in chronological order and marked as a weight distribution curve. For each weight distribution curve, high-frequency jitter of weight values ​​caused by acquisition noise is eliminated. Through first-order difference, the value with the largest attention weight is selected as the main peak of the weight distribution curve, and the motion corresponding to the main peak is recorded. The frame sequence number is the peak temporal position. If the weight distribution curve of a spatiotemporal connection path has two or more main peaks with similar attention weights, indicating a multi-peak pattern, the peak temporal position of that spatiotemporal connection path is directly marked as abnormal and included in the anomaly count. Then, a consistency check is performed. If at least 2 / M spatiotemporal connection paths have misaligned peak temporal positions, the misalignment is determined by the following criteria: the peak temporal position of a spatiotemporal connection path falls outside the standard peak temporal interval, or the deviation from the center frame number of the interval exceeds ±wc frames, and wc is usually set to 3. In this case, the peak temporal position of the spatiotemporal connection path is determined to be abnormal, and an abnormal force response is triggered. At the same time, the number of peak temporal position anomalies in all spatiotemporal connection paths is counted, and the action frame intervals corresponding to all spatiotemporal connection paths with peak temporal anomalies are marked as force conflict temporal intervals in the motion knowledge graph. The corresponding posture nodes are located and marked as force conflict positions. Otherwise, it is determined to be normal force exertion, and the current action is marked as a normal training state label in the motion knowledge graph.

[0023] Through a first screening process, PCA dimensionality reduction and Davidson-Bolding index calculation are used to quickly eliminate completely erroneous action frameworks and avoid redundant calculations of invalid actions, thereby improving judgment efficiency to a certain extent. Through a second screening process, peak analysis of attention weights is performed and consistency checks are conducted to quickly identify misalignments and anomalies in the timing of force exertion during strength training, further refining and improving judgment efficiency.

[0024] The force analysis model is constructed as follows: The time corresponding to the first peak time-series abnormal action frame is marked as the start timestamp, and the time corresponding to the last peak time-series abnormal action frame is marked as the end timestamp, forming a time period for triggering abnormal force exertion. This time period is divided into multiple time-series intervals, with each interval corresponding to multiple action frames. From a pre-constructed motion knowledge graph, historically labeled action frame samples of the target object are retrieved. These samples include three types of force exertion state labels: normal force exertion, compensation warning, and risk injury. These samples serve as training samples for the force analysis model. The training samples are randomly divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used to solve for model parameters, the validation set is used to monitor model overfitting and adjust hyperparameters, and the test set is used to verify the model's final generalization ability. Using the key feature vectors of the training samples as input and the corresponding force exertion state labels as output, the maximum likelihood estimation method is used to train the model parameters, completing the pre-construction and validation of the model. It should be noted that normal force exertion indicates that the movement framework is compliant, the target muscle group's dominant force exertion sequence is correct, the core joint stress is stable within the safety risk threshold, and there is no muscle compensation or risk of injury – a compliant training state. The compensation warning state indicates that the movement framework is basically compliant, but the force exertion is deviated, the core joint stress is close to the safety risk threshold, there is a potential risk of sports injury, and the movement needs to be adjusted in time – a warning level state. The injury risk state indicates that the force exertion is seriously deviated, the core joint stress exceeds the safety risk threshold, or the joint range of motion is exceeded, and the muscles are uncontrolled activated, there is a clear risk of acute muscle, ligament, and joint injury, and the training needs to be stopped immediately – a high-risk state. Multiple time intervals are used as target variables and imported into a pre-defined logistic regression model to calculate the classification probability of the force exertion state label corresponding to each time interval. This includes a first probability value for a normal training state, a second probability value for a compensation warning state, and a third probability value for an injury risk state, with the sum of the first, second, and third probability values ​​being 1. Simultaneously, the percentage of occurrences of each force exertion state within the current time period is calculated. It should be noted that if the percentage of injury risk states is greater than or equal to 20%, the entire target training movement is determined to be in an injury risk state; if the percentage of compensation warning states is greater than or equal to 30% and there are no injury risk intervals, the entire movement is determined to be in a compensation warning state; all other cases are determined to be normal training states with local anomalies. The force exertion state of each time interval is synchronously updated to the motion knowledge graph, and each time interval with abnormal force exertion is labeled with a corresponding force exertion state label, while simultaneously locating the corresponding force exertion conflict location within that interval. It should be noted that the numerical values ​​mentioned above are merely illustrative examples; specific settings should be implemented based on actual conditions and will not be elaborated upon here. The force application state coefficient is obtained by multiplying and summing the classification probability and the frequency proportion. This involves multiplying the first probability value, the first weight, and the frequency proportion corresponding to the normal training state label; multiplying the second probability value, the second weight, and the frequency proportion corresponding to the compensation warning state label; and multiplying the third probability value, the third weight, and the frequency proportion corresponding to the damage risk state label. The results of these three multiplications are then summed. It should be noted that dimensionless processing is performed during the calculation process to ensure that the calculation results conform to physical meaning. In the motion knowledge graph, posture nodes whose force application state coefficient is greater than or equal to the standard risk state threshold within the corresponding time interval are located to determine if a force application conflict has occurred. The first, second, and third weights are all dynamically changing variables, determined using a genetic algorithm. Specifically, the process involves: In a high-dimensional target space, an initial population is randomly generated, consisting of multiple particles. A certain number of individuals are randomly generated during initialization, and each individual is labeled with a weight combination consisting of the first, second, and third weights. All weight values ​​are ensured to be within a suitable range, between 0 and 1, and the sum of the weight values ​​is 1. Individuals with high fitness are selected as parents. A roulette wheel selection strategy is used to sequentially select particles from the population and place them into a mating pool until the number of particles in the mating pool reaches its maximum. New individuals are obtained through selection, crossover, and mutation operations and added to the population to update its composition. After updating the population, if the average fitness or the optimal solution no longer changes significantly, the result is output, and training stops. The standard risk state threshold is derived from: collecting the force state coefficients of the current target training action under and without force conflict; statistically analyzing the collected data to determine the mean and standard deviation of the force state coefficients under the condition of no force conflict; and setting the standard risk state threshold based on the statistical results, which is the mean of the force state coefficients under the condition of no force conflict plus the standard deviation of 2 or 3 times the standard deviation. It should be noted that this multiple is just an example, and the specific value can be set and adjusted according to the actual situation, which will not be elaborated here.

[0025] By setting up a force analysis model and considering three types of force state labels—normal training, compensation warning, and damage risk—the model is pre-built and validated. After training and analysis, the model can accurately locate the position of force conflict.

[0026] S3: Based on the location and time interval of the force conflict, retrieve multi-source motion data from the same time interval, trigger the secondary force determination mechanism, construct multiple types of force combinations, analyze the linear correlation with the joint force intensity, and correct the location of the force conflict. The secondary force determination mechanism includes: collecting the time interval of the force conflict location, extracting multiple statistical features based on multi-source motion data, including but not limited to time domain features and frequency domain features, constructing a feature candidate set, and extracting data corresponding to b statistical features from the feature candidate set through permutation and combination to generate multiple types of force combinations C(n, b); where n is the total number of feature candidate sets, and n and b are both positive integers; For example, taking a yoga scenario, based on the aforementioned force determination mechanism, the temporal interval of the force conflict location in yoga posture training is obtained. Assuming the force conflict location is the left knee joint node, the corresponding temporal interval is from the 12th to the 18th second of the posture holding phase. Assuming only temporal and frequency domain features are extracted, and that the temporal features include the temporal change rate of joint angles and the temporal integral value of electromyography (EMG), and that the frequency domain features include the median frequency of EMG, the frequency of muscle tremors, and the frequency domain features of center of gravity swaying, then the total number of features in the feature candidate set is 5. Taking b=2, C(5,2)=10. A correlation analysis model is configured for each type of force application combination. The linear correlation between posture nodes, body parameters, and joint stress intensity is retrieved and analyzed, and a differentiated weighted rule base is constructed. If the current force application combination and joint stress intensity are linearly correlated, the weighted rule base is used to synthesize the joint stress intensity curve in real time. Linear correlation includes strong and weak linear correlation. Strong linear correlation is typically defined as a Pearson correlation coefficient greater than 0.7, indicating that changes in the characteristics of this force application combination directly lead to significant changes in joint stress intensity, and strong linear correlation is assigned a higher weight. Weak linear correlation is typically defined as a Pearson correlation coefficient less than or equal to 0.7 and greater than or equal to 0.3, indicating that changes in the characteristics of this force application combination have some impact on joint stress intensity, but the impact is limited. The model assigns lower weights to weakly linearly correlated components. It performs a first-order difference on the joint stress intensity curve, calculates the slope change of the curve, finds the maximum extreme point where the first-order difference changes from positive to negative (this is the maximum peak value of the joint stress intensity), and extracts the time corresponding to the maximum peak value. It retrieves the initial time interval marked by the initial judgment mechanism, calculates the time offset from the initial time interval, translates and aligns the current time interval, and remaps the attitude nodes to correct the force conflict position. If the current force combination has no linear correlation with the joint stress intensity (typically, a Pearson correlation coefficient less than 0.3 is used), indicating that the characteristic changes of the force combination are not significantly related to the joint stress intensity, and assigns a weight of 0 without any processing. The correlation analysis model uses the Pearson similarity coefficient. The corrected information is synchronously updated to the motion knowledge graph, covering the labeling results of the first judgment mechanism; at the same time, the analysis results of linear correlation, the differentiated weighted rule base, and the force combination data are added to the corresponding training database to continuously optimize the subsequent correlation analysis model and weighted rule base, thereby achieving the self-iterative optimization of the system.

[0027] By employing a secondary force determination mechanism, the system focuses on the time interval where the force conflict location is located and conducts targeted analysis. By comprehensively considering time domain and frequency domain characteristics and combining different force combination types to perform differentiated adaptation, the system improves the linear correlation of the analysis and adjusts the corresponding calculation weights. In the process of correcting the force conflict location, the correction effect is better, the adaptability is better, and the response speed is faster.

[0028] S4: Based on the corrected force conflict location and corresponding force state, combined with the motion knowledge graph, a strength training test report is generated.

[0029] Example 2: This invention provides a strength training force detection system based on motion capture; Figure 2 This is a schematic diagram of a strength training force detection system according to an exemplary embodiment; please refer to... Figure 2 The system includes: a determination module, a first determination module, a second determination module, and a report generation module, and the determination module, the first determination module, the second determination module, and the report generation module are connected in communication. The determination module determines multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected during the motion capture process; wherein the starting pose node and the target pose node are contained in different motion frames. The first judgment module extracts key feature vectors from the pre-constructed motion knowledge graph based on each spatiotemporal connection path, performs feature splicing, activates the force determination mechanism, builds a force analysis model under the condition of triggering abnormal force exertion, generates the force exertion state of strength training, and marks the force conflict position and time interval in the motion knowledge graph. The second determination module retrieves multi-source motion data from the same time interval based on the location and time interval of the force conflict, triggers a secondary force determination mechanism, constructs multiple types of force combinations, analyzes the linear correlation with the joint force intensity, and corrects the location of the force conflict. The report generation module generates a strength training assessment report based on the corrected force conflict locations and corresponding force states, combined with a motion knowledge graph.

[0030] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0031] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0032] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0033] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A strength training force detection method based on motion capture, characterized in that, The method includes: during motion capture, determining multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected; wherein the starting pose node and the target pose node are contained in different motion frames; Based on the spatiotemporal connection paths, key feature vectors are extracted from the pre-constructed motion knowledge graph, feature splicing is performed, the force determination mechanism is activated, and under the condition of triggering abnormal force exertion, a force analysis model is built to generate the force exertion state of strength training, and the force conflict location and time interval are marked in the motion knowledge graph. Based on the location and time interval of the force conflict, multi-source motion data in the same time interval are retrieved to trigger a secondary force determination mechanism. By constructing multiple types of force combinations, the linear correlation with the joint force intensity is analyzed, and the location of the force conflict is corrected. Based on the corrected force conflict location and corresponding force state, combined with the motion knowledge graph, a strength training test report is generated.

2. The strength training force detection method based on motion capture according to claim 1, characterized in that, Determine multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected, including: Obtain the coordinate sequence of the skeletal key points of the object to be detected in consecutive action frames, perform single-frame topology construction, and generate a single-frame pose graph by using the set of all skeletal key points in a single frame as pose nodes and adjacent skeletal key points with association as endpoints to construct topological connection edges in a single frame. Using the action frame timing as the axis, the single-frame pose graphs corresponding to all action frames are temporally associated through the chain operation of the same bone key points between adjacent frames. The temporal motion association edges between frames are constructed with the pose nodes of two adjacent frames as endpoints to obtain a dynamic topology graph. Then, pose node locking is performed to determine the starting pose node and the target pose node. A bidirectional breadth-first search algorithm is adopted. In the dynamic topology graph, the starting attitude node is used as the search starting point and the target attitude node is used as the search ending point. The algorithm traverses the associated edges in both forward and reverse time sequences to obtain multiple spatiotemporal connection paths.

3. The strength training force detection method based on motion capture according to claim 1, characterized in that, Pre-built motion knowledge graphs include: By configuring multiple sensing units, real-time information on device operation, body parameters of the target object, and training needs during motion capture is acquired. By defining the roles and functions of each sensing unit, NLP technology is used to extract the relationships between the sensing units and establish a motion knowledge graph connecting them. Each sensing unit is linked and mapped to the corresponding node in the motion knowledge graph. Based on at least one of the extracted device operation information, body parameters, and training needs, a basic layer in the motion knowledge graph is constructed. The basic layer is then matched and associated with real-time multi-source motion data through spatiotemporal connection paths. The multi-source motion data includes at least the coordinate sequence of skeletal key points, angular acceleration of limb segments, and electromyographic signal intensity.

4. The strength training force detection method based on motion capture according to claim 1, characterized in that, Perform feature concatenation, including: Using each spatiotemporal connection path as an index, node matching and localization are performed in the motion knowledge graph to lock the starting posture node, target posture node, and all action frames between the starting posture node and the target posture node corresponding to each spatiotemporal connection path, so as to determine the first feature vector, the second feature vector, and the third feature vector. The first spliced ​​feature is obtained by feature concatenation. Starting from the second action frame, take the pose nodes corresponding to N intermediate action frames in sequence, and extract multiple fourth feature vectors from the knowledge motion graph and concatenate them in a temporal manner to obtain the second concatenated feature; where N is a positive integer; The first and second splicing features are sequentially aligned and fully spliced ​​to obtain the feature splicing sequence.

5. The strength training force detection method based on motion capture according to claim 4, characterized in that, Triggering abnormal force exertion includes: For the feature splicing sequence of each spatiotemporal connection path, an attention learning mechanism is introduced to obtain the path feature representation of all action frame pairs on each spatiotemporal connection path; the current training requirement is retrieved and the target training action is identified. For the same target training action, the corresponding path feature representation under M spatiotemporal connection paths is extracted, PCA dimensionality reduction is performed to obtain M two-dimensional feature points, which are then aggregated into a feature point set; where M is a positive integer. Calculate the Davidson-Bolding index of the feature point set and perform a first screening. If the Davidson-Bolding index is greater than the preset quantile threshold, it is determined to be an invalid action; otherwise, it is determined to be a valid action and a second screening is performed. Extract the attention weight set corresponding to each action frame for each path feature representation, construct a one-dimensional change curve of attention weight distribution in temporal order, and mark it as the weight distribution curve. Extract the main peak of the weight distribution curve and perform consistency verification, including: if the peak temporal positions of at least 2 / M spatiotemporal connection paths are different, it is determined to be abnormal force exertion.

6. The strength training force detection method based on motion capture according to claim 1, characterized in that, Build a force analysis model, including: Identify the time period that triggers abnormal force exertion, divide the time period into multiple time intervals, and each time interval corresponds to multiple action frames; collect historically labeled action frame samples, and the samples contain three types of force exertion state labels: normal force exertion, compensation warning, and risk damage. Using the key feature vector of the sample as input and the corresponding force exertion state label as output, the maximum likelihood estimation method is used to train the model parameters to complete the pre-construction of the force exertion analysis model. Multiple time intervals are used as target variables and imported into a pre-defined logistic regression model to calculate the classification probability of the force application state label corresponding to each time interval. The frequency percentage of each force application state corresponding to the time interval is also counted within the current time period. The classification probability and frequency percentage are multiplied and summed to generate the force application state coefficient. The posture node with the force application state coefficient greater than or equal to the standard risk state threshold within the corresponding time interval is located in the motion knowledge graph to determine the occurrence of force application conflict.

7. The strength training force detection method based on motion capture according to claim 1, characterized in that, The secondary force determination mechanism includes: Collect the time interval of the force conflict location, extract multiple statistical features based on multi-source motion data, including at least time domain features and frequency domain features, construct a feature candidate set, and extract the data corresponding to b statistical features from the feature candidate set through permutation and combination to generate multiple types of force combination C(n, b); where n is the total number of feature candidate sets, and n and b are both positive integers; A correlation analysis model is configured for each type of force application combination. The linear correlation between posture nodes, body parameters, and joint force intensity is retrieved and analyzed, and a differentiated weighted rule base is constructed. If the current force application combination is linearly correlated with the joint force intensity, the weighted rule base is called to synthesize the joint force intensity curve in real time, and the time corresponding to the maximum peak value is extracted. The time offset from the start time of the time series interval is calculated, the current time series interval is shifted and aligned, and the posture nodes are remapped to correct the position of force conflict. If the current force application combination is not linearly correlated with the joint force intensity, no processing is performed. The correlation analysis model uses the Pearson similarity coefficient.

8. The strength training force detection method based on motion capture according to claim 7, characterized in that, Linear correlation includes strong linear correlation and weak linear correlation.

9. A strength training force detection system based on motion capture, characterized in that, The system includes: The determination module determines multiple spatiotemporal connection paths between the starting pose node and the target pose node of the object to be detected during the motion capture process; wherein the starting pose node and the target pose node are contained in different motion frames. The first judgment module extracts key feature vectors from the pre-constructed motion knowledge graph based on each spatiotemporal connection path, performs feature splicing, activates the force determination mechanism, builds a force analysis model under the condition of triggering abnormal force exertion, generates the force exertion state of strength training, and marks the force conflict position and time interval in the motion knowledge graph. The second determination module retrieves multi-source motion data from the same time interval based on the location and time interval of the force conflict, triggers a secondary force determination mechanism, constructs multiple types of force combinations, analyzes the linear correlation with the joint force intensity, and corrects the location of the force conflict. The report generation module generates a strength training assessment report based on the corrected force conflict locations and corresponding force states, combined with a motion knowledge graph.

Images