A deep learning-based trainer posture recognition system and method

Abstract

The present application relates to human motion posture recognition technical field, specifically to a kind of training personnel posture recognition system and method based on deep learning, comprising the following steps: S1, obtains training scene ultra-wideband radio frequency imaging data, extracts microscale dielectric disturbance signal induced by neuromuscular electrical activity, and the neuromuscular electrical activity space-time atlas is generated by clustering muscle group activity;S2, the space-time atlas is input into neural activity analysis network, identifies muscle group synergistic activation relationship and constructs neural synergistic drive graph;S3, neural synergistic drive graph and macroscopic motion characteristics extracted from video are jointly input into posture generation prediction model, predict motion trajectory and output action recognition result and training quality evaluation index.The present application can realize the advanced identification of the action of training personnel and training quality evaluation, so as to improve the accuracy, foresight and explainability of posture recognition.

Description

Technical Field

[0001] This invention relates to the field of human motion posture recognition technology, and in particular to a training posture recognition system and method based on deep learning. Background Technology

[0002] With the development of artificial intelligence, computer vision, and human motion analysis technologies, posture recognition technology based on video or multi-sensor data is increasingly being applied to scenarios such as training behavior analysis, movement standardization evaluation, and training quality monitoring. Current technologies typically collect motion videos of trainees using camera equipment and utilize deep learning models to extract and identify key points, skeletal structures, or joint movement trajectories, thereby achieving automatic identification and analysis of training movement categories. This is of great significance in improving the digitalization and intelligence of the training process.

[0003] However, most existing posture recognition methods rely on macroscopic motion information at the visible level, such as the position of key points in the human body or changes in skeletal structure, lacking effective perception of the internal neuromuscular drive mechanisms. When different actions are similar in appearance, relying solely on external posture features is prone to recognition errors, and it is difficult to reflect the coordinated activation relationships and drive propagation patterns between different muscle groups during the action formation process. Furthermore, existing methods can usually only recognize actions after they have occurred, lacking the ability to predict the action generation process in advance, and it is also difficult to quantitatively evaluate the standardization and coordination of training actions from the perspective of neural drive stability, thus affecting the accuracy of posture recognition results and the reliability of training quality assessment. Summary of the Invention

[0004] This invention provides a deep learning-based system and method for trainee posture recognition. By acquiring microscale dielectric perturbation information induced by human muscle and nerve activity, a spatiotemporal map of neuromuscular electrical activity is constructed and a neural co-driving map is further established. The human neural driving mechanism is fused and analyzed with macroscopic motion structure features. The human posture change path is constrained in the posture generation and prediction model, enabling early recognition of trainee movements and evaluation of training quality, thereby improving the accuracy, foresight, and interpretability of posture recognition.

[0005] A deep learning-based method for training human pose recognition includes the following steps: S1. Acquire ultra-wideband radio frequency imaging data of the training scene, perform human tissue dielectric property inversion processing on each frame in the ultra-wideband radio frequency imaging data, extract microscale dielectric perturbation signals induced by neuromuscular electrical activity of human muscle tissue from the inversion results, and perform cluster analysis on neuromuscular activity of multiple muscle group regions according to the variation law of the spatial distribution and temporal evolution of the microscale dielectric perturbation signals, and generate a spatiotemporal map of neuromuscular electrical activity characterizing the distribution and evolution relationship of nerve driving activities of different muscle groups. S2, the spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The muscle group association analysis network identifies the synergistic activation relationship and its temporal sequence between different muscle groups. Based on the synergistic activation intensity and drive propagation sequence between muscle groups, a neural synergistic drive map reflecting the human motor neural drive mechanism is constructed. The neural synergistic drive map is used to describe the propagation path and drive hierarchy of different muscle group neural drives during the formation of movement in trainees. S3, the neural co-driving graph and the macroscopic motion structure features extracted from the training scene video stream are jointly input into the posture generation prediction model to generate constraints on the human posture change path, thereby predicting the action generation trajectory of the trainee. Based on the action generation trajectory, the action category recognition result that is ahead of the actual human movement and the training quality evaluation index based on neural driving stability are output.

[0006] Optionally, S1 includes: S11, In the training scenario, an ultra-wideband transmitting array and a receiving array are deployed. The transmitting array transmits a wideband pulse signal to the target training area, and the receiving array collects the echo signal after penetrating the human body. The original echo of each transmitting-receiving channel is subjected to time delay correction, channel gain normalization, and background static clutter suppression to obtain ultra-wideband radio frequency imaging data of the training scenario. Based on the ultra-wideband radio frequency imaging data, the dielectric properties of human tissue are inverted for each frame to obtain the dielectric response reconstruction value of the target human body area. S12, perform differential analysis on the reconstructed dielectric response values ​​of adjacent frames to extract the time-varying microscale dielectric perturbation signal within the human muscle tissue region. Based on the prior anatomical template or the location results of human structures in the training area, the target human body is divided into multiple muscle group regions, and the perturbation response value of each muscle group region is calculated separately. , construct the first Temporal feature vectors of muscle group regions ; S13. Based on the temporal feature vectors of each muscle group region, the activity similarity between different muscle group regions is calculated. Cluster analysis is performed on multiple muscle group regions to obtain several muscle group activity clusters with synergistic change trends. Combining the spatial distribution of each muscle group activity cluster and the temporal response sequence, a spatiotemporal map of neuromuscular electrical activity is generated to characterize the distribution and evolution of neural drive activities of different muscle groups.

[0007] Optionally, S11 includes: S111, Ultra-wideband transmitting and receiving arrays are deployed on both sides, front and back, or around the target training area in the training scenario. These arrays form a multi-channel penetrating detection path covering the target human body. The transmitting arrays emit broadband pulse signals into the target training area according to a preset transmission sequence, while the receiving arrays synchronously collect the raw echo signals after penetrating the human body from each transmitting-receiving channel, forming a multi-channel raw echo dataset for the current frame. ; S112, the transmit-receive channel signals in the multi-channel raw echo dataset are sequentially processed with time delay correction, channel gain normalization, and background static clutter suppression to obtain ultra-wideband radio frequency imaging data of the training scene. ; S113, based on the first The ultra-wideband radio frequency imaging data of the frame is used to perform human tissue dielectric property inversion processing on each spatial sampling point within the target training area, and the reconstructed dielectric response value of each spatial sampling point is calculated. The dielectric response distribution of the target human body region is formed by reconstructing the dielectric response values ​​of all spatial sampling points. .

[0008] Optionally, S2 includes: S21, the spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The neural drive signals of each muscle group region in the spatiotemporal map of neuromuscular electrical activity are feature-encoded in the time dimension. The temporal correlation and degree of synergistic change between different muscle groups are calculated by the muscle group association analysis unit in the neural activity analysis network. In this way, the synergistic activation relationship of multiple muscle groups being activated simultaneously or continuously during the action formation process is identified, and the synergistic activation intensity matrix between each muscle group is obtained to characterize the degree of synergy of neural drive activities between different muscle groups. S22, based on the co-activation intensity matrix and the activation sequence of nerve drive signals of each muscle group in the time dimension, the drive propagation path between different muscle groups is analyzed. Each muscle group region is taken as a node in the graph structure, and the co-activation relationship and drive propagation direction between muscle groups are taken as edge relationships. The direction and weight of the edges are determined according to the temporal relationship of drive propagation, thereby constructing a neural co-drive graph that reflects the human motor nerve drive mechanism, which is used to describe the propagation path and drive hierarchy relationship of nerve drive of different muscle groups in the process of movement formation of trainees.

[0009] Optionally, S21 includes: S211, Input the spatiotemporal map of neuromuscular electrical activity into the neural activity analysis network, and extract the neural drive signal sequences of each muscle group region in the time dimension from the spatiotemporal map. The neural drive signal sequence is then temporally encoded to obtain neural drive feature encoding vectors characterizing each muscle group region. ; S212 utilizes the muscle group association parsing unit in the neural activity parsing network to perform correlation analysis on the neural drive feature encoding vectors between different muscle groups, thereby calculating the temporal correlation coefficient between each muscle group. This characterizes the temporal correlation and degree of coordinated change over time. S213, based on the temporal correlation coefficient between different muscle groups, identifies the synergistic activation relationship between muscle groups. Greater than the collaborative threshold At that time, the judgment of the first The muscle group region and the first There is a synergistic activation relationship among various muscle groups during the movement formation process, and a synergistic activation intensity matrix among these muscle groups is constructed. It is used to characterize the degree of coordination of neural drive activities of different muscle groups.

[0010] Optionally, S22 includes: S221, Based on the perturbation response values ​​of each muscle group region, extract the main activation time of each muscle group within the target time window. Based on the sequential relationship of the principal activation times of different muscle groups, the driving sequence between muscle groups is determined, which is used to characterize the starting and ending points of the neural driving activities of each muscle group in the time dimension during the formation of movement, and to calculate the first... The muscle group region and the first Activation time difference between individual muscle groups ,when When, it indicates the first The muscle group region is earlier than the first When individual muscle groups are activated, When, it indicates the first The muscle group region is earlier than the first Several muscle groups were activated; S222: Based on the co-activation intensity matrix and the activation time difference between different muscle group regions, muscle group pairs with co-activation relationships are screened, and the direction of the edges is determined according to the activation order, while the edge weights are determined according to the co-activation intensity and the temporal relationship. This allows for the analysis of the drive propagation pathways between different muscle groups; S223, using each muscle group region as a node in the graph structure, and the determined drive propagation edges and their weights as directed edges in the graph structure, a neural collaborative drive graph reflecting the human motor neural drive mechanism is constructed. The driving level score is calculated based on the input and output edge weights of each node. ,when At that time, the judgment of the first Each muscle group region corresponds to a driver source node, when At that time, the judgment of the first Each muscle group region corresponds to a corresponding node as a response node, where... The threshold for determining the driving level is determined by scoring the driving level of each node. By comparison, the hierarchical relationship of the driving force in different muscle groups during the movement formation process can be determined.

[0011] Optionally, S3 includes: S31, the neural co-driving graph is aligned and jointly encoded with the macroscopic motion structure features extracted from the training scene video stream. The macroscopic motion structure features include the spatial distribution features of human joint nodes, the joint angle change features, and the limb movement speed features. The driving hierarchy relationship, driving propagation path, and edge weight in the neural co-driving graph are used as prior constraints and input into the posture generation prediction model. The temporal constraint unit in the posture generation prediction model constrains the motion change path of each joint node in the human body in the time dimension, so that the human posture generation process simultaneously satisfies the macroscopic motion structure law and the neural driving propagation law, thereby obtaining candidate posture change paths that conform to the neural driving mechanism. S32, based on the candidate posture change path, the motion state of each joint node of the human body in the future time window is predicted by the posture generation prediction model, and the action generation trajectory of the trainee in the future time window is generated. According to the matching result between the action generation trajectory and the preset action template library, the action category recognition result that is ahead of the actual human movement is output. At the same time, combined with the stability index of each driving node in the neural collaborative driving graph and the continuity of driving propagation between different muscle groups, the neural driving stability of the current training action is evaluated, thereby generating a training quality evaluation index that reflects the standardization and coordination of the training action.

[0012] Optionally, S31 includes: S311, extract the spatial distribution features of human joint nodes, joint angle change features, and limb movement speed features from the training scene video stream, and perform spatiotemporal correspondence between the macroscopic motion structure features and muscle group nodes in the neural co-drive graph to complete the feature alignment between neural drive information and macroscopic motion information, thereby obtaining the joint-muscle group alignment feature vector. ; S312, the first The macroscopic motion structure features of the frame are jointly encoded with the joint-muscle group alignment feature vector, and the driving hierarchy, driving propagation path and edge weights in the neural co-driving graph are embedded into the joint features to obtain the pose fusion feature sequence under neural driving constraints. ; S313, The posture fusion feature sequence is input into the temporal constraint unit of the posture generation prediction model. The motion change path of each joint node of the human body in the time dimension is constrained by the neural co-driving graph to generate candidate posture change paths that satisfy the macroscopic motion structure law and the neural driving propagation law, specifically including: S3131, Construct a human posture state vector based on the spatial coordinate information of human joint nodes extracted from the video stream. , used to characterize the human body's posture state in the time dimension; S3132, the first in the pose fusion feature sequence Jointly encoded feature vector of frames The current human posture state vector is input into the temporal prediction unit of the posture generation and prediction model to predict the human posture state at the next moment. S3133, according to the neural synergistic driving diagram, the first The muscle group node and the first Edge weights between muscle group nodes Calculate neural drive constraint terms ; S3134 combines the attitude prediction error with the neural drive constraint term to construct the optimization objective function of the attitude generation prediction model; S3135, by minimizing the objective function to solve the sequence of human posture changes, candidate posture change paths that satisfy the laws of neural drive propagation are obtained. .

[0013] Optionally, S32 includes: S321, Based on the candidate posture change path, the posture generation prediction model is used to recursively predict the motion state of each joint node of the human body within a future time window, obtaining the human posture state vector at each moment. The human posture state vectors at each moment within the future time window are then combined in chronological order to generate the trainee's motion generation trajectory. Extracting action to generate trajectory feature vectors ; S322, perform similarity matching between the motion generation trajectory feature vector and the feature vectors of each motion template in the preset motion template library to determine the motion template category that best matches the motion generation trajectory, and output the corresponding motion template category as the motion category recognition result that is ahead of the actual human movement. Specifically, this includes: S3221, assuming the preset action template library contains a total of Action template, number The feature vector of the action template is Then the action-generated trajectory feature vector and the first Matching similarity between action templates; S3222, Take the action template category with the highest matching similarity as the action category recognition result. ; S323, combining the stability indices of each drive node in the neural synergistic drive diagram and the continuity of drive propagation between different muscle groups, assesses the neural drive stability of the current training movement, and generates training quality assessment indicators reflecting the standardization and coordination of the training movement based on the assessment results. Specifically, it includes: S3231, according to the... The driving characteristic values ​​of each driving node within the target time window Calculate the first Stability metrics of each driver node ; S3232, Calculate the propagation continuity index This characterizes the continuity of drive propagation between different muscle groups; S3233, based on the stability index and drive propagation continuity index of each driving node, calculates the training quality evaluation index. The training quality assessment indicators are used to reflect the standardization and coordination of training movements. The higher the value, the higher the degree of neural drive stability of the training movement, the more standardized the movement execution, and the more coordinated the muscle group synergy.

[0014] A deep learning-based trainer pose recognition system, used to implement the aforementioned deep learning-based trainer pose recognition method, includes the following modules: Data acquisition and neural electrical activity reconstruction module: Acquires ultra-wideband radio frequency imaging data of the training scene, and performs human tissue dielectric property inversion processing on each frame of the ultra-wideband radio frequency imaging data. Extracts microscale dielectric perturbation signals induced by neuromuscular electrical activity of human muscle tissue from the inversion results. Based on the variation law of the spatial distribution and temporal evolution of the microscale dielectric perturbation signals, performs cluster analysis on the neuromuscular activity of multiple muscle group regions, and generates a spatiotemporal map of neuromuscular electrical activity characterizing the distribution and evolution relationship of neural drive activities of different muscle groups. Neural drive relationship analysis module: The spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The muscle group association analysis network identifies the synergistic activation relationship and its temporal sequence between different muscle groups. Based on the synergistic activation intensity and drive propagation sequence between muscle groups, a neural synergistic drive map reflecting the human motor neural drive mechanism is constructed. Posture prediction and action recognition module: The neural co-driving graph and the macroscopic motion structure features extracted from the training scene video stream are jointly input into the posture generation prediction model to generate constraints on the human posture change path, thereby predicting the action generation trajectory of the trainee, and outputting action category recognition results that are ahead of the actual human movement and training quality evaluation indicators based on neural driving stability based on the action generation trajectory.

[0015] The beneficial effects of this invention are: This invention, by employing a combined approach of ultra-wideband radio frequency imaging, inversion of the dielectric properties of human tissue, and extraction of microscale dielectric perturbation signals, can obtain deep prior information about the movement of trainees from the neuromuscular activity level of muscle groups, which is difficult to directly represent in conventional videos. Furthermore, it constructs a spatiotemporal map of neuromuscular electrical activity, thereby overcoming the limitations of existing posture recognition methods that mainly rely on externally visible postures or skeletal key points, and effectively improving the perception of hidden muscle group activation, early movement intentions, and the formation process of complex movements.

[0016] This invention analyzes the spatiotemporal map of neuromuscular electrical activity to identify the synergistic activation relationships, activation sequence, drive propagation paths, and drive hierarchy among different muscle groups, and constructs a neural synergistic drive map. This explicitly introduces the neural drive propagation patterns in training movements into the posture recognition process. It not only depicts the synergistic and transmission relationships among different muscle groups during movement formation but also avoids misidentification problems caused by judging solely based on macroscopic movement appearance. This improves the ability to distinguish between similar-looking movements, continuous transitional movements, and complex combinations of movements, thereby enhancing the accuracy and interpretability of posture recognition results.

[0017] This invention aligns and jointly encodes neural co-driving graphs with macroscopic motion structure features, and introduces neural driving constraints, motion generation trajectory prediction, motion template matching, and training quality assessment mechanisms into the posture generation prediction model. This not only outputs motion category recognition results that are ahead of actual human motion, but also simultaneously generates training quality assessment indicators that reflect the standardization and coordination of training motions. Therefore, compared with existing technologies that only perform post-event recognition, it can simultaneously achieve advance motion prediction and quantitative assessment of training quality, which is beneficial to improving the real-time performance, foresight, and assessment accuracy of training process monitoring. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the identification method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system functional modules according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figure 1 As shown, a deep learning-based method for training human pose recognition includes the following steps: S1. Acquire ultra-wideband radio frequency imaging data of the training scene, perform inversion processing of the dielectric properties of human tissue on each frame of the ultra-wideband radio frequency imaging data, extract the microscale dielectric perturbation signal induced by neuromuscular electrical activity of human muscle tissue from the inversion results, and perform cluster analysis on the neuromuscular activity of multiple muscle groups based on the variation law of the microscale dielectric perturbation signal in spatial distribution and temporal evolution, and generate a spatiotemporal map of neuromuscular electrical activity that characterizes the distribution and evolution relationship of nerve driving activities of different muscle groups. S2. The spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The muscle group association analysis network identifies the synergistic activation relationship and its temporal sequence between different muscle groups. Based on the synergistic activation intensity and drive propagation sequence between muscle groups, a neural synergistic drive map reflecting the human motor neural drive mechanism is constructed. The neural synergistic drive map is used to describe the propagation path and drive hierarchy of different muscle group neural drives during the formation of movement in trainees. S3 inputs the neural co-driving graph and the macroscopic motion structure features extracted from the training scene video stream into the posture generation prediction model to generate constraints on the human posture change path, thereby predicting the trainee's action generation trajectory. Based on the action generation trajectory, it outputs the action category recognition result that is ahead of the actual human movement and the training quality evaluation index based on neural drive stability.

[0022] S1 includes: S11. In the training scenario, an ultra-wideband transmitting array and a receiving array are deployed. The transmitting array transmits a wideband pulse signal to the target training area, and the receiving array collects the echo signal after penetrating the human body. The original echo of each transmitting-receiving channel is subjected to time delay correction, channel gain normalization, and background static clutter suppression to obtain ultra-wideband radio frequency imaging data of the training scenario. Based on the ultra-wideband radio frequency imaging data, the dielectric properties of human tissue are inverted for each frame to obtain the dielectric response reconstruction value of the target human body area. S12, perform differential analysis on the reconstructed dielectric response values ​​of adjacent frames to extract the time-varying microscale dielectric perturbation signal within the human muscle tissue region. Based on the prior anatomical template or the location results of human structures in the training area, the target human body is divided into multiple muscle group regions, and the perturbation response value of each muscle group region is calculated separately. , construct the first Temporal feature vectors of muscle group regions , represented as: ; ; ; in, , The first , Frame at pixel Dielectric response reconstruction value at the location, For the first The set of spatial points corresponding to each muscle group region Muscle group area The total number of spatial points in the space, , , , Disturbance response value The mean, standard deviation, temporal gradient mean, and energy accumulation value within the target time window; ; ; ; ; in, The total number of frames within the target time window. For the first The muscle group region in the first The frame's perturbation response value; S13, based on the temporal feature vectors of each muscle group region, calculate the activity similarity between different muscle group regions, perform cluster analysis on multiple muscle group regions to obtain several muscle group activity clusters with synergistic change trends, and combine the spatial distribution of each muscle group activity cluster and the temporal response sequence to generate a spatiotemporal atlas of neuromuscular electrical activity characterizing the distribution and evolution of neural drive activities of different muscle groups, specifically including: (1) The similarity of activity between different muscle group regions is represented as follows: ; in, For the first The muscle group region and the first The characteristic distance between individual muscle groups For the first Temporal feature vectors of individual muscle groups For the first Temporal feature vectors of individual muscle group regions; (2) Clustering is performed using the cluster center assignment method. The category label for each muscle group region is represented as follows: ; in, for The cluster category to which each muscle group region belongs. To preset the total number of cluster categories, For the first The feature vectors of each cluster center; ; ; ; in, For the first The number of muscle groups contained in each cluster. For indicator functions, when The time value is Otherwise, the value is ; (3) The spatiotemporal map of neuromuscular electrical activity is represented as follows: ; in, For the first The spatiotemporal map of neuromuscular electrical activity corresponding to the frame. For the first The cluster category to which each muscle group region belongs. For the first Spatial coordinates of each muscle group region This represents the total number of muscle groups.

[0023] S11 includes: S111, Ultra-wideband transmitting and receiving arrays are deployed on both sides, front and back, or around the target training area in the training scenario. These arrays form a multi-channel penetrating detection path covering the target human body. The transmitting arrays emit broadband pulse signals into the target training area according to a preset transmission sequence, while the receiving arrays synchronously collect the raw echo signals after penetrating the human body from each transmitting-receiving channel, forming a multi-channel raw echo dataset for the current frame. , is represented as: ; in, For the first The first frame The transmitting antenna and the first The original echo signal of each receiving antenna corresponding to the channel. This represents the total number of transmitting antennas. This represents the total number of receiving antennas. S112, the transmit-receive channel signals in the multi-channel raw echo dataset are sequentially processed with time delay correction, channel gain normalization, and background static clutter suppression to obtain the ultra-wideband radio frequency imaging data of the training scene. Specifically, it includes: (1) Time delay correction: for the first The transmitting antenna and the first The original echo signals of each receiving antenna corresponding to a channel are time-aligned to obtain the time-delay corrected echo signals, represented as: ; in, For the first The first frame - The echo signal after time delay correction of the channel, For the first - The system propagation delay compensation amount of the channel relative to the reference channel. This is the result after time shifting the original echo signal; (2) Channel gain normalization: The amplitude of the echo signal after time delay correction is normalized to obtain the signal with normalized channel gain, which is expressed as: ; in, For the first The first frame - The echo signal after channel gain normalization. For the first - Channel gain calibration factor; When the peak amplitude of this channel during the preset calibration phase is used as the calibration basis, the gain calibration coefficient... Represented as: ; in, For the first - The time-delay corrected echo signal of the channel under reference calibration conditions; (3) The gain-normalized echo signal is differentially divided with the static background reference signal to obtain the effective echo signal and generate ultra-wideband radio frequency imaging data. , is represented as: ; ; in, For the first The first frame - The effective echo signal of the channel after background static clutter suppression. For the first - The static background reference signal corresponding to the channel, For the first Frames of ultra-wideband radio frequency imaging data; S113, based on the first The ultra-wideband radio frequency imaging data of the frame is used to perform human tissue dielectric property inversion processing on each spatial sampling point within the target training area, and the reconstructed dielectric response value of each spatial sampling point is calculated. The dielectric response distribution of the target human body region is formed by reconstructing the dielectric response values ​​of all spatial sampling points. , is represented as; ; ; ; ; in, For the first - Channel for spatial sampling points The propagation weighting coefficient, For the signal from the first Each transmitting antenna propagates to a space sampling point. Then spread to the The total propagation delay of each receiving antenna For the first One transmitting antenna to the space sampling point The distance of transmission Spatial sampling points To the The propagation distance of each receiving antenna, This is the equivalent propagation speed of electromagnetic waves in the target human body and the environmental medium. The path decay factor, This is the set of spatial sampling points corresponding to the target human body region.

[0024] S2 includes: S21, the spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The neural drive signals of each muscle group region in the spatiotemporal map of neuromuscular electrical activity are feature-encoded in the time dimension. The temporal correlation and degree of synergistic change between different muscle groups are calculated through the muscle group association analysis unit in the neural activity analysis network. In this way, the synergistic activation relationship of multiple muscle groups being activated simultaneously or continuously during the action formation process is identified, and the synergistic activation intensity matrix between each muscle group is obtained, which is used to characterize the degree of synergy of neural drive activities between different muscle groups. S22, based on the co-activation intensity matrix and the activation sequence of nerve drive signals of each muscle group in the time dimension, analyzes the drive propagation path between different muscle groups, takes each muscle group region as a node in the graph structure, takes the co-activation relationship between muscle groups and the drive propagation direction as the edge relationship, and determines the direction and weight of the edge according to the temporal relationship of drive propagation, thereby constructing a neural co-drive graph that reflects the human motor neural drive mechanism, which is used to describe the propagation path of nerve drive of different muscle groups and its drive hierarchy relationship during the formation of movement in trainees.

[0025] S21 includes: S211, Input the spatiotemporal map of neuromuscular electrical activity into the neural activity analysis network, and extract the neural drive signal sequences of each muscle group region in the time dimension from the spatiotemporal map. The neural drive signal sequence is then temporally encoded to obtain neural drive feature encoding vectors characterizing each muscle group region. , is represented as: ; ; in, Indicates the first The muscle group region in the first The perturbation response value of the frame. , The total number of frames within the target time window. For encoding temporal features in neural activity parsing networks; Temporal feature encoding function in neural activity parsing network Represented as: ; ; ; in, For the first The hidden feature vector of a frame, For the input weight matrix, For the recursive weight matrix, For bias vectors, The hyperbolic tangent activation function is used. This is the initial hidden state vector; S212 utilizes the muscle group association parsing unit in the neural activity parsing network to perform correlation analysis on the neural drive feature encoding vectors between different muscle groups, thereby calculating the temporal correlation coefficient between each muscle group. The temporal correlation and degree of co-change in the time dimension are represented as follows: ; in, For the first The muscle group region and the first Temporal correlation coefficients between different muscle groups For the first The muscle group region in the first The perturbation response value of the frame. For the first The muscle group region in the first The perturbation response value of the frame. For the first The mean of the neural drive signal sequence of each muscle group region For the first The mean value of the neural drive signal sequence of each muscle group region; S213, based on the temporal correlation coefficient between different muscle groups, identifies the synergistic activation relationship between muscle groups. Greater than the collaborative threshold At that time, the judgment of the first The muscle group region and the first There is a synergistic activation relationship among various muscle groups during the movement formation process, and a synergistic activation intensity matrix among these muscle groups is constructed. The degree of coordination of neural drive activities of different muscle groups is used to characterize the degree of coordination, and is expressed as: ; ; in, , These represent the mean and standard deviation of the temporal correlation coefficients between all muscle groups within the target time window. This is the threshold adjustment coefficient.

[0026] S22 includes: S221, Based on the perturbation response values ​​of each muscle group region, extract the main activation time of each muscle group within the target time window. Based on the sequential relationship of the principal activation times of different muscle groups, the driving sequence between muscle groups is determined, which is used to characterize the starting and ending points of the neural driving activities of each muscle group in the time dimension during the formation of movement, and to calculate the first... The muscle group region and the first Activation time difference between individual muscle groups ,when When, it indicates the first The muscle group region is earlier than the first When individual muscle groups are activated, When, it indicates the first The muscle group region is earlier than the first The activation of a muscle group region is represented as follows: ; ; in, , The first The muscle group area, the first The primary activation time of each muscle group region; S222: Based on the co-activation intensity matrix and the activation time difference between different muscle group regions, muscle group pairs with co-activation relationships are screened, and the direction of the edges is determined according to the activation order, while the edge weights are determined according to the co-activation intensity and the temporal relationship. This allows us to analyze the drive propagation pathways between different muscle groups, represented as follows: ; ; in, For the first The muscle group region and the first Are there connections between different muscle groups? point to The driving propagation edge; S223, using each muscle group region as a node in the graph structure, and the determined drive propagation edges and their weights as directed edges in the graph structure, a neural collaborative drive graph reflecting the human motor neural drive mechanism is constructed. The driving level score is calculated based on the input and output edge weights of each node. It is used to determine the hierarchical relationship of the driving force of each muscle group region during the movement formation process. At that time, the judgment of the first Each muscle group region corresponds to a driver source node, when At that time, the judgment of the first Each muscle group region corresponds to a corresponding node as a response node, where... The threshold for determining the driving level is determined by scoring the driving level of each node. By comparing and determining the hierarchical relationship of different muscle groups during the movement formation process, it can be represented as: ; in, Let be the set of nodes in the graph. For the first The nodes corresponding to each muscle group region Let be the set of directed edges in the graph. Let be the set of weights for each directed edge; ; in, For the first The total edge weight output by each node to other nodes Point other nodes to the first The total input edge weights of each node; .

[0027] S3 includes: S31, the neural co-driving graph is aligned and jointly encoded with the macroscopic motion structure features extracted from the training scene video stream. The macroscopic motion structure features include the spatial distribution features of human joint nodes, the joint angle change features, and the limb movement speed features. The driving hierarchy relationship, driving propagation path, and edge weight in the neural co-driving graph are used as prior constraints input to the posture generation prediction model. The temporal constraint unit in the posture generation prediction model constrains the motion change path of each joint node in the human body in the time dimension, so that the human posture generation process simultaneously satisfies the macroscopic motion structure law and the neural driving propagation law, thereby obtaining candidate posture change paths that conform to the neural driving mechanism. S32, based on candidate posture change paths, predicts the motion state of each joint node of the human body within a future time window through a posture generation prediction model, generates the action generation trajectory of the trainee within the future time window, and outputs the action category recognition result that is ahead of the actual human movement based on the matching result between the action generation trajectory and the preset action template library. At the same time, it evaluates the neural drive stability of the current training action by combining the stability index of each drive node in the neural collaborative drive graph and the continuity of drive propagation between different muscle groups, thereby generating a training quality evaluation index that reflects the standardization and coordination of the training action.

[0028] S31 includes: S311: Extract spatial distribution features of human joint nodes, joint angle change features, and limb movement speed features from the training scene video stream. Then, perform spatiotemporal correspondence between the macroscopic motion structure features and muscle group nodes in the neural co-drive graph to complete feature alignment between neural drive information and macroscopic motion information, and obtain joint-muscle group alignment feature vectors. Specifically, it includes: (1) Let the first There are a total of human figures in the frame The joint node, the first The spatial coordinates of each joint node are: Then the first The spatial distribution characteristics of human joint nodes in a frame are represented as follows: ; in, For the first Spatial distribution characteristics of human joint nodes in a frame; (2) Let the first Each joint connection unit consists of joint nodes. With joint nodes Composition, then the first The joint angle change features corresponding to the frame are represented as follows: ; in, For the first The joint connection unit in the first The characteristics of joint angle changes in frames. , These are the two skeleton vectors that constitute this joint angle; (3) Let the first The reference node coordinates for each limb unit are: Then the first The limb movement velocity features corresponding to the frame are represented as follows: ; in, For the first The limb unit in the first frame motion speed, , For the first The limb unit reference node at the th Frame coordinates; (4) The spatial distribution characteristics of human joint nodes, the characteristics of joint angle changes, and the characteristics of limb movement speed are spliced ​​together to generate the first... Macroscopic motion structure features of frames , is represented as: ; (5) Establish a correspondence between muscle group nodes in the neural synergistic drive diagram and joint nodes in the macroscopic motion structure features, and define the joint-muscle group mapping matrix as follows: ,in, For the first The first joint node and the first Mapping weights between muscle group nodes; ; ; in, This is the distance attenuation coefficient. For the first The first joint node and the first Spatial distance between muscle group nodes Indicates the first The position vectors of each joint node. Indicates the first Each muscle group node corresponds to the center position vector of the muscle group region; (6) Based on the joint-muscle group mapping matrix as , obtained the The joint-muscle alignment feature vector of a frame is represented as: ; in, This is the joint-muscle group alignment feature vector. For the first The frame is a muscle group drive feature vector extracted from the neural co-drive graph; ; ; in, For the first The muscle group node at the first The driving feature value of the frame, For the first The muscle group region in the first The perturbation response value of the frame. For the first The driving hierarchy score of individual muscle group nodes in the neural synergistic driving map. For the first The propagation driving force received by a muscle group node from its upstream associated muscle group. , , These are the corresponding weighting coefficients; S312, the first The macroscopic motion structure features of the frame are jointly encoded with the joint-muscle group alignment feature vector, and the driving hierarchy, driving propagation path and edge weights in the neural co-driving graph are embedded into the joint features to obtain the pose fusion feature sequence under neural driving constraints. Specifically, it includes: (1) According to the first Macroscopic motion structure features of frames Feature vectors aligned with joint-muscle groups , construct the first The initial fusion features of the frame are represented as: ; in, For the first Initial fusion features of the frame, The mapping weight matrix represents the macroscopic motion structure characteristics. This is the mapping weight matrix for joint-muscle group alignment features. It is the bias vector; (2) Calculate the driving level score vector of the neural synergistic driving graph. With edge weight matrix , is represented as: ; ; in, , , The first , , Driver hierarchy score for individual muscle group nodes For the first The muscle group node points to the first Edge weights of each muscle group node; (3) Map the driving hierarchy and driving propagation relationships to the fusion features to construct the first... Jointly encoded feature vector of frames And the joint encoded feature vector of each frame Composition of pose fusion feature sequence , is represented as: ; in, The mapping weight matrix that drives the hierarchical scoring vector. The edge weight matrix The derived propagation constraint mapping matrix; ; S313, inputting the posture fusion feature sequence into the temporal constraint unit of the posture generation prediction model, constrains the motion change paths of each joint node in the human body in the time dimension through the neural co-driving graph, generating candidate posture change paths that satisfy the macroscopic motion structure law and the neural driving propagation law, specifically including: S3131, Construct a human posture state vector based on the spatial coordinate information of human joint nodes extracted from the video stream. This characterizes the human body's posture state over time, and is represented as: ; S3132, the first in the pose fusion feature sequence Jointly encoded feature vector of frames The current human pose state vector is input into the temporal prediction unit of the pose generation and prediction model to predict the human pose state at the next moment, as shown below: ; in, For the first The predicted pose state vector of the frame. For the temporal prediction function in the attitude generation prediction model, For the first The predicted pose state vector of the frame; The temporal prediction function in the pose generation and prediction model is expressed as: Update Gate: ; Reset Door: ; Candidate hidden states: ; Hidden state update: ; Next-moment attitude prediction: ; in, For the first The hidden state vector at time t. For the first The hidden state vector updated at each time step. For the first The candidate hidden state vector at time t. For the first Update the gate vector at each time step. For the first Reset gate vector at time step , , These are the weight matrices corresponding to the update gates. , , These are the weight matrices corresponding to the reset gates. , , These are the weight matrices corresponding to the candidate hidden states. To output the mapping weight matrix, , , , These are the corresponding bias vectors. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. For element-wise vector multiplication, This represents the result of subtracting the corresponding element value from each element of the update gate vector by taking 1; S3133, according to the neural synergistic driving diagram, the first The muscle group node and the first Edge weights between muscle group nodes Calculate neural drive constraint terms , is represented as: ; in, , The first The first muscle group node, the first The amount of posture change corresponding to each muscle group node. To drive the propagation ratio coefficient; S3134 combines the pose prediction error with the neural drive constraint term to construct the optimization objective function of the pose generation and prediction model, expressed as: ; ; in, For attitude prediction error, This is the true attitude state vector. To predict the attitude state vector, Generate the total loss function for the pose prediction model. These are the neural drive constraint weight coefficients; S3135, by minimizing the objective function to solve the sequence of human posture changes, candidate posture change paths that satisfy the laws of neural drive propagation are obtained. , is represented as: ; in, For the first The predicted human pose vector at time t. , To predict the length of the time window.

[0029] S32 includes: S321, based on candidate posture change paths, uses a posture generation prediction model to recursively predict the motion state of each joint node of the human body within a future time window, obtaining the human posture state vector at each moment. Then, the human posture state vectors at each moment within the future time window are combined in chronological order to generate the trainee's motion generation trajectory. Extracting action to generate trajectory feature vectors , is represented as: ; ; ; in, , The first The joint node at the _ ... Predicted horizontal and vertical coordinates at each time point , This represents the total number of joint nodes. This represents the change in attitude state between two adjacent time points; S322, perform similarity matching between the feature vector of the generated motion trajectory and the feature vectors of each motion template in the preset motion template library, determine the motion template category that best matches the generated motion trajectory, and output the corresponding motion template category as the motion category recognition result that precedes the actual human movement. Specifically, this includes: S3221, assuming the preset action template library contains a total of Action template, number The feature vector of the action template is Then the action-generated trajectory feature vector and the first The matching similarity between action templates is represented as: ; in, Generate trajectory for action and the first Matching similarity between action templates; S3222, Take the action template category with the highest matching similarity as the action category recognition result. , is represented as: ; S323, combining the stability indices of each drive node in the neural synergistic drive diagram and the continuity of drive propagation between different muscle groups, assesses the neural drive stability of the current training movement, and generates training quality assessment indicators reflecting the standardization and coordination of the training movement based on the assessment results. Specifically, it includes: S3231, according to the... The driving characteristic values ​​of each driving node within the target time window Calculate the first Stability metrics of each driver node , is represented as: ; ; ; in, , The first The mean and standard deviation of the feature sequence driven by each driving node. To prevent extremely small positive numbers with a denominator of zero; S3232, Calculate the propagation continuity index This characterizes the continuity of drive propagation between different muscle groups, and is represented as: ; ; in, The total number of driving nodes, For the first The moment is determined by the first The number of driver nodes up to the first The propagation consistency indicator of each driving node. The propagation ratio coefficient is determined by the driving propagation path; S3233, based on the stability index and drive propagation continuity index of each driving node, calculates the training quality evaluation index. Training quality assessment indicators are used to reflect the standardization and coordination of training movements. A higher value indicates a higher degree of neural drive stability for the training movement, more standardized movement execution, and more coordinated muscle group synergy, expressed as: ; in, , These are the corresponding weighting coefficients.

[0030] like Figure 2 As shown, a deep learning-based trainer pose recognition system, used to implement the aforementioned deep learning-based trainer pose recognition method, includes the following modules: Data Acquisition and Neural Electrical Activity Reconstruction Module: Acquires ultra-wideband radio frequency imaging data of the training scene, performs inversion processing of the dielectric properties of human tissue in each frame of the ultra-wideband radio frequency imaging data, extracts microscale dielectric perturbation signals induced by neuromuscular electrical activity of human muscle tissue from the inversion results, and performs cluster analysis on the neuromuscular activity of multiple muscle groups based on the changes in the spatial distribution and temporal evolution of the microscale dielectric perturbation signals, generating a spatiotemporal map of neuromuscular electrical activity that characterizes the distribution and evolution relationship of neural drive activities of different muscle groups. Neural drive relationship analysis module: Input the spatiotemporal map of neuromuscular electrical activity into the neural activity analysis network, identify the synergistic activation relationship and its temporal sequence between different muscle groups through the muscle group association analysis network, and construct a neural synergistic drive map reflecting the human motor neural drive mechanism based on the synergistic activation intensity and drive propagation sequence between muscle groups. The posture prediction and action recognition module inputs the neural co-driving graph and the macroscopic motion structure features extracted from the training scene video stream into the posture generation prediction model. It generates constraints on the path of human posture change, thereby predicting the trajectory of the trainee's action generation. Based on the action generation trajectory, it outputs the action category recognition result that is ahead of the actual human movement and the training quality evaluation index based on neural driving stability.

[0031] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0032] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for training personnel posture recognition, characterized in that, Includes the following steps: S1. Acquire ultra-wideband radio frequency imaging data of the training scene, perform human tissue dielectric property inversion processing on each frame in the ultra-wideband radio frequency imaging data, extract microscale dielectric perturbation signals induced by neuromuscular electrical activity of human muscle tissue from the inversion results, and perform cluster analysis on neuromuscular activity of multiple muscle group regions according to the variation law of the spatial distribution and temporal evolution of the microscale dielectric perturbation signals, and generate a spatiotemporal map of neuromuscular electrical activity characterizing the distribution and evolution relationship of nerve driving activities of different muscle groups. S2, the spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The muscle group association analysis network identifies the synergistic activation relationship and its temporal sequence between different muscle groups. Based on the synergistic activation intensity and drive propagation sequence between muscle groups, a neural synergistic drive map reflecting the human motor neural drive mechanism is constructed. The neural synergistic drive map is used to describe the propagation path and drive hierarchy of different muscle group neural drives during the formation of movement in trainees. S3, the neural co-driving graph and the macroscopic motion structure features extracted from the training scene video stream are jointly input into the posture generation prediction model to generate constraints on the human posture change path, thereby predicting the action generation trajectory of the trainee. Based on the action generation trajectory, the action category recognition result that is ahead of the actual human movement and the training quality evaluation index based on neural driving stability are output.

2. The deep learning-based method for trainee pose recognition according to claim 1, characterized in that, S1 includes: S11, In the training scenario, an ultra-wideband transmitting array and a receiving array are deployed. The transmitting array transmits a wideband pulse signal to the target training area, and the receiving array collects the echo signal after penetrating the human body. The original echo of each transmitting-receiving channel is subjected to time delay correction, channel gain normalization, and background static clutter suppression to obtain ultra-wideband radio frequency imaging data of the training scenario. Based on the ultra-wideband radio frequency imaging data, the dielectric properties of human tissue are inverted for each frame to obtain the dielectric response reconstruction value of the target human body area. S12, perform differential analysis on the reconstructed dielectric response values ​​of adjacent frames to extract the time-varying microscale dielectric perturbation signal within the human muscle tissue region. Based on the prior anatomical template or the location results of human structures in the training area, the target human body is divided into multiple muscle group regions, and the perturbation response value of each muscle group region is calculated separately. , construct the first Temporal feature vectors of muscle group regions ; S13. Based on the temporal feature vectors of each muscle group region, the activity similarity between different muscle group regions is calculated. Cluster analysis is performed on multiple muscle group regions to obtain several muscle group activity clusters with synergistic change trends. Combining the spatial distribution of each muscle group activity cluster and the temporal response sequence, a spatiotemporal map of neuromuscular electrical activity is generated to characterize the distribution and evolution of neural drive activities of different muscle groups.

3. The deep learning-based method for trainee pose recognition according to claim 2, characterized in that, S11 includes: S111, Ultra-wideband transmitting and receiving arrays are deployed on both sides, front and back, or around the target training area in the training scenario. These arrays form a multi-channel penetrating detection path covering the target human body. The transmitting arrays emit broadband pulse signals into the target training area according to a preset transmission sequence, while the receiving arrays synchronously collect the raw echo signals after penetrating the human body from each transmitting-receiving channel, forming a multi-channel raw echo dataset for the current frame. ; S112, the transmit-receive channel signals in the multi-channel raw echo dataset are sequentially processed with time delay correction, channel gain normalization, and background static clutter suppression to obtain ultra-wideband radio frequency imaging data of the training scene. ; S113, based on the first The ultra-wideband radio frequency imaging data of the frame is used to perform human tissue dielectric property inversion processing on each spatial sampling point within the target training area, and the reconstructed dielectric response value of each spatial sampling point is calculated. The dielectric response distribution of the target human body region is formed by reconstructing the dielectric response values ​​of all spatial sampling points. .

4. The deep learning-based method for trainee pose recognition according to claim 1, characterized in that, S2 includes: S21, the spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The neural drive signals of each muscle group region in the spatiotemporal map of neuromuscular electrical activity are feature-encoded in the time dimension. The temporal correlation and degree of synergistic change between different muscle groups are calculated by the muscle group association analysis unit in the neural activity analysis network. In this way, the synergistic activation relationship of multiple muscle groups being activated simultaneously or continuously during the action formation process is identified, and the synergistic activation intensity matrix between each muscle group is obtained to characterize the degree of synergy of neural drive activities between different muscle groups. S22, based on the co-activation intensity matrix and the activation sequence of nerve drive signals of each muscle group in the time dimension, the drive propagation path between different muscle groups is analyzed. Each muscle group region is taken as a node in the graph structure, and the co-activation relationship and drive propagation direction between muscle groups are taken as edge relationships. The direction and weight of the edges are determined according to the temporal relationship of drive propagation, thereby constructing a neural co-drive graph that reflects the human motor nerve drive mechanism, which is used to describe the propagation path and drive hierarchy relationship of nerve drive of different muscle groups in the process of movement formation of trainees.

5. The deep learning-based method for trainee pose recognition according to claim 4, characterized in that, S21 includes: S211, Input the spatiotemporal map of neuromuscular electrical activity into the neural activity analysis network, and extract the neural drive signal sequences of each muscle group region in the time dimension from the spatiotemporal map. The neural drive signal sequence is then temporally encoded to obtain neural drive feature encoding vectors characterizing each muscle group region. ; S212 utilizes the muscle group association parsing unit in the neural activity parsing network to perform correlation analysis on the neural drive feature encoding vectors between different muscle groups, thereby calculating the temporal correlation coefficient between each muscle group. This characterizes the temporal correlation and degree of coordinated change over time. S213, based on the temporal correlation coefficient between different muscle groups, identifies the synergistic activation relationship between muscle groups. Greater than the collaborative threshold At that time, the judgment of the first The muscle group region and the first There is a synergistic activation relationship among various muscle groups during the movement formation process, and a synergistic activation intensity matrix among these muscle groups is constructed. It is used to characterize the degree of coordination of neural drive activities of different muscle groups.

6. The deep learning-based method for trainee pose recognition according to claim 5, characterized in that, S22 includes: S221, Based on the perturbation response values ​​of each muscle group region, extract the main activation time of each muscle group within the target time window. Based on the sequential relationship of the principal activation times of different muscle groups, the driving sequence between muscle groups is determined, which is used to characterize the starting and ending points of the neural driving activities of each muscle group in the time dimension during the formation of movement, and to calculate the first... The muscle group region and the first Activation time difference between individual muscle groups ,when When, it indicates the first The muscle group region is earlier than the first When individual muscle groups are activated, When, it indicates the first The muscle group region is earlier than the first Several muscle groups were activated; S222: Based on the co-activation intensity matrix and the activation time difference between different muscle group regions, muscle group pairs with co-activation relationships are screened, and the direction of the edges is determined according to the activation order, while the edge weights are determined according to the co-activation intensity and the temporal relationship. This allows for the analysis of the drive propagation pathways between different muscle groups; S223, using each muscle group region as a node in the graph structure, and the determined drive propagation edges and their weights as directed edges in the graph structure, a neural collaborative drive graph reflecting the human motor neural drive mechanism is constructed. The driving level score is calculated based on the input and output edge weights of each node. ,when At that time, the judgment of the first Each muscle group region corresponds to a driver source node, when At that time, the judgment of the first Each muscle group region corresponds to a corresponding node as a response node, where... The threshold for determining the driving level is determined by scoring the driving level of each node. By comparison, the hierarchical relationship of the driving force in different muscle groups during the movement formation process can be determined.

7. The deep learning-based method for trainee pose recognition according to claim 6, characterized in that, S3 includes: S31, the neural co-driving graph is aligned and jointly encoded with the macroscopic motion structure features extracted from the training scene video stream. The macroscopic motion structure features include the spatial distribution features of human joint nodes, the joint angle change features, and the limb movement speed features. The driving hierarchy relationship, driving propagation path, and edge weight in the neural co-driving graph are used as prior constraints and input into the posture generation prediction model. The temporal constraint unit in the posture generation prediction model constrains the motion change path of each joint node in the human body in the time dimension, so that the human posture generation process simultaneously satisfies the macroscopic motion structure law and the neural driving propagation law, thereby obtaining candidate posture change paths that conform to the neural driving mechanism. S32, based on the candidate posture change path, the motion state of each joint node of the human body in the future time window is predicted by the posture generation prediction model, and the action generation trajectory of the trainee in the future time window is generated. According to the matching result between the action generation trajectory and the preset action template library, the action category recognition result that is ahead of the actual human movement is output. At the same time, combined with the stability index of each driving node in the neural collaborative driving graph and the continuity of driving propagation between different muscle groups, the neural driving stability of the current training action is evaluated, thereby generating a training quality evaluation index that reflects the standardization and coordination of the training action.

8. The deep learning-based method for trainee pose recognition according to claim 7, characterized in that, S31 includes: S311, extract the spatial distribution features of human joint nodes, joint angle change features, and limb movement speed features from the training scene video stream, and perform spatiotemporal correspondence between the macroscopic motion structure features and muscle group nodes in the neural co-drive graph to complete the feature alignment between neural drive information and macroscopic motion information, thereby obtaining the joint-muscle group alignment feature vector. ; S312, the first The macroscopic motion structure features of the frame are jointly encoded with the joint-muscle group alignment feature vector, and the driving hierarchy, driving propagation path and edge weights in the neural co-driving graph are embedded into the joint features to obtain the pose fusion feature sequence under neural driving constraints. ; S313, The posture fusion feature sequence is input into the temporal constraint unit of the posture generation prediction model. The motion change path of each joint node of the human body in the time dimension is constrained by the neural co-driving graph to generate candidate posture change paths that satisfy the macroscopic motion structure law and the neural driving propagation law, specifically including: S3131, Construct a human posture state vector based on the spatial coordinate information of human joint nodes extracted from the video stream. , used to characterize the human body's posture state in the time dimension; S3132, the first in the pose fusion feature sequence Jointly encoded feature vector of frames The current human posture state vector is input into the temporal prediction unit of the posture generation and prediction model to predict the human posture state at the next moment. S3133, according to the neural synergistic driving diagram, the first The muscle group node and the first Edge weights between muscle group nodes Calculate neural drive constraint terms ; S3134 combines the attitude prediction error with the neural drive constraint term to construct the optimization objective function of the attitude generation and prediction model; S3135, by minimizing the objective function to solve the sequence of human posture changes, candidate posture change paths that satisfy the laws of neural drive propagation are obtained. .

9. The deep learning-based method for trainee pose recognition according to claim 8, characterized in that, S32 includes: S321, Based on the candidate posture change path, the posture generation prediction model is used to recursively predict the motion state of each joint node of the human body within a future time window, obtaining the human posture state vector at each moment. The human posture state vectors at each moment within the future time window are then combined in chronological order to generate the trainee's motion generation trajectory. Extracting action to generate trajectory feature vectors ; S322, perform similarity matching between the motion generation trajectory feature vector and the feature vectors of each motion template in the preset motion template library to determine the motion template category that best matches the motion generation trajectory, and output the corresponding motion template category as the motion category recognition result that is ahead of the actual human movement. Specifically, this includes: S3221, assuming the preset action template library contains a total of Action template, number The feature vector of the action template is Then the action-generated trajectory feature vector and the first Matching similarity between action templates; S3222, Take the action template category with the highest matching similarity as the action category recognition result. ; S323, combining the stability indices of each drive node in the neural synergistic drive diagram and the continuity of drive propagation between different muscle groups, assesses the neural drive stability of the current training movement, and generates training quality assessment indicators reflecting the standardization and coordination of the training movement based on the assessment results. Specifically, it includes: S3231, according to the... The driving characteristic values ​​of each driving node within the target time window Calculate the first Stability metrics of each driver node ; S3232, Calculate the propagation continuity index This characterizes the continuity of drive propagation between different muscle groups; S3233, based on the stability index and drive propagation continuity index of each driving node, calculates the training quality evaluation index. The training quality assessment indicators are used to reflect the standardization and coordination of training movements. The higher the value, the higher the degree of neural drive stability of the training movement, the more standardized the movement execution, and the more coordinated the muscle group synergy.

10. A deep learning-based trainee posture recognition system, used to implement the deep learning-based trainee posture recognition method as described in any one of claims 1-9, characterized in that, Includes the following modules: Data acquisition and neural electrical activity reconstruction module: Acquires ultra-wideband radio frequency imaging data of the training scene, and performs human tissue dielectric property inversion processing on each frame of the ultra-wideband radio frequency imaging data. Extracts microscale dielectric perturbation signals induced by neuromuscular electrical activity of human muscle tissue from the inversion results. Based on the variation law of the spatial distribution and temporal evolution of the microscale dielectric perturbation signals, performs cluster analysis on the neuromuscular activity of multiple muscle group regions, and generates a spatiotemporal map of neuromuscular electrical activity characterizing the distribution and evolution relationship of neural drive activities of different muscle groups. Neural drive relationship analysis module: The spatiotemporal map of neuromuscular electrical activity is input into the neural activity analysis network. The muscle group association analysis network identifies the synergistic activation relationship and its temporal sequence between different muscle groups. Based on the synergistic activation intensity and drive propagation sequence between muscle groups, a neural synergistic drive map reflecting the human motor neural drive mechanism is constructed. Posture prediction and action recognition module: The neural co-driving graph and the macroscopic motion structure features extracted from the training scene video stream are jointly input into the posture generation prediction model to generate constraints on the human posture change path, thereby predicting the action generation trajectory of the trainee, and outputting action category recognition results that are ahead of the actual human movement and training quality evaluation indicators based on neural driving stability based on the action generation trajectory.

Images