Intelligent dummy hitting action standardization recognition method based on deep learning

Abstract

The present application relates to the technical field of action recognition, and more particularly to a kind of intelligent dummy hitting action norm identification method based on deep learning, to solve the problem of strong subjectivity and low precision of hitting action norm evaluation.Method first obtains hitting action video stream, obtains three-dimensional skeleton joint space-time trajectory data by pose estimation network;Then construct human skeleton space-time graph structure, through channel topological refinement graph convolution network, fusion dynamic correlation matrix and global learning topological matrix form channel dynamic graph topology, extract cross-limb non-local collaborative feature;Subsequently sent into action specificity attention mechanism network, dynamically allocate weight in time axis and space axis, capture force phase saliency feature;Finally, with joint coordinate, skeleton vector, motion speed and other multi-stream features are fused, sent into the action quality evaluation model based on skeleton contrast learning framework, output dynamics chain coordination score and action norm quantization result, realize the accurate identification of hitting action norm.

Description

Technical Field

[0001] This invention relates to the field of action recognition technology, specifically to a method for recognizing the normative nature of striking actions by an intelligent dummy based on deep learning. Background Technology

[0002] Intelligent dummy testing typically covers a variety of single or continuous movements, including punches, kicks, knee and elbow strikes, and combinations of punches. The same dummy often has to withstand impact tasks from different testers and under different conditions. Currently, most common evaluation methods only set judgment rules based on force sensor peak values, contact duration, or score levels, and then rely on manual playback of video to verify the quality of the movements frame by frame. The entire process is cumbersome and time-consuming, and the results are highly subjective, making it difficult to meet the requirements of objective and consistent evaluation in intelligent dummy testing.

[0003] Existing motion assessment methods mostly rely on force sensor peak values, contact time, or score values ​​to establish judgment rules, and manually review video frame by frame to verify motion quality. When the same dummy performs continuous cross-subject assessments on different test subjects, static threshold rules struggle to distinguish between compensatory and non-compensatory strikes that appear similar but have different force chains. They also struggle to identify coordination deviations in power transmission between non-adjacent limbs, making it easier to confuse striking force with motion standardization. This can lead to actions with broken force chains being misjudged as standard because the end-contact force meets the standard, affecting the assessment credibility of intelligent dummy systems and the quality of test subject training feedback. Existing motion recognition methods based on skeletal graph convolutional networks mostly rely on constructing adjacency matrices using human anatomical physical connections. Feature extraction is limited to physically adjacent joints, making it difficult to capture non-local coordination features across limbs. Furthermore, since they aim at motion classification, they cannot provide continuous, standardized, quantitative assessment results.

[0004] Therefore, there is an urgent need for a method that can accurately quantify and assess the standardization of the impact actions that intelligent dummies can withstand.

[0005] To address this, a deep learning-based method for recognizing the normative nature of striking actions by intelligent dummies is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for recognizing the normative nature of striking actions using a deep learning-based intelligent dummy. This method achieves accurate quantitative evaluation of the normative nature of striking actions through topological refined graph convolution and multi-stream contrastive learning.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for recognizing the normative nature of striking actions in intelligent dummy robots includes: The video stream of the striking action is acquired and fed into the pose estimation network to obtain the spatiotemporal trajectory data of the three-dimensional skeletal joints. A spatiotemporal graph structure of the human skeleton is constructed. Based on the spatiotemporal trajectory data of the three-dimensional skeletal joints, a channel topology refined graph convolutional network is used to form a channel dynamic graph topology by using dynamic correlation matrix and global learning topology matrix to extract cross-limb non-local collaborative features between non-physically connected joints. The cross-limb non-local collaborative features are fed into the action-specific attention mechanism network, and weights are dynamically allocated on the time and space axes to capture the salient features of the force application phase. The salient features are fused with multi-stream feature data including joint coordinates, skeletal vectors, and motion velocity. The fused feature data is then fed into a motion quality assessment model trained based on a skeletal contrastive learning framework to obtain kinetic chain coordination scores and quantitative assessment results of striking motion standardization.

[0008] Preferably, the pose estimation network specifically involves: after acquiring the video stream data of the striking action, extracting the frame-by-frame two-dimensional joint coordinates and confidence scores of the human body using a preset two-dimensional pose estimation model to construct a two-dimensional skeleton sequence; feeding the two-dimensional skeleton sequence into a three-dimensional pose enhancement network based on a spatiotemporal attention mechanism, and combining the camera intrinsic parameters to predict the joint depth coordinates in three-dimensional space to obtain the spatiotemporal trajectory data of the three-dimensional skeleton joints containing the three-dimensional spatial coordinates and time dimension coordinates of each joint.

[0009] Preferably, constructing the spatiotemporal graph structure of the human skeleton includes: using three-dimensional skeletal joints as nodes of the graph structure to establish a set of joints containing a spatial dimension and a set of frames containing a temporal dimension; for the spatial dimension, constructing spatial edges based on the natural connection relationships of human anatomy to establish the physical topological relationship between physically adjacent joints within the same frame; for the temporal dimension, establishing temporal edges between adjacent video frames for corresponding nodes of the same joint at different time points to establish the spatiotemporal continuity relationship of joint evolution over time; and constructing a three-dimensional skeletal spatiotemporal graph structure representing the entire process of the striking action based on the nodes, the spatial edges, and the temporal edges.

[0010] Preferably, the extraction of cross-limb nonlocal collaborative features between non-physically connected joints through the channel topology refined graph convolutional network includes: inputting the three-dimensional skeletal spatiotemporal graph structure and the three-dimensional skeletal joint spatiotemporal trajectory data into the channel topology refined graph convolutional network; adaptively capturing real-time motion associations between joints using a dynamic correlation matrix, and learning global motion prior topology using a global learning topology matrix; fusing the dynamic correlation matrix and the global learning topology matrix to generate independent channel dynamic graph topologies for different feature channels; and using the channel dynamic graph topology to establish nonlocal feature channel links for non-anatomically physically connected cross-limb joint pairs, and extracting cross-limb nonlocal collaborative features of different limb joints during the striking force application process.

[0011] Preferably, the action-specific attention mechanism network includes a temporal attention unit, a spatial attention unit, and a spatiotemporal feature fusion unit. The temporal attention unit receives the cross-limb nonlocal collaborative features, calculates the correlation weights of different video frames on the time axis of the striking action, generates a temporal attention weight map, and performs temporal weighting on the cross-limb nonlocal collaborative features to establish the temporal salience of the key phase of striking force. The spatial attention unit calculates the correlation weights of different joints on the human body's spatial structure, generates a spatial attention weight map, and performs spatial weighting on the cross-limb nonlocal collaborative features to establish the spatial salience of the core joint of striking force. The spatiotemporal feature fusion unit performs element-wise multiplication fusion of the temporal salience features weighted by the temporal attention unit and the spatial salience features weighted by the spatial attention unit to extract the salience features of the force phase.

[0012] Preferably, the motion quality assessment model includes a positive and negative sample construction unit, a contrastive representation learning unit, a kinetic chain coordination evaluation unit, and a motion standardization quantification regression unit. Specifically, the positive and negative sample construction unit uses standard striking motion data as anchor samples and non-standard striking motion data generated after perturbation of the standard striking motion data as negative samples to construct skeletal motion contrastive learning sample pairs. The contrastive representation learning unit receives the fused feature data and calculates the feature distances between anchor samples, positive samples, and negative samples based on the contrastive loss function within the contrastive learning framework. It then minimizes the feature distances of similar samples and maximizes the feature distances of dissimilar samples through gradient backpropagation, mapping to obtain a higher-order motion representation vector. The kinetic chain coordination evaluation unit feeds the higher-order motion representation vector into a preset kinetic chain evaluation branch, calculates the phase delay and energy transfer continuity of each joint in the human body during force transmission, and provides a kinetic chain coordination score. The motion standardization quantification regression unit concatenates the higher-order motion representation vector with the kinetic chain coordination score and maps it through a regression network to obtain the quantitative evaluation result of the striking motion standardization.

[0013] Preferably, the saliency feature is fused with multi-stream feature data including joint coordinates, bone vectors, and motion velocity, including: In the spatiotemporal trajectory data of 3D skeletal joints, the 3D absolute coordinate flow features corresponding to each joint, the physical skeleton vector flow features composed of adjacent joints, and the motion velocity flow features between adjacent frames of the joints are obtained respectively. The above three types of flow features are temporally aligned by a multi-scale one-dimensional convolutional network. The aligned multi-flow feature data and salient features are concatenated along the channel dimension, and cross-modal feature fusion is performed by pixel-by-pixel convolution to obtain the fused feature data.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a human skeletal spatiotemporal graph structure from 3D skeletal joint spatiotemporal trajectory data. It adaptively captures real-time motion associations between joints using a dynamic correlation matrix and learns global motion prior topology using a global learning topology matrix. By fusing these two methods, it generates independent channel dynamic graph topologies for different feature channels. Non-local feature channel links are established for cross-limb joint pairs that are not physically connected in human anatomy, extracting cross-limb non-local collaborative features of different limb joints during the striking force application process. This approach enables feature extraction to overcome the limitations of physical connections in human anatomy, incorporating the dynamic transmission and collaborative patterns between non-adjacent limbs, such as the supporting leg-striking arm and the contralateral trunk-force application joint, into the graph convolutional receptive field. This avoids the omission of cross-limb collaborative information caused by modeling solely based on physical skeletal adjacency relationships, while preserving the differentiated expressive ability of different feature channels for motion associations, improving the completeness of striking action representation and the joint coverage of force chain modeling.

[0015] 2. This invention first feeds cross-limb non-local collaborative features into an action-specific attention mechanism network: on the time axis, the temporal attention unit identifies the salience of the key phase of striking force; on the spatial axis, the spatial attention unit identifies the salience of the core force-generating joints; the two results are then fused by element-wise dot product by the spatiotemporal feature fusion unit. Subsequently, the obtained salience features are aligned temporally with the three data streams—joint coordinate stream, skeletal vector stream, and motion velocity stream—through multi-scale one-dimensional convolution, and then concatenated and fused according to the channel dimension. In this way, the force-generating keyframes and core force-generating joints can obtain differentiated response weights, while complementary modal information such as the absolute spatial position of the joints, limb geometry, and instantaneous motion velocity are also preserved. This avoids the dilution of instantaneous force-generating features by global average pooling and the representation degradation of single-stream features when occluded or self-overlapping, thus improving the temporal sensitivity and multimodal robustness of the evaluation.

[0016] 3. This invention uses standard striking motion data as anchor samples and non-standard motion data generated after perturbing the standard motion as negative samples to form sample pairs for skeletal motion comparative learning. Within the comparative learning framework, the feature distance of similar samples is brought closer together, while the feature distance of dissimilar samples is pushed further apart, thus mapping a higher-order motion representation vector. This vector is then fed into the kinetic chain evaluation branch to calculate the phase delay and energy transfer continuity of each joint during force transmission, obtaining a kinetic chain coordination score. Finally, the higher-order motion representation vector and this coordination score are concatenated and passed through a regression network to obtain a quantitative evaluation result of the striking motion's standardization. The advantages of this approach are that motion quality evaluation can capture the compensatory differences among similar striking motions that have similar appearances but different force chain qualities. Furthermore, it replaces discrete motion classification labels with continuous kinetic chain coordination scores and standardization scores, avoiding the shortcomings of classification models that only identify motion types but cannot distinguish between the quality of completion. Simultaneously, it embeds the physical constraints of force transmission into the evaluation indicators, thereby improving the scoring resolution and cross-tester consistency of the intelligent dummy evaluation system. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for recognizing the normative nature of intelligent dummy striking actions based on deep learning proposed in this invention. Figure 2 This is a flowchart of the intelligent dummy striking motion norm recognition method based on deep learning proposed in this invention. Figure 3 This is a network structure diagram of the action-specific attention mechanism network proposed in this invention. Detailed Implementation

[0018] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 This embodiment uses a smart dummy impact testing platform as an example to illustrate the complete implementation process of the method described in this invention. The testing platform includes three synchronously triggered RGB cameras with a resolution of 1920×1080 and a frame rate of 60 Hz, arranged at a 120° angle around the test area. The shooting distance between the cameras and the center point of the dummy is 3.2 m. A separate set of three-axis force sensors is embedded in the chest-abdomen connection of the dummy to collect the triaxial contact force at the moment of impact. This sensor serves as a separate reference signal for motion standardization and is not involved in model inference. The dummy's outer layer is a silicone rubber skin, filled with EVA foam material. The joints are connected by ball joints, providing basic range of motion for the shoulder, elbow, hip, and knee. The camera intrinsic parameter matrix and distortion coefficients are obtained using the Zhang Zhengyou calibration method. The calibration board is an 11×8 grid pattern with a side length of 30 mm. Each set of cameras captures 16 different poses, which are then solved using the OpenCV calibrateCamera function. Human skeletal joints were defined using COCO 17, including the nose, both eyes, both ears, both shoulders, both elbows, both wrists, both hips, both knees, and both ankles. The test covered four basic striking techniques: straight punch, hook punch, front kick, and sweeping kick. Three participants with a Level 2 or higher Sanda athlete certificate each performed 30 repetitions of each technique. Three amateur participants without athlete certificates performed the same techniques as a non-standard control group. Each strike was initiated from the starting position after the participant heard the electronic trigger tone and returned to the starting position after the strike was completed.

[0020] Please see Figures 1 to 3 This invention provides a method for recognizing the normative nature of striking actions by an intelligent dummy based on deep learning. The technical solution is as follows: A deep learning-based method for recognizing the normative nature of striking actions in intelligent dummies, such as Figures 1-2 As shown, it includes: The video stream of the striking action is acquired and fed into the pose estimation network to obtain the spatiotemporal trajectory data of the three-dimensional skeletal joints. A spatiotemporal graph structure of the human skeleton is constructed. Based on the spatiotemporal trajectory data of the three-dimensional skeletal joints, a channel topology refined graph convolutional network is used to form a channel dynamic graph topology by using dynamic correlation matrix and global learning topology matrix to extract cross-limb non-local collaborative features between non-physically connected joints. The cross-limb non-local collaborative features are fed into the action-specific attention mechanism network, and weights are dynamically allocated on the time and space axes to capture the salient features of the force application phase. The salient features are fused with multi-stream feature data including joint coordinates, skeletal vectors, and motion velocity. The fused feature data is then fed into a motion quality assessment model trained based on a skeletal contrastive learning framework to obtain kinetic chain coordination scores and quantitative assessment results of striking motion standardization.

[0021] Furthermore, the pose estimation network specifically involves: after acquiring the video stream data of the striking action, extracting the frame-by-frame two-dimensional joint coordinates and confidence scores of the human body using a preset two-dimensional pose estimation model to construct a two-dimensional skeleton sequence; feeding the two-dimensional skeleton sequence into a three-dimensional pose enhancement network based on a spatiotemporal attention mechanism, and combining the camera intrinsic parameters to predict the joint depth coordinates in three-dimensional space to obtain the spatiotemporal trajectory data of the three-dimensional skeleton joints containing the three-dimensional spatial coordinates and time dimension coordinates of each joint.

[0022] The preset two-dimensional pose estimation model uses HRNet-W48 pre-trained on the COCO 2017 keypoint detection dataset. This model uses high-resolution representation as the backbone, maintains high-resolution branches throughout the feature extraction process, and fuses multi-scale features step by step. It has high localization accuracy for small-scale joints (such as wrists and ankles) under self-occlusion and low-light conditions.

[0023] The model's pre-training weights come from the MMPose official model library. The parameters for the first three stages are frozen, and only the keypoint detection head for the last stage is fine-tuned using the dummy impact video frames collected in this embodiment. The 3D pose enhancement network consists of two parts: a temporal encoder and a pose decoder. On the temporal encoder side, the coordinates of each joint point in each frame of the 2D skeleton sequence are first flattened into a 34-dimensional vector, and then fed into a 4-layer Transformer encoder along the frame dimension. Each layer of the encoder is equipped with an 8-head self-attention mechanism and a feedforward network. The hidden layer width of the feedforward network is set to 1024, and finally, a temporal context embedding vector is output for each joint point.

[0024] After obtaining the embedding vector, the pose decoder first regresses the depth values ​​Z of each joint in the camera coordinate system through a fully connected layer, and then uses the camera intrinsic parameter matrix to convert the pixel coordinates... Together with the predicted depth Z, it is back-projected back to the three-dimensional spatial coordinates. .

[0025] The back projection formula is: ; ; in , For camera focal length parameters, , The coordinates of the principal point are obtained from the aforementioned calibration.

[0026] During the training of the 3D pose enhancement network, the coordinates of real 3D joints on the ground are used as the supervision signal. The loss function is the L2 norm between the predicted 3D coordinates and the ground truth. Training stops when the loss on the validation set no longer decreases for 5 consecutive epochs. In this embodiment, each video is continuously acquired at 60Hz, and the average duration of the striking action from the starting posture to the return to the starting posture is about 1.8 seconds. Therefore, 108 frames are extracted as the input sequence for each striking action. The three video streams are independently fed into the pose estimation network to obtain three sets of spatiotemporal trajectory data of 3D skeletal joints. The arithmetic mean of the three sets of coordinates for each joint is taken as the fused coordinate. When the confidence score of a joint at a certain viewpoint is lower than 0.3, the component of that viewpoint is not included in the averaging. This method enhances the multi-view 2D skeleton to 3D through spatiotemporal attention and then performs weighted fusion, avoiding the depth ambiguity of monocular estimation in self-occlusion scenes and improving the spatial accuracy and multi-view consistency of 3D skeletal joint trajectory.

[0027] Furthermore, constructing the spatiotemporal graph structure of the human skeleton includes: using three-dimensional skeletal joints as nodes of the graph structure, establishing a set of joints containing a spatial dimension and a set of frames containing a temporal dimension; for the spatial dimension, constructing spatial edges based on the natural connection relationships of human anatomy, establishing the physical topological relationship between physically adjacent joints within the same frame; for the temporal dimension, establishing temporal edges between adjacent video frames for the corresponding nodes of the same joint at different time points, establishing the spatiotemporal continuity relationship of the joint evolution over time; and constructing a three-dimensional skeletal spatiotemporal graph structure representing the entire process of the striking action based on the nodes, the spatial edges, and the temporal edges.

[0028] Specifically, a set of joints is constructed based on the COCO 17 joint definition: ; in The frame number, This refers to the total number of frames in a single hit animation. This is an index for joints. The natural connections in human anatomy refer to the physical connections between adjacent joints within the skeleton, formed by skeletal segments: nose-eyes-ears form the head chain, shoulders-elbows-wrists form the arm chain, hips-knees-ankles form the leg chain, and shoulders-hips are represented virtually by the trunk. Spatial boundary sets. Includes the following Undirected edges: nose-left eye, nose-right eye, left eye-left ear, right eye-right ear, left shoulder-right shoulder, left shoulder-left elbow, left elbow-left wrist, right shoulder-right elbow, right elbow-right wrist, left shoulder-left hip, right shoulder-right hip, left hip-right hip, left hip-left knee, left knee-left ankle, right hip-right knee, right knee-right ankle. Compared to the standard COCO skeleton, a left shoulder-right shoulder edge is added here to introduce synchronous linkage constraint between the two shoulders, and a left hip-right hip edge is added to capture the opening and closing posture of the pelvis.

[0029] Time edge set The method of establishment is: for each joint point In the adjacent number Frame and the Between frames, connect two nodes at the same joint point and A directed temporal edge is formed, with the direction uniformly pointing from the previous frame to the next. In terms of weighting, the spatial edge uses the normalized L2 norm of the corresponding bone vector to reflect the spatial distance between adjacent joints in the current frame; the temporal edge uses the L2 norm of the cross-frame displacement of the same joint to reflect the motion amplitude of that joint between two frames. In this embodiment, the number of frames T is 108, and the total number of nodes... There are 108 × 19 = 2052 spatial edges and 107 × 17 = 1819 temporal edges. The completed 3D skeletal spatiotemporal graph structure is fed into the subsequent graph convolutional network in the form of a combination of an adjacency matrix and a node feature matrix.

[0030] This approach unifies the anatomical connections of the human body with the temporal relationships between frames into a single spatiotemporal graph, and uses the spatial distance of skeletal vectors and the cross-frame motion amplitude to assign weights to the edges, thereby avoiding the loss of limb geometric information by the binary adjacency matrix and improving the fidelity of the graph structure in representing the spatiotemporal evolution of the striking action.

[0031] Furthermore, the non-local collaborative features between non-physically connected joints are extracted through the channel topology refined graph convolutional network, including: inputting the three-dimensional skeletal spatiotemporal graph structure and the three-dimensional skeletal joint spatiotemporal trajectory data into the channel topology refined graph convolutional network; adaptively capturing the real-time motion association between joints using a dynamic correlation matrix, and learning the global motion prior topology using a global learning topology matrix; fusing the dynamic correlation matrix and the global learning topology matrix to generate independent channel dynamic graph topologies for different feature channels; and using the channel dynamic graph topology to establish non-local feature channel links for non-anatomically physically connected cross-limb joint pairs, and extracting the non-local collaborative features between different limb joints during the striking force application process.

[0032] Specifically, the channel topology refined graph convolutional network consists of three stacked graph convolutional layers, with each layer outputting 64, 128, and 256 feature channels respectively. Each layer is followed by Batch Normalization and ReLU activation, and Dropout is added between layers at a ratio of [ratio missing]. Dropout layer. Dynamic association matrix. The acquisition process is as follows: extract the features of the input nodes in the current layer. Calculate any two nodes and eigenvectors and Normalized dot product similarity between them: ; in and It is a learnable linear transformation matrix. Scaling factor along Direction normalization, making The sum of the elements in each row is 1; this matrix adaptively captures the real-time correlation strength between any two joints in the same striking motion based on the current motion pattern. Global learning topology matrix. Initialized to obey A Gaussian distributed random matrix, which serves as a learnable parameter of the network and is updated during training, encodes the general action prior topology between keypoints in each striking subject in the training set after convergence. During fusion, for each feature channel... Calculate separately: ; in For channel-specific learnable scalar parameters, Function constraints to Within the interval. For non-physically connected cross-limb joint pairs. ,like Exceeding the preset connection threshold In this channel A non-local feature channel link is established, and the feature aggregation of the key points is incorporated into the message passing of the graph convolution in the l-th layer.

[0033] Among non-physically connected cross-limb joint pairs, typical examples include right shoulder-left hip, right wrist-left ankle, and right elbow-left knee, all of which exhibit strong cooperative motion phase locking during straight punches. After the input is processed by a three-layer channel-topology refined graph convolutional network, the output features are globally averaged along the channel dimension to obtain the cross-limb non-local cooperative feature vectors for each joint in each frame. This method adaptively fuses the dynamic correlation matrix and the global prior topology along the channel, establishing non-local feature channel links for cross-limb joint pairs. This compensates for the shortcomings of standard skeletal graph convolution, which only covers physically connected joint pairs and misses cross-limb cooperative information, resulting in more complete feature coverage of the power transmission link during a punch.

[0034] Furthermore, such as Figure 3 As shown, the action-specific attention mechanism network includes a temporal attention unit, a spatial attention unit, and a spatiotemporal feature fusion unit; Specifically, the temporal attention unit features cross-limb nonlocal collaborative characteristics. As input, first perform max pooling along the keypoint dimension to obtain time series features. Then Feed into a two-branch structure: local branches are based on the kernel size. One-dimensional convolutions slide along the time axis to extract local temporal patterns, and global branches are based on the convolution kernel size. Channel-level one-dimensional convolution achieves the receptive field of the entire sequence. The outputs of the two branches are broadcast and then added together. Function generates temporal attention weight graph Spatial attention unit for input Spatial features are obtained by performing average pooling along the frame dimension. Calculate the key points With key points Multi-head self-attention weights between: ; A four-head attention mechanism is used, with each head calculating independently and then concatenating the results. The attention values ​​of all j are summed over i. Compression yields the spatial attention weight map Time-domain weighted operation: Spatial weighted operation: ,in Element-wise multiplication is broadcast along the corresponding dimension. Spatiotemporal feature fusion unit pairs. Perform element-wise dot product fusion: The dot product operation further enhances the feature sites that simultaneously acquire high temporal and spatial weights, while suppressing sites with only high single-dimensional weights. This allows for the extraction of salient coupling features between keyframes and core joints in the striking motion. The force phase is defined as the continuous frame interval from the start frame of the body's center of mass shifting forward to the frame where the fist / foot makes contact with the dummy's skin. The starting frame of the center of mass shifting forward is determined by the frame where the velocity component of the midpoint between the left and right hips in the sagittal plane first exceeds 0.15 m / s. The contact frame is determined by the frame where the axial force first exceeds 5 N as recorded by the dummy's built-in three-part force sensor. Taking a straight punch as an example, the force phase duration of a straight punch is approximately 17–21 frames. Within this interval, the four joints of the right wrist, right elbow, right shoulder, and left hip acquire high spatial attention weights, and the corresponding frame intervals acquire high temporal attention weights. These joints precisely constitute the power transmission path for the straight punch.

[0035] This method captures keyframes and core joints of force application using dual-branch temporal attention and multi-head spatial attention respectively, and enhances the coupling saliency features through dot product fusion. This avoids the dilution of instantaneous differences in force application caused by global pooling, which would otherwise overwhelm the normalized fine-grained discrimination information, thereby improving the spatiotemporal positioning accuracy and saliency feature discrimination power of the force application phase.

[0036] Furthermore, the motion quality assessment model includes a positive and negative sample construction unit, a contrastive representation learning unit, a kinetic chain coordination evaluation unit, and a motion standardization quantification regression unit. The positive and negative sample construction unit uses standard striking motion data as anchor samples and non-standard striking motion data generated after perturbation of the standard striking motion data as negative samples to construct skeletal motion contrastive learning sample pairs. The contrastive representation learning unit receives the fused feature data and calculates the feature distances between anchor samples, positive samples, and negative samples based on the contrastive loss function within the contrastive learning framework. It then minimizes the feature distances of similar samples and maximizes the feature distances of dissimilar samples through gradient backpropagation, mapping to obtain a higher-order motion representation vector. The kinetic chain coordination evaluation unit feeds the higher-order motion representation vector into a preset kinetic chain evaluation branch, calculates the phase delay and energy transfer continuity of each joint in the human body during force transmission, and provides a kinetic chain coordination score. The motion standardization quantification regression unit concatenates the higher-order motion representation vector with the kinetic chain coordination score and maps it through a regression network to obtain the quantitative evaluation result of the striking motion standardization.

[0037] Specifically, the standard striking action data in the positive and negative sample construction unit is taken from the striking records of three athletes in various subjects. Strikes where the peak force recorded by the three force sensors in each subject is within ±5% of the subject's historical peak force in that subject are used as standard samples. Standard samples are paired with each other as positive samples. The perturbation methods include the following six categories, each generating an equal number of negative samples: Firstly, temporal perturbation—randomly cropping or inserting 1 to 4 frames into the skeletal sequence and performing linear interpolation resampling to simulate the deviation in striking rhythm; Secondly, spatial perturbation—the superposition of the three-dimensional coordinates of a single key point follows a certain order. Gaussian noise, Take 0.05 times the bone length at that joint point to simulate joint vibration; Third, dynamic chain disturbance—the velocity vector directions of adjacent joints are swapped within the force application phase, resulting in an incorrect simulated force application sequence; Fourth, posture perturbation—replacing the angles of several key points in the initial striking frame with the corresponding angles of amateur subjects in the same subject to simulate the deviation of the initial posture; Fifth, center of mass perturbation—a low-frequency sinusoidal offset with a peak value of 2% of the total bone height is superimposed along the trajectory of the human body's center of mass to simulate body swaying; Sixth, amplitude perturbation—multiplying the displacement of the distal joint of the wrist or ankle in the direction of force by a random factor of 0.7 to 0.9, the simulation of insufficient force exertion.

[0038] The above perturbations can be applied in combination within the same skeletal sequence, but no more than two classes are allowed. When combining, the parameters of each perturbation are independently and randomly sampled. The contrastive loss function uses... form: ; in For cosine similarity, For temperature coefficient, This represents the total number of negative samples within a single batch. This is a 128-dimensional normalized vector after being mapped by the projection head. The projection head is a two-layer fully connected network with a hidden layer dimension of 256 and an output dimension of 128, activated by ReLU in between. The preset kinetic chain evaluation branch takes the higher-order motion representation vector as input. This branch is a three-layer fully connected network with output dimensions of 128, 64, and 2 respectively, and the last layer is deactivated. The meaning of the two output channels is as follows: Channel 1 outputs the force phase delay of each joint relative to the striking end (wrist or ankle). The physical meaning is key point. Peak speed moment Peak velocity at the end joint Difference: ; A positive value indicates that the joint reaches peak velocity before the distal end. In standard power transmission, the proximal hip and shoulder... The value should be a relatively large positive value, followed by the elbow and knee at the distal end, and zero at the wrist and ankle at the distal end; the energy transfer efficiency of the adjacent joint link output by channel two. Defined as a key point Kinetic energy increment and key points The absolute value of the ratio of kinetic energy decay: ; in Joint quality The scores were scaled down to the subject's height and weight. For example, for a subject who is 1.75 m tall and weighs 70 kg, the mass of their forearm and hand would be 2.52% of their body weight, or 1.76 kg. (This refers to the overall kinetic chain coordination score.) Defined as a weighted average of the energy transfer efficiency across all adjacent node links: ; The number of joint pairs in the kinetic chain is 10 pairs in this embodiment. Furthermore, the dispersion of phase delay is included as a penalty in the coordination score: if... Alternating signs between adjacent joints along the kinetic chain indicate either force retraction or a secondary force application midway through the process. Subtract 0.15. The regression network is a three-layer fully connected network, with output dimensions of 64, 32, and 1 respectively. The first two layers are connected to ReLU activation and Dropout rate = 0.1, and the last layer has no activation and outputs directly. The continuous normative scoring is used, with manually labeled normative scores as the supervision signal during training, and the loss function is Huber Loss. The score is set to 5.0. The manual annotation standard is as follows: Three judges holding Sanda referee certificates independently watch the synchronized three-view replay video of each strike. They score the strikes according to four dimensions: kinetic chain coherence (40 points), end-effector trajectory accuracy (25 points), posture stability (20 points), and rhythmic accuracy (15 points). The scores are then summed, and the median score of the three judges is used as the action accuracy annotation value for that strike, which is used as the supervision signal for training the regression network. Model training is divided into two phases: the first phase trains only the contrastive representation learning unit and the kinetic chain evaluation branch until the InfoNCE loss converges; the second phase freezes the layers before the projection head of the contrastive representation learning unit and jointly trains the kinetic chain evaluation branch and the regression network until the Huber Loss does not decrease for 5 consecutive epochs on the validation set.

[0039] This approach constructs negative samples using six types of perturbations and learns higher-order representations using contrastive loss. After quantifying phase delay and energy transfer efficiency through kinetic chain branching, it outputs continuous normative scores via a regression network. This avoids the situation where classification models can only identify the type of action but cannot quantify the quality of completion. At the same time, it embeds the physical constraints of force transmission into the evaluation index, improving the resolution and cross-subject consistency of the normative scores.

[0040] Furthermore, the salient features are fused with multi-stream feature data including joint coordinates, bone vectors, and motion velocity, including: In the spatiotemporal trajectory data of 3D skeletal joints, the 3D absolute coordinate flow features corresponding to each joint, the physical skeleton vector flow features composed of adjacent joints, and the motion velocity flow features between adjacent frames of the joints are obtained respectively. The above three types of flow features are temporally aligned by a multi-scale one-dimensional convolutional network. The aligned multi-flow feature data and salient features are concatenated along the channel dimension, and cross-modal feature fusion is performed by pixel-by-pixel convolution to obtain the fused feature data.

[0041] Specifically, three-dimensional absolute coordinate flow features Directly extract the key points in each frame Coordinates are normalized with the center of the dummy base as the origin. The normalization factor is the reciprocal of the subject's height H to eliminate the influence of differences in body shape among different subjects. Physical skeletal vector flow characteristics. It consists of directed skeletal segment vectors corresponding to the aforementioned 19 spatial edges, and the skeletal vectors... In the code implementation, the unit direction vector (3D) is concatenated with the L2 norm of the bone segment length to form a 4D feature, which is then projected onto 64D. Motion velocity flow features. Bit removal of key points between adjacent frames to frame interval Find: ; in With the velocity of the last frame assigned to a zero vector, the peak instantaneous velocity at the point of impact can reach 8 to 12 m / s.

[0042] The multi-scale one-dimensional convolutional network consists of three parallel temporal convolutional branches: branch one has a kernel size of 3, branch two has a kernel size of 5, and branch three has a kernel size of 7. Each branch has 64 output channels. The outputs of the three branches are concatenated along the channel dimension to obtain a 192-dimensional multi-scale temporal aligned feature.

[0043] The basis for multi-scale design is that, in striking actions, the momentum changes of rapid joints such as the wrist occur within a short time window (3 to 5 frames), making small convolutional kernels suitable for capturing instantaneous acceleration features; while the posture adjustments of trunk joints such as the hip take longer (more than 10 frames), making large convolutional kernels suitable for smoothing and extracting slow-changing displacement trends. The pixel-by-pixel convolution refers to using the convolutional kernel size as... Two-dimensional convolution operates along the feature channel direction, its function being to maintain the spatiotemporal dimensions without changing the spatiotemporal size. Under the premise of learning the optimal linear combination of three-stream features and saliency features across channels, the number of operational parameters is: , Let 256 be the number of feature channels fed into the contrastive representation learning unit after fusion. Before concatenating multi-stream features and saliency features by channel dimension, the channel number of each type of feature is projected to 64 dimensions through a fully connected layer, so that the total number of channels after concatenation is... Then, pixel by pixel Convolutional compression reduces the dimension to 256. Each spatiotemporal location of the fused features... It contains a 256-dimensional vector that encodes the key points. In frame The absolute spatial position, the instantaneous configuration of adjacent bone segments, the speed of movement, and the salience information of the force phase from the action-specific attention mechanism complement each other to form a complete multimodal representation of the striking action.

[0044] This approach aligns the three-stream temporal features with multi-scale one-dimensional convolution and then performs cross-modal fusion using pixel-by-pixel convolution. This avoids the degradation of single-stream features in scenarios with rapid limb overlap or short-term self-occlusion, and improves the modal complementarity and temporal modeling robustness of the fused features.

[0045] This invention achieves accurate quantitative evaluation of the standardization of striking actions through topological refined graph convolution and multi-stream comparative learning.

[0046] Example 2 This embodiment is implemented on the intelligent dummy hitting test platform described in Embodiment 1, illustrating three processes: skeletal normalization and migration adaptation for test subjects of different body types, automatic segmentation and subject attribution determination for continuous hitting of multiple subjects, and joint completion and robust reasoning in self-occlusion scenarios.

[0047] Three synchronously triggered RGB cameras capture video streams of the impact action, which are then fed into an HRNet-W48 2D pose estimation model to obtain frame-by-frame coordinates and confidence scores for 17 key points. These coordinates are then enhanced to 3D spatiotemporal trajectories of the skeletal joints via a 4-layer Transformer temporal encoder and a fully connected pose decoder. The three-view confidence scores are then weighted and averaged for fusion. (Please refer to COCO for more details.) A spatiotemporal graph structure of the human skeleton is constructed using 17 key points, 19 spatial edges, and temporal edges of adjacent frames, with edge weights assigned based on 3D coordinates and bone displacement. A three-layer channel topology-refined graph convolutional network is used to generate a channel dynamic graph topology through adaptive fusion of a dynamic correlation matrix and a global learning topology matrix. Non-local feature links are established for cross-limb key point pairs to extract collaborative features. Temporal dual-branch attention and spatial multi-head attention are then fed into the network to capture saliency features of the force phase. After dot-multiplication fusion, these features are concatenated with three types of multi-flow features: coordinate flow, bone vector flow, and velocity flow. A pixel-by-pixel 1×1 convolution is used for cross-modal fusion to achieve 256 dimensions. The fused features are fed into a skeleton contrastive learning framework that constructs negative samples with six types of perturbations to map higher-order action representations. The phase delay and energy transfer efficiency are calculated through the kinetic chain evaluation branch to obtain the kinetic chain coordination score. After concatenation with the higher-order representations, a three-layer fully connected regression network outputs a 0-100 normalization score.

[0048] Furthermore, the skeletal normalization and transfer adaptation process for cross-body type test subjects: After the contrastive representation learning unit and regression network of the motion quality assessment model described in Example 1 were trained using hitting data from three athlete-level subjects and three amateur subjects, when a new subject used the intelligent dummy testing platform, their height, limb length, and weight differed from those of the training set subjects. Directly inputting three-dimensional skeletal coordinates would introduce a body type distribution shift unrelated to motion standardization, causing the standardization score output by the regression network to systematically deviate from the labeled distribution during training. Therefore, after the pose estimation network outputs three-dimensional skeletal joint spatiotemporal trajectory data and before constructing the human skeletal spatiotemporal map structure, a skeletal normalization step is added. This maps the three-dimensional skeletal joint coordinates of each subject to a unified reference skeletal coordinate system, and simultaneously provides the normalized score and original body type parameters in the report to the subject after the evaluation is completed.

[0049] Specifically, the unified reference skeleton was based on the static standing posture skeleton of the subject with the median height among the six subjects in the training set. The average coordinates of each joint point were taken from 50 frames of skeleton data with the subject's feet together and arms hanging naturally at their sides as the reference template. Skeletal normalization was performed in two steps: The first step is scale normalization—taking the trunk height from the midpoint between joint 1 to joints 9 and 10. The shoulder width from the midpoint of joints 6 and 7 to the midpoint of joints 5 and 6. As a scale factor unique to the subject, all three-dimensional joint coordinates are calculated according to... and After dividing by the corresponding value of the reference template, the average scaling ratio is taken to eliminate the linear differences in body shape (height, weight, and build).

[0050] The second step, posture alignment—using the center of the pelvis as the origin and the vector from the midpoint of both shoulders to the center of the pelvis as the torso's longitudinal axis, calculates the rotation matrix between the subject's current frame's torso's longitudinal axis and the reference template's torso's longitudinal axis. A Procrustes orthogonal transformation is then used to rotate the subject's skeleton to the orientation of the reference template, eliminating the subject's orientation relative to the camera. Posture alignment uses only the torso's longitudinal axis and not the limb joints to avoid the Procrustes transformation incorrectly rotating and eliminating the force posture information when the limbs are already in a force-generating posture during a striking action. The normalized 3D skeletal joint spatiotemporal trajectory data continues to be fed into subsequent human skeletal spatiotemporal map structure construction and all downstream modules.

[0051] In terms of transfer adaptation, the normalized skeletal data of new subjects is first used to perform forward inference using the previously trained contrastive representation learning units to obtain higher-order action representation vectors. These higher-order representations share the same embedding space as the higher-order representations of the subjects in the training set. Subsequently, if the new subjects agree to use their data for model adaptation, after completing 10 consecutive strike normativity assessments and obtaining corresponding manually labeled scores, the last layer of the regression network is fine-tuned using these 10 samples. The fine-tuning uses the AdamW optimizer and a learning rate of [missing information]. The model was trained for 10 epochs to adaptively compensate for the body size differences of the new subject. After fine-tuning, the model retained copies of the original training weights, while the weights specific to the new subject were stored independently. In this embodiment, a new subject who was 1.88 m tall and weighed 92 kg completed a straight punch standardization assessment using normalized skeletal data. Before normalization, the Euclidean distance from the pelvic center to the wrist joint was about 13% longer than the digits in the training set. The standardization score before normalization was systematically lower by about 6.2 points. After normalization, the score deviation was reduced to less than 1.1 points, and after 10 fine-tuning sample adaptations, it was further reduced to less than 0.5 points.

[0052] Furthermore, the automatic segmentation and subject classification process for continuous multi-subject strikes: The method described in Example 1 performs standardized evaluation based on a single strike action. When a subject continuously performs multiple subjects such as straight punch, hook punch, front kick, and sweep kick in the same video clip, the video stream needs to be pre-segmented into independent single strike segments, and each segment needs to be labeled with its subject classification before it can be sent to the subsequent evaluation process. To address this requirement, this embodiment adds a preprocessing step for continuous strike segmentation after acquiring the three-dimensional skeletal joint spatiotemporal trajectory data. The start and end frames of each strike are automatically detected based on the sagittal plane displacement velocity of the pelvic center and the combined velocity of the distal joint. The detection method for the start and end frames of a strike segment is based on the combination of the velocity vector direction at the end of the strike and the force-generating joint. The pelvic center is defined as the average three-dimensional coordinates of the left and right hip joints. The displacement difference between adjacent frames of the pelvic center in the sagittal plane is calculated and divided by the frame interval. The sagittal velocity of the pelvis was obtained. .when The moment the baseline first rises above the 0.12 m / s threshold, it is marked as the start of the hit frame. Subsequently, the combined velocity at the distal joints was monitored—the maximum value of the L2 norm of the wrist velocity vector and the maximum value of the L2 norm of the ankle velocity vector were taken. The resultant velocity was considered when it fell from its peak value to below 30% of its peak value. When the frame falls back to the baseline level, it is marked as the end of the hit. The baseline value is automatically updated during the static standing period before each strike. The update method is to skip 12 frames after the end frame of the previous strike and take the average sagittal velocity of the pelvis in the following 10 frames as the baseline for the next strike. The method for determining the subject classification is as follows: within the force phase interval of the strike segment, extract the direction angle of the velocity vector of the distal joint in the horizontal plane. And the average kinetic energy percentage of each joint point within the force application phase. The criteria for judging a straight punch are: Within ±22.5°, and with the ratio of right or left wrist kinetic energy to total distal end kinetic energy exceeding 0.70; the criteria for determining a hook punch are: The absolute values ​​range from 22.5° to 67.5°, with the wrist kinetic energy accounting for more than 0.60 and the shoulder kinetic energy accounting for a significantly higher percentage than the straight punch. The criteria for determining a proper leg drive are: the maximum descent of the vertical displacement of the pelvic center within the force application phase exceeds 15% of the subject's lower leg length, and the direction angle of the velocity vector of the unilateral ankle is [not specified]. Within ±30°, the ankle kinetic energy accounts for more than 0.65%; the criteria for judging a sweeping leg are: The absolute value exceeds 67.5°, the ankle kinetic energy ratio exceeds 0.60, and the force phase duration is significantly longer than the normal kick. If a segment does not meet any of the above criteria, it is marked as pending manual review and will not be sent to the automatic evaluation process.

[0053] Furthermore, during the execution of the striking action, when the subject adopts a side-standing position or a horizontal swing punch, the striking arm may obscure the opposite shoulder, hip, or torso, causing the confidence score of the two-dimensional pose estimation model for the output of the obscured joints to be significantly reduced or even completely lost in some viewpoints, and the three-dimensional skeletal joints after the weighted fusion of the three viewpoints show spatial position jitter or missing.

[0054] If a skeletal sequence containing missing values ​​is directly fed into a subsequent graph convolutional network, the missing node features will spread along the message passing path of the graph convolution, resulting in distortion of cross-limb collaborative feature extraction on the associated pathway. To address this, this embodiment adds a joint completion step after the multi-view fusion of the 3D skeletal joint spatiotemporal trajectory data is completed and before it is fed into the human skeletal spatiotemporal graph structure: missing joints are determined by confidence-driven judgment, and completion is completed by combining mirror filling of contralateral joints with interpolation of previous and subsequent frames. In the completed graph structure, a mask mark is added to the completed joints so that the attention unit can adjust their weights.

[0055] Specifically, the criteria for determining missing joints are: a joint's confidence score is below 0.25 in all three viewpoints, or the confidence score in at least two viewpoints is below 0.15, or a joint is not detected in a viewpoint because it falls outside the image boundary. A joint meeting any of these conditions is marked as missing. The completion operation has three priorities: the first priority—contralateral mirror completion—is suitable for unilateral missing joints with paired joints. The longitudinal axis of the human torso is taken as the mirror axis. The 3D coordinates of the non-missing contralateral joint are reflected along the mirror axis, and the reflected coordinates are directly filled into the missing joint location. The mirror axis is formed by the pelvic center O and the direction vector of the longitudinal axis of the torso. Determine the mirror image of key point P. ,in To and Orthogonal horizontal plane unit normal vector. Second priority—temporal interpolation completion, applicable to unilateral missing joints where contralateral joints are also missing or where paired joints do not exist. The missing joint is taken from the previous non-missing frame. With the next non-missing frame The three-dimensional coordinates are used for linear interpolation with the frame number as the independent variable. The interpolation formula is: When more than 8 consecutive frames are missing, linear interpolation becomes unreliable. Instead, the predicted values ​​from the previous 3 frames are used. The state vector of the Kalman filter is derived from the 3D coordinates and 3D velocity of the key points. The process noise covariance is also considered. Observation noise covariance The third priority—kinematic chain estimation—is applicable to cases where one side is missing, the contralateral joints are also missing, and the interpolation span is too long. The estimated position is obtained by adding the coordinates of the missing joint's parent joint to the average bone vector of the limb segment in adjacent non-missing frames. After completion, a differentiable confidence mask value is added to the completed joints. An additional mask channel is concatenated into the node features of the joint in that frame. This mask channel is then fed into the graph convolutional network so that subsequent attention units can adjust the contribution of the completed joint when calculating the spatial attention weights. In this embodiment, when the subject performs a right-side punch, the right arm sweeps across the front of the torso, occluding the left shoulder and left hip during frames 42 to 47. The left shoulder is completed using first-priority mirroring, and the left hip is completed using second-priority temporal interpolation. The completed sequence is then fed into the complete evaluation process. The deviation between the standardized score and the score of the unoccluded control group is 1.3 points, which is within the allowable range of scoring error.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing the normative nature of striking actions by an intelligent dummy based on deep learning, characterized in that, include: The video stream of the striking action is acquired and fed into the pose estimation network to obtain the spatiotemporal trajectory data of the three-dimensional skeletal joints. A spatiotemporal graph structure of the human skeleton is constructed. Based on the spatiotemporal trajectory data of the three-dimensional skeletal joints, a channel topology refined graph convolutional network is used to form a channel dynamic graph topology by using dynamic correlation matrix and global learning topology matrix to extract cross-limb non-local collaborative features between non-physically connected joints. The cross-limb non-local collaborative features are fed into the action-specific attention mechanism network, and weights are dynamically allocated on the time and space axes to capture the salient features of the force application phase. The salient features are fused with multi-stream feature data including joint coordinates, skeletal vectors, and motion velocity. The fused feature data is then fed into a motion quality assessment model trained based on a skeletal contrastive learning framework to obtain kinetic chain coordination scores and quantitative assessment results of striking motion standardization.

2. The method for recognizing the normative nature of intelligent dummy striking actions based on deep learning according to claim 1, characterized in that, The posture estimation network specifically involves: after acquiring the video stream data of the striking action, extracting the coordinates and confidence scores of the human body's two-dimensional joints frame by frame using a preset two-dimensional posture estimation model, and constructing a two-dimensional skeleton sequence; feeding the two-dimensional skeleton sequence into a three-dimensional posture enhancement network based on a spatiotemporal attention mechanism, and combining the camera intrinsic parameters to predict the joint depth coordinates in three-dimensional space, thereby obtaining the spatiotemporal trajectory data of the three-dimensional skeleton joints containing the three-dimensional spatial coordinates and time dimension coordinates of each joint.

3. The method for recognizing the normative nature of intelligent dummy striking actions based on deep learning according to claim 1, characterized in that, Constructing the spatiotemporal graph structure of the human skeleton includes: using three-dimensional skeletal joints as nodes of the graph structure, establishing a set of joints containing a spatial dimension and a set of frames containing a temporal dimension; for the spatial dimension, constructing spatial edges based on the natural connections in human anatomy, establishing the physical topological relationship between physically adjacent joints within the same frame; for the temporal dimension, establishing temporal edges between adjacent video frames for corresponding nodes of the same joint at different time points, establishing the spatiotemporal continuity of joint evolution over time; and constructing a three-dimensional skeletal spatiotemporal graph structure representing the entire process of the striking action based on the nodes, the spatial edges, and the temporal edges.

4. The method for recognizing the normative nature of intelligent dummy striking actions based on deep learning according to claim 3, characterized in that, Extracting cross-limb nonlocal collaborative features between non-physically connected joints using the channel topology refined graph convolutional network includes: inputting the 3D skeletal spatiotemporal graph structure and 3D skeletal joint spatiotemporal trajectory data into the channel topology refined graph convolutional network; adaptively capturing real-time motion associations between joints using a dynamic correlation matrix, and learning global motion prior topology using a global learning topology matrix; fusing the dynamic correlation matrix and the global learning topology matrix to generate independent channel dynamic graph topologies for different feature channels; and using the channel dynamic graph topology to establish nonlocal feature channel links for non-anatomically physically connected cross-limb joint pairs, extracting cross-limb nonlocal collaborative features of different limb joints during the striking force application process.

5. The method for recognizing the normative nature of intelligent dummy striking actions based on deep learning according to claim 1, characterized in that, The action-specific attention mechanism network includes a temporal attention unit, a spatial attention unit, and a spatiotemporal feature fusion unit. The temporal attention unit is used to receive the cross-limb non-local collaborative features, calculate the correlation weights of different video frames on the time axis of the striking action, generate a temporal attention weight map, and perform temporal weighting on the cross-limb non-local collaborative features to establish the temporal saliency of the key phase of striking force. The spatial attention unit is used to calculate the correlation weights of different joints on the human body spatial structure, generate a spatial attention weight map, and perform spatial weighting on the cross-limb non-local collaborative features to establish the spatial salience of the core joints of striking force. The spatiotemporal feature fusion unit is used to perform element-level dot product fusion of the temporal salient features weighted by the temporal attention unit and the spatial salient features weighted by the spatial attention unit to extract the salient features of the force exertion phase.

6. The method for recognizing the normative nature of intelligent dummy striking actions based on deep learning according to claim 1, characterized in that: The motion quality assessment model includes a positive and negative sample construction unit, a contrastive representation learning unit, a kinetic chain coordination evaluation unit, and a motion standardization quantification regression unit. The positive and negative sample construction unit uses standard striking motion data as anchor samples and non-standard striking motion data generated after perturbation of the standard striking motion data as negative samples to construct skeletal motion contrastive learning sample pairs. The contrastive representation learning unit receives the fused feature data and calculates the feature distances between anchor samples, positive samples, and negative samples based on the contrastive loss function within the contrastive learning framework. Through gradient backpropagation, it minimizes the feature distances of similar samples and maximizes the feature distances of dissimilar samples, mapping to obtain a higher-order motion representation vector. The kinetic chain coordination evaluation unit feeds the higher-order motion representation vector into a pre-defined kinetic chain evaluation branch, calculates the phase delay and energy transfer continuity of each joint in the human body during force transmission, and provides a kinetic chain coordination score. The motion standardization quantification regression unit concatenates the higher-order motion representation vector with the kinetic chain coordination score and maps it through a regression network to obtain the quantitative assessment result of the striking motion standardization.

7. The method for recognizing the normative nature of intelligent dummy striking actions based on deep learning according to claim 1, characterized in that: The salient features are fused with multi-stream feature data including joint coordinates, bone vectors, and motion velocity, including: In the spatiotemporal trajectory data of 3D skeletal joints, the 3D absolute coordinate flow features corresponding to each joint, the physical skeleton vector flow features composed of adjacent joints, and the motion velocity flow features between adjacent frames of the joints are obtained respectively. The three types of flow features are temporally aligned by a multi-scale one-dimensional convolutional network. The aligned multi-flow feature data and salient features are concatenated along the channel dimension, and cross-modal feature fusion is performed by pixel-by-pixel convolution to obtain the fused feature data.

Images