A standardized long sequence action recognition classification method

Abstract

The application relates to a standardized long-sequence action recognition classification method, which comprises the following steps: S1, acquiring a long-sequence human body skeleton point feature sequence, and adopting a sliding window mode to cut the long sequence into multiple local windows; S2, for each sliding window, simultaneously outputting a stage probability distribution and a boundary score through a window-level prediction network; S3, projecting and fusing the outputs of the windows onto a complete video time axis to obtain a frame-by-frame stage probability distribution sequence and a frame-by-frame boundary score sequence; S4, under the constraint of a monotonic stage transition, adopting a Viterbi dynamic programming algorithm to solve an optimal stage path; and S5, according to the stage switching positions in the optimal path, generating time intervals of each standardized action stage. The application solves the problems of how to realize high-precision and real-time stage classification of standardized long-sequence actions and further recover the action stage intervals on a complete time axis, and realizes efficient stage classification and complete interval recovery of standardized long-sequence actions.

Description

Technical Field

[0001] This invention relates to computer vision and intelligent sports technology, and in particular to a method for stage recognition and classification of standardized long sequences of human movements, which is especially applicable to long sequence standardized movement recognition scenarios with fixed movement order and clear stage division, such as martial arts routines, radio calisthenics, rhythmic gymnastics, and fitness exercises. Background Technology

[0002] In the fields of smart sports and motion quality assessment, accurately identifying and segmenting standardized long sequences of continuously executed movements is a crucial foundation for achieving automated scoring, motion correction, and training assistance. Taking martial arts routines as an example, a complete routine typically contains dozens of movement stages executed in a fixed order, with individual videos lasting tens of seconds or even longer. The movements have continuous transitions, and the stage boundaries often do not manifest as abrupt changes in a single frame.

[0003] The existing related technologies have the following main shortcomings:

[0004] Directly modeling complete long sequences end-to-end is computationally expensive, making it difficult to meet real-time processing requirements and hindering the learning of local action patterns. Only window-level or frame-level classification is performed, lacking explicit modeling of stage switching boundaries, which can easily cause classification jitter in the transition area between adjacent stages. Window-level prediction results are scattered across a large number of overlapping sliding windows, lacking an effective global timeline fusion and sequence constraint mechanism, making it difficult to recover the start and end intervals of each action stage on the complete video. Relying solely on classification confidence to determine stage switching is susceptible to local noise, leading to stage misordering, missed detections, or boundary shifts.

[0005] Therefore, there is an urgent need for a technical solution that can perform high-precision stage classification of standardized long sequence actions while meeting real-time requirements, and further recover the action stage intervals on the complete timeline. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing related technologies and solve the problem of how to achieve high-precision, real-time stage classification of standardized long sequence actions and further recover the action stage intervals on the complete timeline. It provides a standardized long sequence action recognition and classification method, which achieves efficient stage classification and complete interval recovery of standardized long sequence actions through a two-stage technical route of "window-level dual-task prediction + timeline fusion + monotonic constraint global decoding", providing a reliable foundation for subsequent action quality assessment, automatic scoring and training feedback.

[0007] The technical solution of this invention is as follows: A standardized long-sequence action recognition and classification method includes: Step S1: Obtain a long sequence of human skeletal point feature sequences, the feature sequences including normalized skeletal point coordinates and derived velocity and acceleration dynamic features; divide the feature sequences into multiple local windows using a sliding window method; Step S2: For each sliding window, perform dual-task prediction through a window-level prediction network: use the stage label of the window center frame as classification supervision, and construct Gaussian soft labels near the stage transition time to guide boundary regression, and simultaneously output the stage probability distribution and boundary score corresponding to the window; the window-level prediction network performs the following processing in sequence: perform motion encoding on the features within the window, fuse spatiotemporal position embedding, window relative position embedding and window size embedding to obtain a high-dimensional frame-level representation; perform multi-scale enhancement through a temporal-series feature pyramid structure; process through a multi-layer stacked spatiotemporal joint attention module and a multi-expert feedforward network; after attention pooling aggregation, output the results through the stage classification head and boundary regression head respectively; during training, a joint loss function of weighted classification loss and boundary regression loss is used, and a dynamic weight mechanism is introduced to suppress the dominance of boundary samples on the classification loss; Step S3: Backproject and fuse the stage probability distribution and boundary scores output by each window into the complete video timeline: For the stage probability distribution, accumulate and normalize the distribution by confidence level according to the center frame position of the window to obtain the frame-by-frame stage probability distribution sequence; For the boundary scores, use the same accumulation and normalization strategy, and perform one-dimensional linear interpolation and Gaussian smoothing on the uncovered frames, and then perform Softmax transformation to obtain the frame-by-frame boundary score sequence. Step S4: Under the monotonic stage transition constraint, the Viterbi dynamic programming algorithm is used to solve for the optimal stage path; the monotonic transition constraint allows the state to remain or advance to the next stage, and supports finite forward jump transitions and additional penalties; the observation score is defined as the probability value of the corresponding category in the current frame stage probability distribution, and the transition score is defined as being composed of the monotonic transition penalty and the boundary score cue: the transition score in the non-switching state is negatively correlated with the boundary score, and the transition score in the switching state is positively correlated with the boundary score; Step S5: Backtrack the optimal path of the Viterbi dynamic programming to obtain the optimal stage sequence frame by frame, detect the position of stage number change to determine the stage switching point, and generate the time interval set of each standardized action stage.

[0008] Furthermore, step S1 specifically includes: Let the input long sequence of human skeleton point features be X = {x_1, x_2, ..., x_t, ..., x_T}, where x_t represents the features of the t-th frame, including the normalized skeleton point coordinates and derived velocity and acceleration dynamic features; a sliding window approach is adopted, with a window length of W and a step size of S, and the i-th window is represented as X_i = {x_{b_i}, ..., x_{b_i+W}. 1}}, the center frame index is c_i=b_i+W / 2.

[0009] Furthermore, step S2 specifically includes: Each window uses the stage label of the center frame as classification supervision; at the same time, a boundary regression task is introduced: Gaussian soft labels are constructed near the stage transition time so that the boundary scores are continuously distributed near the transition boundary, and the boundary label of the i-th window is taken from the Gaussian soft label value at the center frame; the window-level prediction network outputs the stage probability distribution P_i and the boundary score B_i simultaneously; Motion coding and embedding fusion: Motion coding is performed on the frame-by-frame skeletal point features within the window, explicitly calculating velocity and acceleration, and introducing the average relative position of keypoints and bone length features. These features are then encoded separately by independent linear layers and concatenated to obtain a high-dimensional frame-level representation. Three types of embedding information are further fused: spatiotemporal position embedding: distinguishing different time steps and keypoint positions within the window; window relative position embedding: encoding the relative position of the window in the entire sequence, utilizing prior information about the fixed order of standardized actions; and window size embedding: supporting multi-window length training and inference, enabling the network to explicitly perceive the current time scale. Multi-scale enhancement: To address the significant differences in duration between different stages in standardized actions, a temporal-like feature pyramid structure is used to enhance window features at multiple scales. Multi-layer features are constructed by downsampling layer by layer along the time dimension, and then the high-level long-range context is backpropagated to the low layer through top-down fusion. Finally, the residual is fused with the original input, so that the model can retain both local action details and action organization information over a longer range. Spatiotemporal attention mechanism and dual-head output: The enhanced features are fed into a multi-layer stacked spatiotemporal joint attention module, each layer containing a spatiotemporal joint attention mechanism and a multi-expert feedforward network; the spatiotemporal joint attention simultaneously models the dynamic motion patterns of key points in the temporal and spatial dimensions; the multi-expert feedforward network selects and integrates different experts through a routing mechanism to adapt to the feature pattern differences of multi-stage actions. After being aggregated into window-level representations by attention pooling, the stage probability distribution and boundary score are output through the stage classification head and boundary regression head, respectively. Joint loss: Cross-entropy is used for classification loss, and binary cross-entropy is used for boundary regression loss; to address the problem of difficulty in classifying samples near the boundary, a dynamic weighting mechanism is introduced to suppress the dominance of boundary samples on classification loss; the total loss is the weighted sum of weighted classification loss and boundary regression loss.

[0010] Furthermore, step S3 specifically includes: Time-axis fusion: Since a long video sequence is divided into a large number of overlapping sliding windows, a single frame often corresponds to the prediction results of multiple windows. The window-level outputs need to be mapped and fused onto the original time axis. For stage classification output: Maintain an accumulation vector and accumulation weight for each frame on the time axis, write the output of each window according to the center frame position, accumulate it with weighted by the window confidence, and then normalize each frame to obtain a frame-by-frame stage probability distribution sequence. For the boundary score output: the same strategy is used to accumulate and normalize; frames not covered by the window center are completed by one-dimensional linear interpolation, then smoothed by Gaussian to suppress jitter, and finally converted into a frame-by-frame boundary score sequence by Softmax.

[0011] Furthermore, step S4 specifically includes: Global range recovery: Monotonic transition constraint: The stage number is used as the hidden state and the time index is used as the sequence step; since standardized actions have a fixed execution order, a monotonic constraint is imposed on stage transitions: only the current stage can be maintained or the next stage can be advanced; at the same time, limited forward jump transitions are supported and additional penalties are imposed to handle cases where individual stages are missed. Observation score and transition score: For each time step t and stage state k, the observation score is defined as the probability value of the corresponding class in the stage probability distribution of that frame; the transition score includes monotonic transition penalty and boundary score cue: when no stage switch occurs, the transition score is negatively correlated with the current frame boundary score (preferring to stay in the current stage); when a stage switch occurs, the transition score is positively correlated with the current frame boundary score (preferring to switch at the location with strong boundary signal).

[0012] Furthermore, step S5 specifically includes: Viterbi dynamic programming is used to find the optimal path with the largest cumulative score among all feasible paths that satisfy monotonic constraints. After backtracking to obtain the optimal stage sequence frame by frame, the set of transition times is extracted by detecting the change position of the stage number, and then the set of time intervals for each stage is generated.

[0013] Preferred parameter settings for method implementation: The Halpe-26 keypoint skeleton sequence was extracted from the video using the RTMPose pose estimation method and normalized to human scale. The video was processed by extracting one frame every two frames to construct the skeleton point feature sequence. Sliding window settings: window length W=15, step size S=1; boundary soft labels are generated by superimposing Gaussian kernels with standard deviation σ=3 and cutoff range ±9 at the stage switching position; Network parameters: latent space dimension 512, number of attention heads 6, feedforward layer dimension 2048, spatiotemporal attention stacking 8 layers; Halpe-26 skeleton topology is used to construct the adjacency matrix as spatial structure constraint; batch size 64, training for 50 epochs, AdamW optimizer, learning rate 1e-4; boundary loss weight λ=1.0, classification dynamic weight a=2.0, enabled after epoch 10; Global decoding parameters: The initial state constraint of the Viterbi algorithm is stage 1, the boundary score weight λ_b=30, and the single-step forward penalty coefficient γ=0.5.

[0014] The beneficial effects of this invention are as follows: By using a sliding window to segment long sequences, local action patterns can be effectively learned while ensuring real-time performance, thus reducing the computational overhead of modeling the complete sequence. The dual-task joint learning of classification and boundary regression avoids forcibly giving a clear classification in the stage switching region and provides continuous boundary observations for global interval recovery; The collaborative design of motion coding, multi-scale augmentation, spatiotemporal attention, and multi-expert feedforward significantly improves window-level stage classification accuracy and boundary regression fitting ability; The combination of timeline fusion and Viterbi monotonic constraint decoding effectively suppresses local classification jitter and restores the time intervals of each stage in the complete video; The boundary score-guided stage switching judgment makes the transition more likely to occur near the true boundary, improving the precision of interval recovery; The method is universal and can be extended to other standardized long sequence action recognition scenarios with fixed action order. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the standardized long sequence action recognition and classification algorithm of this invention; Figure 2 This is a schematic diagram of the window-level prediction network structure of the present invention (including motion coding, multi-scale enhancement, spatiotemporal attention, and dual-head output); Figure 3 This is a schematic diagram of the multi-scale enhancement module structure of the temporal feature pyramid of the present invention; Figure 4 This is a visual comparison chart of the window-level prediction results and interval recovery results for a single video in this invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0017] A standardized long-sequence action recognition and classification method includes three core components: window-level action classification and boundary regression, timeline fusion, and global interval recovery. Figure 1-4 As shown, the overall process is as follows: Step S1: Obtain the long sequence of human skeletal point feature sequences, and divide the long sequence into multiple local windows using a sliding window method; Step S2: For each sliding window, output the stage probability distribution and boundary scores simultaneously through the window-level prediction network; Step S3: Backproject the output of each window and fuse them onto the complete video timeline to obtain the frame-by-frame stage probability distribution sequence and the frame-by-frame boundary score sequence; Step S4: Under the monotonic stage transition constraint, the Viterbi dynamic programming algorithm is used to solve for the optimal stage path; Step S5: Generate the time interval for each standardized action stage based on the stage switching position in the optimal path.

[0018] Example 1: Stage Recognition of Martial Arts Routine Movements Taking the juvenile continuous punch routine as an example, a complete routine consists of 34 action phases performed in a fixed order.

[0019] Data preprocessing: Halpe-26 keypoint skeleton sequences were extracted from the video using pose estimation methods such as RTMPose, and then normalized to human scale. One frame was extracted from every two frames of the video for processing to construct a skeleton feature sequence.

[0020] Sliding window settings: Window length W=15 (corresponding to 30 frames in the original video, approximately 1 second), step size S=1. Boundary soft labels are generated by superimposing Gaussian kernels (standard deviation σ=3, cutoff range ±9) at the stage switching positions.

[0021] Network parameters: latent space dimension 512, attention heads 6, feedforward layer dimension 2048, spatiotemporal attention stacking 8 layers. The Halpe-26 skeleton topology is used to construct the adjacency matrix as a spatial structure constraint. Batch size 64, training 50 epochs, AdamW optimizer, learning rate 1e-4. Boundary loss weight λ=1.0, classification dynamic weight a=2.0 (enabled after epoch 10).

[0022] Global decoding parameters: The initial state constraint of the Viterbi algorithm is stage 1, the boundary score weight λ_b=30, and the single-step forward penalty coefficient γ=0.5.

[0023] Experimental procedure: Collect long sequence of action video data as training dataset, and assemble each step into a complete network model according to the deep learning network model, which can be trained on a single GPU.

[0024] To verify the effectiveness and advancement of the proposed algorithm, systematic experiments were conducted on a standardized long-order action dataset. The experiments focused on three dimensions: window-level classification performance, boundary regression performance, and video-level interval recovery performance. Multiple ablation experiments were used to verify the effectiveness of each core module and key hyperparameters.

[0025] Experimental results: The window-level classification accuracy reached 92.90%, and the boundary regression R² reached 0.6528. After global decoding, the average error at the start of the interval was 2.80 frames, the average error at the end of the interval was 2.81 frames, the average IoU of the interval was 0.8359, and the interval success detection rate was 99.71%, which can completely cover all 34 action stages in the correct order.

[0026] Ablation experiment verification (1) Module ablation: The introduction of three modules—encoding embedding, multi-expert feedforward network, and attention pooling—all brought about a stable performance improvement. The classification accuracy was 88.95% without using any modules, and increased to 92.90% after adding all modules.

[0027] (2) Number of network layers: The 8-layer spatiotemporal attention configuration achieves the best balance between model capacity and training stability.

[0028] (3) Window width: When the window width is 15, the best overall performance is achieved in both classification and boundary regression, taking into account both local discriminativeness and contextual integrity.

[0029] (4) Boundary score weights: Boundary score weights λ in Viterbi decoding b achieves optimal interval recovery around 30; the start / end point error is about 2.95 frames when there is no boundary guidance, which is reduced to about 2.80 frames after the introduction.

[0030] Example 2: Other Standardized Long Sequence Actions This invention is also applicable to standardized long sequence movements with fixed movement order and clear stage divisions, such as radio calisthenics, rhythmic gymnastics, fitness exercises, and Tai Chi routines. Only the number of stages, window parameters, and skeleton key point definitions need to be adjusted according to the specific movement type, while the core algorithm flow (sliding window dual-task prediction, time axis fusion, and Viterbi monotonic constraint decoding) remains unchanged.

[0031] Innovation points: A sliding window dual-task prediction method for standardized long sequence actions is proposed, which simultaneously performs stage classification and boundary regression, balancing real-time performance and recognition accuracy. Design a window-level prediction network that integrates motion coding, multi-scale temporal feature pyramids, spatiotemporal joint attention, and multi-expert feedforward; A two-stage interval recovery method combining window-level output back-projection fusion and Viterbi monotonic constraint global decoding is proposed. By using boundary fraction clues to guide the timing of phase switching, the accuracy and stability of phase interval recovery for long sequence actions can be improved.

[0032] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A standardized long sequence action recognition and classification method, characterized in that, include: Step S1: Obtain a long sequence of human skeletal point feature sequences, the feature sequences including normalized skeletal point coordinates and derived velocity and acceleration dynamic features; The feature sequence is divided into multiple local windows using a sliding window approach; Step S2: For each sliding window, perform dual-task prediction through a window-level prediction network: use the stage label of the window center frame as classification supervision, and construct Gaussian soft labels near the stage transition time to guide boundary regression, and simultaneously output the stage probability distribution and boundary score corresponding to the window; the window-level prediction network performs the following processing in sequence: perform motion encoding on the features within the window, fuse spatiotemporal position embedding, window relative position embedding and window size embedding to obtain a high-dimensional frame-level representation; perform multi-scale enhancement through a temporal-series feature pyramid structure; Processed via multi-layered stacked spatiotemporal joint attention modules and multi-expert feedforward networks; After attention pooling aggregation, the results are output through the stage classification head and the boundary regression head respectively; during training, a joint loss function of weighted classification loss and boundary regression loss is used, and a dynamic weight mechanism is introduced to suppress the dominance of boundary samples on the classification loss; Step S3: Backproject and fuse the stage probability distribution and boundary scores output by each window into the complete video timeline: For the stage probability distribution, accumulate and normalize the distribution by confidence level according to the center frame position of the window to obtain the frame-by-frame stage probability distribution sequence; For the boundary scores, use the same accumulation and normalization strategy, and perform one-dimensional linear interpolation and Gaussian smoothing on the uncovered frames, and then perform Softmax transformation to obtain the frame-by-frame boundary score sequence. Step S4: Under the monotonic stage transition constraint, the Viterbi dynamic programming algorithm is used to solve for the optimal stage path; the monotonic transition constraint allows the state to remain or advance to the next stage, and supports finite forward jump transitions and additional penalties; The observation score is defined as the probability value of the corresponding category in the probability distribution of the current frame stage. The transition score is defined as being composed of monotonic transition penalty and boundary score cue: the transition score is negatively correlated with the boundary score in the non-switching state, and positively correlated with the boundary score in the switching state. Step S5: Backtrack the optimal path of the Viterbi dynamic programming to obtain the optimal stage sequence frame by frame, detect the position of stage number change to determine the stage switching point, and generate the time interval set of each standardized action stage.

2. The standardized long-sequence action recognition and classification method according to claim 1, characterized in that, Step S1 specifically includes: Let the input long sequence of human skeleton point features be X = {x_1, x_2, ..., x_t, ..., x_T}, where x_t represents the features of the t-th frame, including the normalized skeleton point coordinates and derived velocity and acceleration dynamic features; a sliding window approach is adopted, with a window length of W and a step size of S, and the i-th window is represented as X_i = {x_{b_i}, ..., x_{b_i+W}. 1}}, the center frame index is c_i=b_i+W / 2.

3. The standardized long-sequence action recognition and classification method according to claim 1, characterized in that, Step S2 specifically includes: Each window uses the stage label of the center frame as classification supervision; at the same time, a boundary regression task is introduced: Gaussian soft labels are constructed near the stage transition time so that the boundary scores are continuously distributed near the transition boundary, and the boundary label of the i-th window is taken from the Gaussian soft label value at the center frame; the window-level prediction network outputs the stage probability distribution P_i and the boundary score B_i simultaneously; Motion coding and embedding fusion: Motion coding is performed on the frame-by-frame skeletal point features within the window, explicitly calculating velocity and acceleration, and introducing the average relative position of keypoints and bone length features. These features are then encoded separately by independent linear layers and concatenated to obtain a high-dimensional frame-level representation. Three types of embedding information are further fused: spatiotemporal position embedding: distinguishing different time steps and keypoint positions within the window; window relative position embedding: encoding the relative position of the window in the entire sequence, utilizing prior information about the fixed order of standardized actions; and window size embedding: supporting multi-window length training and inference, enabling the network to explicitly perceive the current time scale. Multi-scale enhancement: A temporal-like feature pyramid structure is used to enhance the window features at multiple scales; multi-layer features are constructed by downsampling layer by layer along the time dimension, and then the high-level long-range context is back-propagated to the low layer through top-down fusion, and finally fused with the original input in a residual manner; Spatiotemporal attention mechanism and dual-head output: The enhanced features are fed into a multi-layer stacked spatiotemporal joint attention module, each layer containing a spatiotemporal joint attention mechanism and a multi-expert feedforward network; the spatiotemporal joint attention simultaneously models the dynamic motion patterns of key points in the temporal and spatial dimensions; the multi-expert feedforward network selects and integrates different experts through a routing mechanism to adapt to the feature pattern differences of multi-stage actions. After being aggregated into window-level representations by attention pooling, the stage probability distribution and boundary score are output through the stage classification head and boundary regression head, respectively. Joint loss: Cross-entropy is used for classification loss, and binary cross-entropy is used for boundary regression loss; to address the problem of difficulty in classifying samples near the boundary, a dynamic weighting mechanism is introduced to suppress the dominance of boundary samples on classification loss; the total loss is the weighted sum of weighted classification loss and boundary regression loss.

4. The standardized long-sequence action recognition and classification method according to claim 1, characterized in that, Step S3 specifically includes: Time axis fusion: Map and fuse window-level outputs onto the original time axis; For stage classification outputs: Maintain an accumulation vector and accumulation weight for each frame on the time axis, write the output of each window according to the center frame position, accumulate it with weighted by window confidence, and then normalize each frame to obtain a frame-by-frame stage probability distribution sequence. For the boundary score output: the same strategy is used to accumulate and normalize; frames not covered by the window center are completed by one-dimensional linear interpolation, then smoothed by Gaussian to suppress jitter, and finally converted into a frame-by-frame boundary score sequence by Softmax.

5. The standardized long-sequence action recognition and classification method according to claim 1, characterized in that, Step S4 specifically includes: Global range recovery: Monotonic transition constraint: The stage number is used as the hidden state and the time index is used as the sequence step; a monotonic constraint is imposed on stage transitions: only the current stage can be maintained or the next stage can be advanced; at the same time, finite forward jump transitions are supported and additional penalties are imposed to handle cases where individual stages are missed. Observation score and transition score: For each time step t and stage state k, the observation score is defined as the probability value of the corresponding class in the stage probability distribution of that frame; the transition score includes monotonic transition penalty and boundary score cue: when no stage switch occurs, the transition score is negatively correlated with the current frame boundary score; when a stage switch occurs, the transition score is positively correlated with the current frame boundary score.

6. The standardized long-sequence action recognition and classification method according to claim 1, characterized in that, Step S5 specifically includes: Viterbi dynamic programming is used to find the optimal path with the largest cumulative score among all feasible paths that satisfy monotonic constraints. After backtracking to obtain the optimal stage sequence frame by frame, the set of transition times is extracted by detecting the change position of the stage number, and then the set of time intervals for each stage is generated.

7. The standardized long sequence action recognition and classification method according to claim 1, characterized in that, Method implementation parameter settings: The Halpe-26 keypoint skeleton sequence was extracted from the video using the RTMPose pose estimation method and normalized to human scale. The video was processed by extracting one frame every two frames to construct the skeleton point feature sequence. Sliding window settings: window length W=15, step size S=1; boundary soft labels are generated by superimposing Gaussian kernels with standard deviation σ=3 and cutoff range ±9 at the stage switching position; Network parameters: latent space dimension 512, number of attention heads 6, feedforward layer dimension 2048, spatiotemporal attention stacking 8 layers; Halpe-26 skeleton topology is used to construct the adjacency matrix as spatial structure constraint; batch size 64, training for 50 epochs, AdamW optimizer, learning rate 1e-4; boundary loss weight λ=1.0, classification dynamic weight a=2.0, enabled after epoch 10; Global decoding parameters: The initial state constraint of the Viterbi algorithm is stage 1, the boundary score weight λ_b=30, and the single-step forward penalty coefficient γ=0.5.

Images