A skeleton action recognition method of fusing topology prior of dynamic graph convolution

Abstract

The application discloses a skeleton action recognition method based on dynamic graph convolution fusing topological structure priori, and comprises the following steps: generating a topological barcode and converting the topological barcode into a continuous global topological feature vector through a differentiable mapping mode; constructing a node flow dynamic graph and an edge flow dynamic graph; extracting features of four modes of joints, bones, joint motion and bone motion, calculating topological confidence by using the global topological feature vector, and independently generating affine transformation parameters for each mode according to the confidence to obtain multi-modal joint features fusing topological priori; in the shallow stage of the network, the global topological feature vector is injected with high intensity to constrain the bone geometry structure, and in the deep stage of the network, the injection intensity of the topological information is gradually reduced through an inverse S-shaped continuous scheduling function, so that the network focuses on learning high-dimensional semantic discriminative features; and outputting an action recognition result; and the application can accurately distinguish complex or easily confused human behaviors.

Description

Technical Field

[0001] This invention relates to the field of computer vision and pattern recognition technology, and in particular to a skeletal action recognition method that integrates dynamic graph convolution with topological priors. Background Technology

[0002] In recent years, Graph Convolutional Networks (GCNs) have become the mainstream method in the field of skeletal action recognition due to their ability to naturally represent and analyze the spatiotemporal structure and dynamic changes of skeletons. Existing GCN methods typically use joints as graph nodes and bones as graph edges for spatial modeling, and connect joints with the same attribute in different frames for temporal modeling, achieving significant recognition results. However, when facing complex and ever-changing real-world scenarios, existing GCN-based skeletal action recognition algorithms still have the following significant drawbacks:

[0003] (1) Insufficient robustness across viewpoints and high sensitivity to Euclidean coordinates: Existing GCN methods typically rely on predefined human skeleton connections and are highly sensitive to the Euclidean representation of skeleton coordinates. In practical applications, due to frequent changes in camera viewpoints, the distribution of skeleton sequence coordinates for the same action can shift significantly under different viewpoints. This graph convolution model, which is based on fixed topology or Euclidean geometry, struggles to characterize cross-viewpoint consistency, leading to a significant decrease in the model's recognition performance under cross-viewpoint conditions.

[0004] (2) Unstable joint modeling of multimodal features: In order to provide richer discriminative information, existing technologies often introduce multimodal skeletal features (such as joint, bone, joint_motion, and bone_motion) for joint training. However, under complex action or noise interference conditions, traditional feature fusion methods are prone to mutual inhibition and modal conflict between different modalities, which cannot achieve stable, non-competitive adaptive modulation and seriously restricts the model's generalization ability under cross-viewpoint and cross-modal conditions.

[0005] (3) Deep graph convolutional networks face a serious problem of “oversmoothing” of features: In deep GCNs, after multiple aggregations of neighborhood features, the features of graph nodes (skeletal joints) become very similar. This “oversmoothing” phenomenon causes the model to lose its ability to capture action details (e.g., it cannot distinguish between features of hands and feet), which limits high-level semantic expression and greatly reduces the model’s ability to distinguish and discriminate fine-grained, easily confused actions (such as “washing hands” and “rubbing hands”).

[0006] Currently, Topological Data Analysis (TDA), as an analytical tool that starts from the overall structure of data, can extract topological invariants (such as persistent cohomology) that are naturally stable to rotation, translation, and local perturbations. Although some existing studies have attempted to perform "feature-level concatenation" or "posterior fusion" of topological features with traditional features, topological information has not been deeply involved in the core processes of skeletal structure modeling and deep feature propagation. How to enable topological information to participate in feature propagation in a stage-aware and schedulable manner in deep graph convolutional networks, while balancing structural consistency and action discrimination capabilities, remains a pressing technical challenge. Summary of the Invention

[0007] Purpose of the Invention: The purpose of this invention is to provide a skeletal action recognition method based on dynamic graph convolution that integrates topological priors. It introduces persistent cohomology methods from topological data analysis to extract topological invariants that are naturally robust to viewpoint changes such as rotation and translation. These invariants serve as global structural priors to guide dynamic graph convolution, enhancing the model's adaptability to cross-viewpoint changes. Through an uncertainty-aware topological FiLM (TA-FiLM) modulation mechanism, different modal features are non-competitively and adaptively modulated, improving the stability and consistency of multimodal feature fusion. Furthermore, through a stage-aware topology injection scheduling strategy, the strength of the topological prior is controlled at different network depth stages via a continuous scheduling mechanism, alleviating the problem of excessive feature smoothing in deep graph convolutional networks. This approach balances structural consistency and action discrimination ability, addressing issues such as cross-viewpoint geometric perturbations, multimodal feature fusion conflicts, and excessive feature smoothing in deep networks.

[0008] Technical solution: The present invention provides a skeletal action recognition method based on dynamic graph convolution that integrates topological priors, comprising the following steps:

[0009] Step 1: Extract high-dimensional point cloud representations of the joints of each frame from the input bone sequence, analyze the distance changes between joints in the point cloud using the persistent cohomology method, generate a topological barcode to describe the evolution of the overall connected structure and ring structure of the skeleton, and convert the topological barcode into a continuous global topological feature vector through micro-mapping.

[0010] Step 2: Decompose the skeletal input features into node flow sub-features and edge flow sub-features in the channel dimension, construct node flow dynamic graphs and edge flow dynamic graphs respectively, and introduce global topological feature vectors as bias terms in the self-attention mechanism of node flow and edge flow to guide the generation of dynamic adjacency relationships with topological structure information, thereby forming a topology-guided three-branch dynamic graph convolutional network.

[0011] Step 3: Extract features from four modalities: joints, bones, joint motion, and bone motion. Calculate topological confidence using global topological feature vectors and generate affine transformation parameters independently for each modality based on the confidence. Perform non-competitive linear modulation updates on the features of each modality to obtain multimodal joint features that incorporate topological priors.

[0012] Step 4: In the shallow stage of the network, global topological feature vectors are injected with high intensity to constrain the skeletal geometry. In the deep stage of the network, the injection intensity of topological information is gradually reduced through an inverse S-shaped continuous scheduling function, so that the network focuses on learning high-dimensional semantic discriminative features, thereby alleviating the problem of excessive feature smoothing in deep graph convolutional networks.

[0013] Step 5: Input the features after multimodal modulation and stage scheduling into the classifier, and output the action recognition result.

[0014] Furthermore, in step 1, after modeling the skeletal joints as a high-dimensional point cloud, a simple complex sequence with varying distance thresholds is constructed based on the Euclidean distance between the joints. The topological feature sets of each dimension are extracted using a persistent homology method, and the features of all topological barcodes are fused into a continuous global topological feature vector through nonlinear mapping and splicing operations.

[0015] Furthermore, in step 2, the construction of the node flow dynamic graph is as follows: calculate the correlation score between the features of every two key points, add the global topological feature vector to the correlation score to form a correlation score with topological guidance, and then convert it into attention weights between nodes through a normalization function. The weights are used to perform weighted aggregation of the features of neighboring nodes to update the feature representation of the current node.

[0016] Furthermore, in step 2, the construction of the edge flow dynamic graph is as follows: each skeletal edge is represented as the feature difference between two endpoints, then the correlation score between the features of each pair of skeletal edges is calculated, and the global topological feature vector is added to the correlation score. After normalization, the attention weight between edges is obtained, which is used to perform weighted aggregation of the features of neighboring edges to update the feature representation of the current edge.

[0017] Furthermore, in step 3, the topological confidence score is calculated by performing a nonlinear transformation on the global topological feature vector and then normalizing it, and is used to characterize the reliability of the topological structure in different modes.

[0018] Further, step 3 is as follows: For each of the four modalities—joint, bone, joint motion, and bone motion—modal embedding features are extracted. Then, the topological confidence is multiplied element-wise with the modal embedding features. The multiplication results are then input into two independent linear transformation layers to generate scaling and translation factors for the modalities. Finally, the generated scaling and translation factors are used to perform channel-wise linear modulation updates on the original modal features to obtain the modal features after fusing the topological prior.

[0019] Furthermore, in step 4, the inverse S-shaped continuous scheduling function calculates the topology injection strength coefficient based on the current layer of the network. As the number of layers increases, the coefficient value of the inverse S-shaped continuous scheduling function monotonically decreases from close to 1 to close to 0, and the decreasing process exhibits an inverse S-shaped curve characteristic of being slow at first, fast in the middle, and slow again. The inverse S-shaped continuous scheduling function is jointly controlled by two preset parameters: the attenuation slope and the critical layer number. That is, the attenuation slope determines the rate at which the coefficient decreases, and the critical layer number determines the network layer number corresponding to when the coefficient value decreases to the middle value.

[0020] Furthermore, in step 5, the classifier uses the cross-entropy loss function to calculate the error between the predicted class and the true label, and adds a norm regularization term for the modulation parameters to prevent overfitting, and optimizes the network parameters through backpropagation.

[0021] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the methods described herein.

[0022] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0023] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: It embeds a deep graph convolutional network that integrates topological priors and multimodal modulation as the core algorithm in human action sequence analysis. This framework utilizes globally topological features extracted with persistent cohomology as a reference to achieve geometric consistency alignment of the skeletal core structure under cross-view conditions. Furthermore, within the network, it employs an uncertain-aware TA-FiLM module and a stage-aware scheduling strategy, enabling the network to non-competitively and adaptively fuse four modal features: joints, bones, joint motion, and skeletal motion. It also effectively limits excessive feature smoothing at a deep level, achieving a smooth transition from structural constraints to high-dimensional semantic discrimination in complex scenes. This framework solves the problems of existing technologies, such as large camera viewpoint disturbances during feature extraction, easy conflicts in multimodal fusion, and lack of robustness of deep node features, significantly improving the accuracy and cross-view generalization ability of skeletal action recognition. The intelligent behavior analysis method designed based on this algorithm can quickly and stably capture the target person's action intentions and spatiotemporal trajectories, thereby accurately identifying complex or easily confused human behaviors. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention;

[0025] Figure 2 This is a diagram of the three-branch network architecture of the topological dynamic graph convolutional network of the present invention;

[0026] Figure 3 This is a structural decomposition diagram of the uncertainty-aware TA-FiLM modulation mechanism of the present invention;

[0027] Figure 4 This is the evolution curve and schematic diagram of the phase-aware topology injection scheduling strategy of the present invention.

[0028] Figure 5 This is a flowchart of the model training process of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, embodiments of the present invention provide a skeletal action recognition method that integrates topological prior dynamic graph convolution. Specifically:

[0031] Step 1: Preprocess the skeletal sequences in the dataset, extract the global topological auxiliary signal of the skeletal point cloud, and map it into continuous features. Specifically, this involves each skeletal keypoint. initial representation It combines its three-dimensional spatial coordinates with semantic identity. First, the skeletal joint set is modeled as a high-dimensional point cloud. Then, a multilayer perceptron (MLP) is used to embed the keypoint locations into a high-dimensional vector, as shown below:

[0032]

[0033] in Semantic identity of skeletal key points; It is a multilayer perceptron used to embed key point locations into a high-dimensional vector; This is in high-dimensional point clouds The skeletal joints. To extract cross-view-invariant structural information, this invention is based on a distance threshold. Construct simple complex sequences and extract topological barcodes using persistent homology. Let the th... The set of persistent cohomological features of dimension is ,in and These represent the "birth" and "death" times of topological features, respectively.

[0034] To incorporate discrete topological features into end-to-end network training, a differentiable mapping function is introduced to transform them into a global topological feature vector. The mapping process for aggregating all barcode information is updated as follows:

[0035]

[0036] in This represents the total number of features in the topological feature set; It is a non-linear activation function; This represents the weight matrix used to map topological features; Indicates feature splicing and connection; This is the bias term. Vector. This will serve as global prior structural information to guide the generation of subsequent graph structures.

[0037] Step 2: Decouple the skeletal input features and construct a topology-guided three-branch dynamic graph convolutional network to capture higher-order structures of nodes, edges, and regular flows. First, a network consisting of... vertices and a set of edges Composition diagram Input skeleton features It is split into two independent sub-features along the channel dimension. and Each node performs adaptive dynamic graph construction. Taking Node Stream as an example, this is done by computing nodes... With nodes Correlation score between To construct dynamic adjacency relationships:

[0038]

[0039] in, and These are the weight transformation matrices for the query and the key, respectively. and For nodes and nodes Feature representation; The dimension of the key vector is used to scale the dot product result.

[0040] To ensure that the generated dynamic map maintains the connectivity of the skeletal core structure even when the viewpoint changes drastically, the global topological features obtained in Step 1 are used. This is injected into the self-attention mechanism. Simultaneously, the importance weights are normalized using the softmax function, and a topology-guided bias is added.

[0041]

[0042] in These are importance weights with topologically guided bias and normalized using the softmax function. and The correlation score between nodes; This is the weight matrix corresponding to the global topological feature vector; The global topological feature vector is used here as a topological guided bias injection self-attention mechanism. For nodes The set of neighboring nodes. Then, passed to the node. Local messages can be aggregated by combining the features of neighbors with relevant attention weights, and finally the node at the 1st rank can achieve the desired result. The representation is updated as follows:

[0043]

[0044] in and The nodes are respectively No. Layer representation and nodes in No. The updated representation of the layer; It is a multilayer perceptron; For nodes The set of neighboring nodes; Neighbor nodes For nodes Relevant attention weights; The weight matrix for the value vector; Neighboring nodes The characteristics are represented.

[0045] Similarly, we can obtain the feature outputs of the Edge Stream and the General Stream, and finally fuse the feature matrices of the three branches to obtain a hybrid feature representation.

[0046] Step 3: A TA-FiLM modulation mechanism based on uncertainty awareness is used to fuse features from four skeletal modalities, and a stage-aware topology injection scheduling strategy is combined to update network features. Details are as follows:

[0047] Based on the global topological information obtained through persistent cohomology in Step 1, spatiotemporal graph features are then extracted from the three-branch structure in Step 2. Furthermore, to comprehensively characterize human movement, this invention extracts and inputs four modal features: joints (represented as...). ), skeleton (represented as) ), joint motion (represented as ) and bone motion (referred to as ).

[0048] Therefore, this invention first constructs an uncertainty-aware TA-FiLM modulation mechanism. To avoid mutual inhibition and conflict among the four modes during joint modeling, topological confidence is introduced. The global uncertainty of topological features is calculated using the following formula:

[0049]

[0050] in It is a non-linear activation function; This is the weight matrix; This represents the global topological feature vector. This is the bias term. Based on this confidence level... For each specific mode Generate independent affine transformation parameters (scaling factors) Translation factor :

[0051]

[0052] in and Indicates mode-specific The weight matrix; For topological confidence; This is an element-wise multiplication operation; Represents the global topological feature vector; and It is modal specific The bias term is then used. Subsequently, each modal feature is updated using non-competitive linear modulation to obtain modal features that fuse topological priors. :

[0053]

[0054] in and These represent the scaling factor and the translation factor, respectively. This represents the input features of a specific modality. At this point, highly stable and complementary multimodal joint features are obtained.

[0055] After resolving the issue of multimodal conflicts, the modal features are combined with topological features, and then the stage-aware topology injection scheduling strategy is improved.

[0056] Deep networks are prone to oversmoothing of features. To control topology information at different network layers... The injection strength coefficient was designed based on the effect intensity. Continuous evolution scheduling is performed using an inverse S-shaped function:

[0057]

[0058] in The decay slope determines how quickly the coefficient decreases; and These represent the current layer number and the critical layer number of the network, respectively. The critical layer number is the network layer number corresponding to when the coefficient of determination drops to the median value. The final output of each graph convolutional block is modulated as follows:

[0059]

[0060] in For dynamic graph convolution block operations; For the first Input features of each graph convolutional block; The injection intensity coefficient for phase decay gating control; For stage-aware parameters; This represents the global topological feature vector. In the shallow layer... Approaching 1, strongly constrained geometry; at a deep level When the value approaches 0, the network degenerates into weak regularization, focusing on distinguishing fine-grained, high-dimensional semantic action features such as "wash hands" and "rub hands".

[0061] Step 4: Calculate the prediction error based on the joint cross-entropy and regularized loss function, optimize the network parameters through backpropagation to obtain the trained action recognition model, and use the model to perform skeletal action recognition.

[0062] This invention evaluates the proposed method on public skeletal motion recognition datasets (such as NTU RGB+D 60 and NTU RGB+D 120), using the cross-entropy loss function to measure the class classification predicted by the model. With real motion tags The differences between them, and the addition of modulation parameters Regularization is used to prevent overfitting. The final total loss function... The calculation is as follows:

[0063]

[0064] in The total number of action categories. Tag it as real action. For the class distribution predicted by the model, For regularization weight parameters, This represents the square of the L2 norm regularization term applied to the modulation parameters (scaling and translation factors) to prevent overfitting. The AdamW optimizer is used for parameter updates, with an initial learning rate of 0.001 and a cosine annealing strategy for decay. Through the complementarity of multimodal features and the constraints of topological priors, the model can converge faster and significantly improve recognition accuracy in cross-viewpoint and complex action contexts.

[0065] In the practical application reasoning stage, the real-time skeletal video sequence to be identified is input into the trained action recognition model, which outputs the target person's action intention and behavior category.

[0066] Example 1:

[0067] This invention provides a skeletal action recognition method that integrates topological prior dynamic graph convolution, comprising the following steps:

[0068] S11: By forming a monitoring and perception network with multiple networked depth image sensors (such as Kinect) or RGB cameras, the captured human behavior videos are stored in real time to cloud servers or edge computing nodes.

[0069] S12: Preprocess the video data in the cloud or on a local workstation. Use a human pose estimation algorithm (such as OpenPose) to extract human skeletal features from the video sequence. Normalize, align, and denoise the 3D coordinates of the joints of each frame of the skeleton, uniformly crop and fill them into a fixed-length skeletal video sequence. These structured skeletal data form a gallery for network training and testing.

[0070] S13: Input the skeleton sequence from the image library into the topology-guided dynamic graph convolutional neural network framework designed in this invention, extract the corresponding spatiotemporal features and action category vectors, and store the processing results and model weights on a cloud server.

[0071] S14: Extract global topological auxiliary signals. The skeletal joints of a single frame are modeled as point clouds in a high-dimensional space. A distance matrix is ​​constructed based on the Euclidean distance between joints, and this matrix is ​​used to generate simplex complex sequences. By calculating persistent homology, a topological barcode that characterizes the evolution of the overall connected components and loop structure of the skeleton is generated. To integrate the discrete barcode into the network, a differentiable mapping function is used to transform it into a continuous global topological feature vector. Let the set consisting of the "birth" and "death" times of the topological feature be denoted as . The mapping process is as follows:

[0072]

[0073] in This represents the total number of features in the topological feature set; It is a non-linear activation function; This represents the weight matrix used to map topological features; Indicates feature splicing and connection; This is a bias term.

[0074] S15: Construct a topology-guided three-branch dynamic graph. For example... Figure 2 As shown, input features Decomposed into nodal flow sub-features and edge flow characteristics In the node flow branch, by computing nodes... and The adjacency matrix is ​​dynamically constructed based on the correlation between the parameters. To calibrate the Euclidean coordinate shift caused by changes in viewpoint, the values ​​obtained in S14 are used... Injected into self-attention computation:

[0075]

[0076] in These are importance weights with topologically guided bias and normalized using the softmax function. and These are the weight transformation matrices for the query and the key, respectively. and For nodes and nodes Feature representation; The dimension of the key vector is used to scale the dot product result; This is the weight matrix corresponding to the global topological feature vector; The global topological feature vector is used here as a topological guided bias injection self-attention mechanism. For nodes The set of neighboring nodes. By introducing a topological bias, the network can maintain correct connectivity of the human body's physical structure even when the camera's viewpoint changes drastically. Then, the three-branch enhanced features are concatenated with channels... Convolution is used to fuse features from multiple viewpoints.

[0077] S16: Multimodal modulation is performed using the uncertain-aware TA-FiLM mechanism, such as... Figure 3 As shown, since a single skeletal feature cannot fully represent movement, four modalities were extracted: Joint, Bone, Joint Motion, and Bone Motion. To avoid conflicts during modal fusion, topological confidence was calculated using global topological features. Next, specific modal embeddings are combined. Generates modes independent of each mode Affine transformation parameters (scaling factors) Translation factor ):

[0078]

[0079] in and Indicates mode-specific The weight matrix; For topological confidence; This is an element-wise multiplication operation; Represents the global topological feature vector; and It is modal specific The bias term. Then, the input mode... Perform non-competitive modulation updates:

[0080]

[0081] in and These represent the scaling factor and the translation factor, respectively. This represents the input features of a specific modality.

[0082] S17: Execution-phase aware topology injection scheduling strategy to mitigate feature "over-smoothing", such as Figure 4 As shown. In the first In each dynamic graph convolutional block (DGBlock), the input features The update formula is subject to stage decay gating. control:

[0083]

[0084] in For dynamic graph convolution block operations; For the first Input features of each graph convolutional block; The injection intensity coefficient for phase decay gating control; For stage-aware parameters; This represents the global topological feature vector. To achieve a smooth transition from early strong dependency structures to later emphasis on semantic discrimination, an inverse S-shaped scheduling function is used for modeling. :

[0085]

[0086] in The decay slope determines how quickly the coefficient decreases; and These represent the current layer number and the critical layer number, respectively. The critical layer number is the network layer number corresponding to the point where the coefficient of determination drops to the median value. Under this scheduling, shallow networks are forced to learn the physical connectivity relationships of the skeleton, while deep networks retain the unique trajectory features of each joint, effectively distinguishing actions with subtle local differences, such as "washing hands" and "rubbing hands".

[0087] S18: After multi-layer graph convolution and global average pooling, a Softmax classifier is used to output the action class probability. Joint cross-entropy and... The regularized loss function measures the difference between the model output and the real action label, and optimizes the model parameters through backpropagation.

[0088] Figure 5The flowchart of the training process for the neural network model provided by this invention is shown, including the following steps: First, the skeletal motion data is preprocessed by expanding the samples through data augmentation methods such as spatial rotation, temporal scaling, and random joint occlusion, and then converted into tensor format for standardization and normalization. After training begins, the data undergoes forward propagation in a three-branch graph convolutional network, followed by TA-FiLM multimodal modulation and stage-aware scheduling to obtain features and predicted labels. Subsequently, the loss function is calculated, and backpropagation is performed to update the weights. Finally, the accuracy is calculated on the validation set and test set (e.g., under the NTU RGB+D 60 / 120 Xsub and Xview settings). If the accuracy is high, the optimal model is saved. After reaching the preset number of training epochs (e.g., 300), training ends and the network parameters are output. This invention can be widely deployed in various scenarios such as intelligent security monitoring, human-computer interaction platforms, medical rehabilitation monitoring, and auxiliary training for athletes.

Claims

1. A skeletal action recognition method based on dynamic graph convolution with fused topological priors, characterized in that, Includes the following steps: Step 1: Extract high-dimensional point cloud representations of the joints of each frame from the input bone sequence, analyze the distance changes between joints in the point cloud using the persistent cohomology method, generate a topological barcode to describe the evolution of the overall connected structure and ring structure of the skeleton, and convert the topological barcode into a continuous global topological feature vector through micro-mapping. Step 2: Decompose the skeletal input features into node flow sub-features and edge flow sub-features in the channel dimension, construct node flow dynamic graphs and edge flow dynamic graphs respectively, and introduce global topological feature vectors as bias terms in the self-attention mechanism of node flow and edge flow to guide the generation of dynamic adjacency relationships with topological structure information, thereby forming a topology-guided three-branch dynamic graph convolutional network. Step 3: Extract features from four modalities: joints, bones, joint motion, and bone motion. Calculate topological confidence using global topological feature vectors and generate affine transformation parameters independently for each modality based on the confidence. Perform non-competitive linear modulation updates on the features of each modality to obtain multimodal joint features that incorporate topological priors. Step 4: In the shallow stage of the network, global topological feature vectors are injected with high intensity to constrain the skeletal geometry. In the deep stage of the network, the injection intensity of topological information is gradually reduced through an inverse S-shaped continuous scheduling function, so that the network focuses on learning high-dimensional semantic discriminative features, thereby alleviating the problem of excessive feature smoothing in deep graph convolutional networks. Step 5: Input the features after multimodal modulation and stage scheduling into the classifier, and output the action recognition result.

2. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 1, characterized in that, In step 1, after modeling the skeletal joints as a high-dimensional point cloud, a simple complex sequence with varying distance thresholds is constructed based on the Euclidean distance between the joints. The topological feature sets of each dimension are extracted using a persistent homology method, and the features of all topological barcodes are fused into a continuous global topological feature vector through nonlinear mapping and splicing operations.

3. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 1, characterized in that, In step 2, the node flow dynamic graph is constructed as follows: the correlation score between the features of every two key points is calculated, and the global topological feature vector is added to the correlation score to form a correlation score with topological guidance. Then, the correlation score is converted into attention weights between nodes through a normalization function. The features of neighboring nodes are weighted and aggregated using the weights to update the feature representation of the current node.

4. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 1, characterized in that, In step 2, the construction of the edge flow dynamic graph is as follows: Each skeletal edge is represented as the feature difference between two endpoints. Then, the correlation score between the features of each pair of skeletal edges is calculated. The global topological feature vector is added to the correlation score and normalized to obtain the attention weight between edges. This weight is used to aggregate the features of neighboring edges to update the feature representation of the current edge.

5. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 1, characterized in that, In step 3, the topological confidence score is calculated by performing a nonlinear transformation on the global topological feature vector and then normalizing it using a normalization function. This score is used to characterize the reliability of the topological structure in different modes.

6. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 5, characterized in that, Step 3 is as follows: For each of the four modalities—joint, bone, joint motion, and bone motion—modal embedding features are extracted. Then, the topological confidence is multiplied element-wise with the modal embedding features. The multiplication results are then input into two independent linear transformation layers to generate scaling and translation factors for the modalities. Finally, the generated scaling and translation factors are used to perform channel-wise linear modulation updates on the original modal features to obtain the modal features after fusing the topological prior.

7. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 1, characterized in that, In step 4, the inverse S-shaped continuous scheduling function calculates the topology injection strength coefficient based on the current layer of the network. As the number of layers increases, the coefficient value of the inverse S-shaped continuous scheduling function decreases monotonically from close to 1 to close to 0, and the decreasing process exhibits an inverse S-shaped curve characteristic of being slow at first, fast in the middle, and slow again. The inverse S-shaped continuous scheduling function is jointly controlled by two preset parameters: the decay slope and the critical layer number. The decay slope determines how fast the coefficient decreases, and the critical layer number determines the number of network layers when the coefficient decreases to the middle value.

8. The skeletal action recognition method based on dynamic graph convolution with fused topological priors as described in claim 1, characterized in that, In step 5, the classifier uses the cross-entropy loss function to calculate the error between the predicted class and the true label, and adds a norm regularization term for the modulation parameters to prevent overfitting. The network parameters are then optimized through backpropagation.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

Images