Two-person skeleton interaction action recognition method and system based on spatiotemporal decoupling and attention compensation

Abstract

This invention proposes a spatiotemporally decoupled, attention-compensated method and system for recognizing two-person skeletonized interactive actions. It introduces positional encoding into video sequences of two-person interactive actions, then divides them into non-overlapping spatiotemporal blocks. Subsequently, it classifies these blocks based on the importance of their respective feature axes in the spatial and temporal dimensions, resulting in dominant and suppressed tokens. The effective information from the dominant token is then directed to the suppressed tokens for further refinement, generating refined features for the current dimension. Finally, the refined features from both dimensions are fused and reclassified to obtain the classification result. This invention proactively compensates for the effective information from the dominant tokens with occluded or disturbed weak response features, preventing the dilution of key local cues and significantly improving the model's robustness in real-world complex scenarios such as viewpoint changes and slight occlusion.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer vision technology, and in particular to a method and system for recognizing interactive actions of two-person skeletons with spatiotemporal decoupling and attention compensation. Background Technology

[0002] Two-person interaction action recognition aims to accurately classify specific interaction behaviors by modeling the coordinated movement relationship between two people in a video sequence, and has significant application value in fields such as intelligent monitoring and security early warning. Compared with RGB images, human skeleton data can explicitly represent the structural topology of human joints and has strong robustness to background interference.

[0003] Existing skeleton-based methods for recognizing two-person interactions typically divide the skeleton sequence into multiple spatiotemporal feature blocks (tokens) and employ a global self-attention mechanism to model the interactions of all feature blocks. However, in practical applications, these existing technologies suffer from the following significant drawbacks: ① Strong spatiotemporal coupling leads to excessively high computational complexity: Existing global attention mechanisms forcibly couple temporal and spatial dependencies, requiring the model to establish global pairwise correlations between all feature blocks. When processing long video sequences, this computational approach generates significant redundancy, resulting in a quadratic increase in computational complexity. This places extremely high demands on hardware computing power, making it difficult to meet the real-time requirements of edge devices.

[0004] ② Susceptible to local interference leading to loss of key features: In complex two-person interaction scenarios (such as those with viewpoint changes, partial occlusion, or subtle movements), the feature signals of some joints or body parts are very weak. Existing global unified modeling methods can dilute (i.e., overwhelm) these weak but crucial local discrimination clues with other redundant global information, resulting in a significant decrease in the model's accuracy and robustness in recognizing complex and similar interactive actions. Summary of the Invention

[0005] In view of the above situation, the main objective of this invention is to propose a spatiotemporally decoupled attention-compensated dual-person skeleton interaction action recognition method and system to solve the above-mentioned technical problems.

[0006] This invention proposes a spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method, the method comprising the following steps: Step 1: Obtain the video sequence of two-person interactive actions, extract the three-dimensional human skeleton joint point sequence data to obtain the skeleton sequence; encode the superimposed position of the skeleton sequence, and then divide it into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and map it into the initial spatiotemporal feature sequence. Step 2: Construct axis representations for the spatial and temporal dimensions of the initial spatiotemporal feature sequence, respectively, to obtain spatial axis representations and temporal axis representations; Step 3: Calculate the axis importance mask of the axis representation; based on the axis importance mask, divide the axis representation into dominant tokens with high response strength and suppressed tokens with weak response strength; Perform targeted compensation using the suppressed token as the query term and the dominant token as the key and value term to obtain the dimension compensation token; The compensation tokens of the dimensions are refined to obtain refined features; Step 4: Using spatial axis representation and temporal axis representation as inputs respectively, repeat step 3 to obtain spatial refinement features and temporal refinement features. Fuse the spatial refinement features and temporal refinement features, and then perform classification to obtain the classification results of the two-person interactive actions.

[0007] This invention also proposes a spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition system, wherein the system applies the spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method described above, and the system includes: The preprocessing module is used for: Obtain video sequences of two-person interactive actions, extract 3D human skeleton joint point sequence data, and obtain skeleton sequence; The spatiotemporal segmentation module is used for: The skeleton sequence is encoded at the superposition position, and then divided into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and mapped to the initial spatiotemporal feature sequence; The spatiotemporal feature decoupling module includes a spatial branching unit and a temporal branching unit with structure alignment; the spatial branching unit is used to obtain spatial refinement features, and the temporal branching unit is used to obtain temporal refinement features. The dual-branch gating fusion and classification module is used for: By fusing spatial refinement features with temporal refinement features and then classifying them, the classification results of two-person interactive actions are obtained.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention abandons the traditional global spatiotemporal attention mechanism and adopts a dual-branch decoupled architecture to model the temporal and spatial relationships separately. This scheme effectively eliminates global computational redundancy in long sequence processing, significantly reduces the number of model parameters and floating-point operations, and improves inference efficiency.

[0009] 2. This invention innovatively proposes a feature strength partitioning and directed feature compensation (DCA) operation. By actively compensating the effective information of the dominant token to the occluded and disturbed weak response features, the dilution of key local clues is avoided, and the robustness of the model in real-world complex scenarios such as viewpoint changes and slight occlusion is significantly improved.

[0010] 3. The continuity of the skeleton structure is ensured through context refinement (CRA) operation; at the same time, the dynamic gating fusion mechanism enables the model to adaptively adjust the emphasis ratio of temporal and spatial features according to the characteristics of different actions, thereby obtaining a more accurate semantic representation of actions.

[0011] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0012] Figure 1 This is the overall architecture of the spatiotemporal decoupling and attention-compensated attention-based dual-person skeleton interaction action recognition method proposed in this invention; Figure 2 This is a schematic diagram of the time-space token construction process of the present invention; Figure 3 This is a schematic diagram of the dual-branch compensated spatiotemporal attention of the present invention; Figure 4 This is a schematic diagram of the directional feature attention of the present invention; Figure 5 This is a schematic diagram illustrating the contextualized attention of the present invention; Figure 6 This is a schematic diagram of the spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition system proposed in this invention. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0015] Please see Figure 1 This embodiment provides a method for recognizing two-person skeleton interaction actions with spatiotemporal decoupling and attention compensation, the method including the following steps: Step 1: Obtain the video sequence of two-person interactive actions, extract the three-dimensional human skeleton joint point sequence data to obtain the skeleton sequence; encode the superimposed position of the skeleton sequence, and then divide it into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and map it into the initial spatiotemporal feature sequence. When performing a two-person action recognition task, the system first acquires the original video sequence and then uses a pose estimation algorithm to extract key points of the human skeleton. The input skeleton sequence is defined as follows: in, These represent batch size, coordinate dimension, time step, number of key points, and number of interacting users, respectively. To enable the model to perceive the topological relationships of different key points in space and the temporal order of actions, the system first superimposes positional encoding (PE) onto the input skeleton sequence to obtain the enhanced feature sequence. The corresponding process follows this relationship: ; in, This represents the skeleton sequence after overlaying position encoding. Represents the skeleton sequence. This indicates an overlay position encoding operation; Subsequently, a spatiotemporal token construction (STC) is performed on the skeleton sequence after overlay position encoding. For example... Figure 2 As shown, the system divides long sequences into a series of fixed-size, non-overlapping spatiotemporal windows.

[0016] During the construction of the spacetime token, the size of its spacetime window is defined as follows: ,in Indicates the length of the time window. Indicates the granularity of spatial joint partitioning. Represents the entity dimension. In this embodiment, the spatiotemporal window uses... As a window configuration, every 20 frames constitute a time block, with each joint acting as an independent unit, while simultaneously preserving information about both individuals. Through this mapping process, the original sequence is transformed into a total of... The local block token space is defined by several local block tokens forming an initial spatiotemporal feature sequence. The total number of tokens is determined by the product of the number of temporal sub-tokens and the number of spatial sub-tokens. , This represents the total number of tokens in the local block. Indicates the number of time-based sub-tokens. Indicates the number of space sub-tokens.

[0017] Step 2: Construct axis representations for the spatial and temporal dimensions of the initial spatiotemporal feature sequence, respectively, to obtain spatial axis representations and temporal axis representations; like Figure 3 As shown, after token construction is completed, this invention no longer uses the traditional global coupled attention mechanism. Instead, it sets up two parallel and topologically consistent spatial and temporal branches to construct a dual-branch compensated spatiotemporal attention mechanism for separate modeling, thus eliminating global computational redundancy in long sequence processing.

[0018] Step 3: Calculate the axis importance mask of the axis representation; based on the axis importance mask, divide the axis representation into dominant tokens with high response strength and suppressed tokens with weak response strength; Perform targeted compensation using the suppressed token as the query term and the dominant token as the key and value term to obtain the dimension compensation token; The compensation tokens of the dimensions are refined to obtain refined features; Taking spatial branching as an example, such as Figure 3 As shown on the left, the initial query terms and initial value terms are first generated using a 1×1×1 three-dimensional convolutional layer. The corresponding process has the following relationship: ; in, This represents the initial query term. Indicates the initial value item. This indicates a 1×1×1 three-dimensional convolutional layer used to generate the initial query terms. This represents a 1×1×1 three-dimensional convolutional layer used to generate the initial value terms; To significantly reduce the computational cost of processing long sequences while preserving the global receptive field, average pooling aggregation is performed on the initial query terms to obtain a spatial axis representation. The corresponding process has the following relationship: ; in, The space axis represents lightweight design. This indicates that average pooling aggregation is performed on both the time and entity dimensions.

[0019] This design lays the data foundation for efficient feature segmentation and directional compensation on the spatial axis. The temporal branch also uses symmetrical logic for feature initialization, aiming to capture dynamic temporal dependencies in subsequent steps.

[0020] In the process of two-person interactive action recognition, due to complex background interference, overlapping of interactive subjects, and inconsistencies in action amplitude, the contribution of different key points or action time segments to the final classification varies significantly. To enable the model to focus on the most discriminative cues, this invention introduces an axis importance estimation mechanism within the decoupled branch. For example... Figure 3 As shown, in the spatial branch, an axis importance estimation algorithm is used to process the lightweight spatial axis representation. Specifically, the system learns the discrimination weights of the current axis through a linear mapping layer and generates a spatial axis importance mask using a sigmoid activation function. The corresponding process has the following relationship: ; in, A mask representing the axial importance of spatial dimensions. This represents the Sigmoid activation function. Represents a linear mapping; Based on the importance mask of spatial dimensions, the feature space is explicitly partitioned mathematically, thereby adaptively decoupling the features into two parts: the "dominant token" with a strong response and the "suppressed token" with a weak or disturbed response. The dominant token includes the dominant matching feature and the dominant value feature. The dominant matching feature represents the action information with high discriminative power in the current axis. The process of obtaining the corresponding feature through element-wise multiplication of the importance mask and the original feature is related by the following formula: ; in, Indicates the dominant matching feature. This represents element-wise multiplication; Correspondingly, the suppressed token represents a part with a low response or semantic ambiguity, and the corresponding process has the following relationship: ; in, Indicates a suppressed token; This partitioning method ensures that the model can clearly identify which regions need "focus" and which regions need "compensation," avoiding the dilution of discriminative signals caused by globally indiscriminate modeling. While partitioning the query features, to provide complete data support for subsequent targeted compensation, the initial value terms are also rearranged and flattened to obtain rearranged value representations. Then, the rearranged value representations are filtered to extract dominant value features containing complete spatiotemporal context. The corresponding process has the following relationship: ; in, Indicates the dominant value characteristic, This represents the rearranged value. This value vector-based filtering mechanism ensures that the information transmitted via the attention mechanism originates from regions with strong and reliable responses. Notably, the temporal branch follows a completely symmetrical processing logic to the spatial branch. By calculating a temporal importance mask and performing feature segmentation on the temporal axis, adaptive locking of key action frames in the temporal dimension is achieved. Through this stage of processing, the system constructs a clear distribution of strong and weak features in both the decoupled spatiotemporal dimensions, laying the structural foundation for the subsequent directional compensation module.

[0021] After feature segmentation is completed through axis importance estimation, directional compensation attention is introduced to perform directional compensation for suppressed tokens. The specific internal structure is detailed in the appendix. Figure 4 As shown. The design motivation for this directional compensating attention is to utilize the highly discriminative cues contained in the dominant token to provide directional information supplementation to suppressed tokens that are obscured or interfered with by interaction. In the implementation of directional compensating attention in the spatial branch, two sets of query vectors and key vectors are mapped from the suppressed tokens and the dominant matching features respectively through learnable linear projections, i.e. and in, This represents the first query vector. Represents the first key vector. This represents the second query vector. This represents the second key vector.

[0022] To ensure the accuracy of information compensation and suppress the influence of environmental noise, the directional compensation attention employs a unique dual-path parallel matching design. First, a basic matching relationship between the suppressed token and the dominant token is established through a Softmax branch. Then, dot-product scaling attention is performed on the first query vector and the first key vector, and a learnable spatial axis bias term is introduced during the dot-product scaling attention process to obtain a normalized correction map. The corresponding process has the following relationship: ; in, This represents the normalized correction plot. Represents the normalized exponential function, This indicates the transpose operation. This represents the learnable spatial axis deviation term. Indicates the scaling factor; Simultaneously, to filter out spurious associations caused by irrelevant movements, a gated correlation branch is introduced. The second query vector and the second key vector are scaled by a dot product, followed by average pooling and activation operations to obtain the gated graph. The corresponding process has the following relationship: ; in, Represents a gated graph. This indicates an average pooling operation; Finally, the system performs element-wise product of the two results to obtain the directed compensation attention map, and the corresponding process has the following relationship: ; in, This represents a targeted, compensated attention map.

[0023] Based on this targeted compensation attention map, the dominant value features carrying complete contextual information are propagated in a targeted manner to calculate the spatial dimension compensation result and obtain the spatial dimension compensation token. The corresponding process has the following relationship: ; in, A compensation token representing the spatial dimension.

[0024] This mechanism ensures that only dominant information that possesses both semantic relevance and high reliability flows to the weak response region. For the time branch, a completely symmetrical computational logic is followed, and directional compensation of key action frame sequences is achieved through directional feature compensation operations on the time axis. In this way, the present invention can effectively solve the problem of missing discrimination clues caused by cross-person interference or partial occlusion in two-person interaction scenarios, and significantly improve the robustness of the recognition algorithm in complex real-world environments.

[0025] After completing the targeted feature compensation operation, contextual refinement attention is introduced to further enhance the compensated feature sequence through contextual refinement operations. For example... Figure 5 As shown, contextual refinement attention is deployed after directional compensation attention. Its core objective is to smooth the compensated feature representation through a self-attention mechanism and enhance its contextual consistency on the current axis.

[0026] In this layer, the system first uses the linear mapping layer to generate query terms, key terms, and value terms for refinement processing; Unlike the directional matching logic of directional feature compensation operations, context refinement operations aim to capture deep semantic relationships within feature blocks. Therefore, a learnable spatial axis bias term, consistent with directional compensation attention, is introduced during the dot product scaling attention process to inject prior information about specific spatial structures or temporal rhythms, resulting in an attention weight map. The corresponding process has the following relationship: ; in, Represents the attention weight map. Indicates the query item. Indicates key item, This represents the learnable spatial axis deviation term; Subsequently, by calculating and weighting the values ​​using the attention weight map, preliminary refined features are obtained. The corresponding process has the following relationship: ; in, This indicates preliminary detailed features; The initial refined features are passed sequentially through layer normalization and a feedforward neural network, followed by residual connections, to obtain spatial refined features. The corresponding process has the following relationship: ; in, This indicates spatial refinement features. Representation layer normalization, This represents a feedforward neural network.

[0027] This design ensures that while enhancing feature representation capabilities, the continuity and stability of the skeletal structure are not compromised. In the temporal branch, the system follows a modeling logic completely symmetrical to the spatial branch, sequentially performing axis importance estimation, directional feature compensation, and context refinement. The temporal branch focuses on capturing the dynamic evolution of token blocks along the timeline, further enhancing the model's perception of changes in action rhythm by augmenting key action segments.

[0028] Step 4: Using spatial axis representation and temporal axis representation as inputs respectively, repeat step 3 to obtain spatial refinement features and temporal refinement features. Fuse the spatial refinement features and temporal refinement features, and then perform classification to obtain the classification results of the two-person interactive actions.

[0029] After the spatial and temporal branches output the spatial and temporal refined features respectively, the feature fusion stage, as shown in the lower right corner of Figure 1, begins. This invention no longer employs the traditional simple weighting or summation method, but instead designs a dynamic gating fusion mechanism to adaptively balance the contributions of the two feature paths. First, the spatial and temporal refined features are concatenated along the channel dimension to obtain the concatenated features. The corresponding process has the following relationship: ; in, Indicates splicing characteristics, This indicates a splicing operation. Indicates time refinement features; The weights are then adaptively calculated using a gated network consisting of two 1×1×1 convolutional layers. Specifically, the concatenated features are input into the gated network and then activated to obtain spatial and temporal weights. The corresponding process follows the following relationship: ; in, Indicates spatial weights, Indicates time weighting, Indicates a gating network; Finally, spatial and temporal weights are used to perform a weighted sum of the spatial and temporal refinement features to obtain the fused features. The corresponding process has the following relationship: ; in, Indicates fusion characteristics; This mechanism enables the model to dynamically determine, based on specific input samples, whether spatial structural information or temporal dynamic information is more discriminative for recognizing the current action. Finally, the fused features are processed through global pooling and linear projection layers to output the classification results of the two-person interaction actions.

[0030] Please refer to Figure 6 This embodiment also provides a spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition system, wherein the system applies the spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method described above, and the system includes: The preprocessing module is used for: Obtain video sequences of two-person interactive actions, extract 3D human skeleton joint point sequence data, and obtain skeleton sequence; The spatiotemporal segmentation module is used for: The skeleton sequence is encoded at the superposition position, and then divided into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and mapped to the initial spatiotemporal feature sequence; The spatiotemporal feature decoupling module includes a spatial branching unit and a temporal branching unit with structure alignment; the spatial branching unit is used to obtain spatial refinement features, and the temporal branching unit is used to obtain temporal refinement features. The dual-branch gating fusion and classification module is used for: By fusing spatial refinement features with temporal refinement features and then classifying them, the classification results of two-person interactive actions are obtained.

[0031] In summary, regarding the optimization of the computational architecture, this invention effectively eliminates the computational redundancy of global correlations in long sequence processing by decomposing the strongly spatiotemporally coupled global attention into a parallel dual-branch decoupled architecture. Experimental data shows that the model parameters of this invention are only 5.25M, and the computational complexity (GFLOPs) is controlled at around 26.69G. Compared with traditional mainstream models based on graph convolutional networks (GCN) or hybrid Transformers, this invention maintains extremely high operating efficiency while significantly reducing computational power consumption, making it particularly advantageous for practical deployment on edge or end-device devices with limited computing power.

[0032] Regarding the accuracy of action recognition, this invention demonstrates superior performance compared to existing technologies on multiple large-scale public datasets. For example, on the highly representative NTU RGB+D 120 dataset, this invention achieved high recognition accuracies of 91.18% and 92.65% under the X-Sub and X-Set evaluation protocols, respectively. This significant performance improvement proves that the model not only possesses the ability to perceive the overall action framework but also accurately captures and distinguishes subtle structural changes and temporal features of similar actions.

[0033] In terms of robustness in complex real-world scenarios, the directional feature compensation mechanism proposed in this invention demonstrates significant practical value. In real-world two-person interaction scenarios, due to visual blind spots, motion occlusion, or cross-person interference, the motion features of some body parts are often very weak. The context refinement mechanism of this invention can proactively identify these weak response features and use effective clues in the dominant token to provide targeted supplementation and enhancement. Experiments show that even under extreme conditions such as strong viewpoint changes, slight local occlusion, small motion amplitude, and video quality fluctuations, the performance degradation of this invention is far less than that of similar algorithms, effectively suppressing noise interference and maintaining extremely high stability in two-person adversarial interaction recognition.

[0034] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present 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 modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A two-person skeleton interaction action recognition method of spatio-temporal decoupled compensation attention, characterized in that, The method includes the following steps: Step 1: Obtain the video sequence of two-person interactive actions, extract the three-dimensional human skeleton joint point sequence data to obtain the skeleton sequence; encode the superimposed position of the skeleton sequence, and then divide it into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and map it into the initial spatiotemporal feature sequence. Step 2: Construct axis representations for the spatial and temporal dimensions of the initial spatiotemporal feature sequence, respectively, to obtain spatial axis representations and temporal axis representations; Step 3: Calculate the axis importance mask of the axis representation; based on the axis importance mask, divide the axis representation into dominant tokens with high response strength and suppressed tokens with weak response strength; Perform targeted compensation using the suppressed token as the query term and the dominant token as the key and value term to obtain the dimension compensation token; The compensation tokens of the dimensions are refined to obtain refined features; Step 4: Using spatial axis representation and temporal axis representation as inputs respectively, repeat step 3 to obtain spatial refinement features and temporal refinement features. Fuse the spatial refinement features and temporal refinement features, and then perform classification to obtain the classification results of the two-person interactive actions.

2. The dual human skeleton interaction action recognition method of claim 1, wherein, In step 1, the positions of the skeleton sequence overlays are encoded, and then divided into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and mapped to the initial spatiotemporal feature sequence. Specifically, this includes the following steps: The encoding of the positions of the skeleton sequence overlay follows the following relationship: ； wherein, represents a superimposed position-encoded skeleton sequence, represents a skeleton sequence, represents a superimposition position encoding operation; Using a window of a preset size, the skeleton sequence after superimposed position encoding is divided into blocks of fixed size that do not overlap, resulting in local block tokens. Several local block tokens form the initial spatiotemporal feature sequence.

3. The dual human skeleton interaction action recognition method of claim 2, wherein, In step 2, a spatial axis representation is constructed for the spatial dimension of the initial spatiotemporal feature sequence, specifically including the following steps: The initial spatiotemporal feature sequence is used to generate initial query terms and initial value terms using a 1×1×1 three-dimensional convolutional layer. The corresponding process has the following relationship: ； wherein, denotes an initial query term, denotes an initial value term, denotes a 1x1x1 three-dimensional convolution layer for generating the initial query term, denotes a 1x1x1 three-dimensional convolution layer for generating the initial value term; Performing average pooling aggregation on the initial query terms yields a spatial axis representation, and the corresponding process follows the following relationship: ； in, The space axis represents lightweight design. This indicates that average pooling aggregation is performed on both the time and entity dimensions.

4. The spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method according to claim 3, characterized in that, In step 3, the axis importance mask of the axis representation is calculated; based on the axis importance mask, the axis representation is divided into dominant tokens with high response strength and suppressed tokens with weak response strength, specifically including the following steps: By sequentially performing linear mapping and activation operations on the spatial axis representations, an axis importance mask in the spatial dimension is generated. The corresponding process has the following relationship: ； in, A mask representing the axial importance of spatial dimensions. This represents the Sigmoid activation function. Represents a linear mapping; The dominant token includes dominant matching features and dominant value features. The dominant matching features are obtained by performing element-wise multiplication between the spatial axis representation and the axis importance mask of the spatial dimension. The corresponding process has the following relationship: ； in, Indicates the dominant matching feature. This represents element-wise multiplication; The initial value items are rearranged and flattened to obtain the rearranged value representation; The rearranged value representation is multiplied element-wise with the spatial dimension's axis importance mask to obtain the dominant value features. The corresponding process has the following relationship: ； in, Indicates the dominant value characteristic, This represents the rearranged value. Taking the complement of the axis importance mask and then multiplying it element-wise with the spatial axis representation yields the suppressed token. The corresponding process has the following relationship: ； in, This indicates a suppressed token.

5. The spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method according to claim 4, characterized in that, In step 3, targeted compensation is performed using the suppressed token as the query term and the dominant token as the key and value term to obtain the dimension compensation token. This specifically includes the following steps: The suppressed token and the dominant matching feature are mapped to two sets of query vectors and two sets of key vectors through learnable linear projections, respectively. The two sets of query vectors include a first query vector and a second query vector, and the two sets of key vectors include a first key vector and a second key vector; dot product scaling attention is performed on the first query vector and the first key vector, and a learnable spatial axis bias term is introduced in the dot product scaling attention process to obtain a normalized correction map; The second query vector and the second key vector are scaled by dot product, and then average pooling and activation operations are performed in sequence to obtain the gated graph. The corresponding process is as follows: By fusing the normalized correction map with the gating map, a directional compensation attention map is obtained. The corresponding process has the following relationship: ； in, This represents the normalized correction plot. This represents a targeted, compensated attention map. Represents a gating diagram; By weighting the dominant value features using a targeted compensation attention map, a compensation token for the spatial dimension is obtained. The corresponding process has the following relationship: ； in, A compensation token representing the spatial dimension.

6. The spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method according to claim 5, characterized in that, The two sets of query vectors include a first query vector and a second query vector, and the two sets of key vectors include a first key vector and a second key vector. Dot product scaling attention is performed on the first query vector and the first key vector, and a learnable spatial axis bias term is introduced during the dot product scaling attention process to obtain a normalized correction map. The corresponding process has the following relationship: ； in, Represents the normalized exponential function, This indicates the transpose operation. This represents the first query vector. Represents the first key vector. This represents the learnable spatial axis deviation term. This represents the scaling factor.

7. The method for recognizing two-person skeleton interaction actions with spatiotemporal decoupling and attention compensation according to claim 6, characterized in that, The second query vector and the second key vector are scaled by a dot product, and then average pooling and activation operations are performed sequentially to obtain the gated graph. The corresponding process has the following relationship: ； in, This represents the second query vector. Represents the second key vector. This indicates the average pooling operation.

8. The spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method according to claim 7, characterized in that, In step 4, the compensation tokens for the spatial dimension are refined to obtain spatial refinement features, specifically including the following steps: Linear mapping is performed on the compensation tokens of the spatial dimension to obtain query items, key items, and value items; Dot product scaling attention is performed on the query and key items, and a learnable spatial axis bias term is introduced during the dot product scaling attention process to obtain the attention weight map. The corresponding process has the following relationship: ； in, Represents the attention weight map. Indicates the query item. Indicates key item, This represents the learnable spatial axis deviation term; By using an attention weight map to weight the values, preliminary refined features are obtained. The corresponding process follows the following relationship: ； in, This indicates preliminary detailed features; The initial refined features are passed sequentially through layer normalization and a feedforward neural network, followed by residual connections, to obtain spatial refined features. The corresponding process has the following relationship: ； in, This indicates spatial refinement features. Representation layer normalization, This represents a feedforward neural network.

9. The method for recognizing two-person skeleton interaction actions with spatiotemporal decoupling and attention compensation according to claim 7, characterized in that, In step 5, the spatial refinement features and temporal refinement features are fused and then classified to obtain the classification result of the two-person interactive actions. Specifically, this includes the following steps: Spatial refinement features and temporal refinement features are concatenated along the channel dimension to obtain concatenated features. The corresponding process has the following relationship: ； in, Indicates splicing characteristics, This indicates a splicing operation. Indicates time refinement features; The concatenated features are input into a gated network and then activated to obtain spatial and temporal weights. The corresponding process has the following relationship: ； in, Indicates spatial weights, Indicates time weight, Indicates a gating network; By using spatial and temporal weights to perform a weighted summation of the spatial and temporal refinement features, the fused features are obtained. The corresponding process follows the following relationship: ； in, Indicates fusion characteristics; The fused features are processed through global pooling and linear projection layers to obtain the classification results of the two-person interaction actions.

10. A spatiotemporally decoupled attention-compensated dual-person skeleton interaction action recognition system, characterized in that, The system employs the spatiotemporal decoupling and attention-compensated dual-person skeleton interaction action recognition method as described in any one of claims 1 to 9, and the system comprises: The preprocessing module is used for: Obtain video sequences of two-person interactive actions, extract 3D human skeleton joint point sequence data, and obtain skeleton sequence; The spatiotemporal segmentation module is used for: The skeleton sequence is encoded at the superposition position, and then divided into non-overlapping spatiotemporal blocks according to the set time window and spatial granularity, and mapped to the initial spatiotemporal feature sequence; The spatiotemporal feature decoupling module includes a spatial branching unit and a temporal branching unit with structure alignment; the spatial branching unit is used to obtain spatial refinement features, and the temporal branching unit is used to obtain temporal refinement features. The dual-branch gating fusion and classification module is used for: By fusing spatial refinement features with temporal refinement features and then classifying them, the classification results of two-person interactive actions are obtained.

Images