融合双注意力机制与空间重建卷积的人体动作识别方法OA

Skeleton based human action recognition is an important research direction in the field of computer vision and has wide applications in intelligent monitoring,human-computer interaction,and other fields.However,most existing methods fail to adequately capture key joint information and human motion patterns in complex scenarios,and they often generate a large number of redundant features during joint feature extraction.To address these issues,a human action recognition method that integrates a dual-attention mechanism with spatial reconstruction convolution is proposed.First,a graph convolutional structure is employed to model joint and bone information within the same frame of the skeleton data,thereby capturing the spatial relationships among joints.Subsequently,channel and spatial attention weighting is applied to the feature maps,enabling the model to adaptively focus on critical joints and motion patterns and enhancing the expressiveness of the extracted features.On this basis,the model further performs holistic joint modeling within each frame and leverages spatial and channel reconstruction convolutions to explore inter-frame dependencies,thus improving its temporal modeling capability.Experimental results show that the proposed model significantly enhances the feature extraction ability and computational efficiency of the Baseline SGN(semantic-guided network).It achieves higher accuracy in human action recognition tasks,with particularly notable improvements in complex scenarios and in handling long-range dependencies.On the NTU RGB+D validation set,the model demonstrates clear advantages over prior approaches,achieving 89.2%accuracy under the cross-subject setting and 95.8%under the cross-view setting,thereby validating its effectiveness and superiority.