基于动作特异图卷积与注意力机制的行为识别方法OA

To address the challenges of human action recognition on resource-constrained edge devices,en-hance model robustness,and alleviate confusion among similar actions,a behavior recognition network based on action-specific graph convolution and attention mechanism network(ASGCA-Net)is proposed.In the tem-poral dimension,a multi-scale temporal convolutional attention module is designed to simultaneously capture short-term local motion patterns and long-term global dependencies,thereby strengthening the temporal mod-eling capability of action sequences.The attention mechanism is further employed to learn the weight impor-tance of each channe.In the spatial dimension,implicit edges are introduced into the fixed topology to supple-ment joint dependencies across physical connections,and a gating mechanism is used to adaptively adjust the weights of structural and implicit edges,enabling action-specific feature modeling for different action types.Finally,ASGCA-Net is evaluated on the NTU RGB+D and the NTU RGB+D 120 datasets.The results show that compared with the baseline network 2s-AGCN,ASGCA-Net achieves substantial accuracy improve-ments on both datasets.