Human Behavior Recognition Based on Multi-Stream Semantic Graph Convolutional Network
Compared to image-based behavior recognition methods,utilizing human skeleton information for recognition effectively overcomes the influence of complex backgrounds,lighting changes,and appearance variations.However,most mainstream human skeleton based methods encounter issues such as large parameter quantity and slow computational speed.To address these issues,this paper proposes a Multi-Stream lightweight Semantic Graph Convolutional Network(MS-SGN)for behavior recognition.The skeleton information is expressed as three data streams:bone length flow,joint flow,and fine-grained flow.Spatial features are extracted from the data stream embedded with semantic information through adaptive graph convolution.Time-domain features are extracted using multi-scale time-domain convolution with different kernels and expansion rates.Finally,the classification results of each stream are weighted and fused.The proposed method achieves an accuracy of 90.0％(X-Sub)and 95.83％(X-View)on large-scale dataset NTU60 RGB+D,and 83.4％(X-Sub)and 84.0％(X-View)on the dataset NTU120 RGB+D,respectively.Comparative experiments demonstrate that the proposed method offers better recognition accuracy than several main methods such as SGN and Logsin-RNN,while maintaining a lightweight network framework.
万方数据
维普
