Pyr-HGCN: Pyramid Hybrid Graph Convolutional Network for Gait Emotion Recognition

Li Jing, Guangchao Yang, Yunfei Yin

Abstract: Gait emotion recognition (GER) plays a crucial role in identifying human emotions. Most previous methods apply Spatial-Temporal Graph Convolutional Networks (ST-GCN) to recognize emotions. However, these methods suffer from two serious problems: (1) they ignore the fact that the similarity between emotions with the similar emotional intensity. Consequently, fine-grained information from the low-layer network, which is essential for accurate emotion recognition, is lost. (2) They ignore that the expression of emotion is a continuous process, that is, failing to model the temporal dimension effectively. To address these issues, a novel Pyramid Hybrid Graph Convolutional Network (Pyr-HGCN) is proposed for GER. Specifically, we first introduce and enhance the pyramid structure in GER to compensate for the missing fine-grained information of the ST-GCN structure. Additionally, we design a novel Spatial-Temporal Hybrid Convolution (STHC) block, which can indirectly and simultaneously capture complex spatio-temporal correlations in long-term regions. Extensive experiments and visualizations were performed on several benchmarks, with an accuracy improvement of 0.01 to 0.02 demonstrating the effectiveness of our approach against state-of-the-art competitors.