STA-GCN:Spatial Temporal Adaptive Graph Convolutional Network for Gait Emotion Recognition

Graph Convolutional Neural Networks (GCNs) recently have been widely used in Gait Emotion Recognition (GER). However, the existing GCNs-based GER methods have two drawbacks that limit the ability to learn discriminative feature. In spatial modeling, context-sensitive affective feature of joint is under-extracted due to the neglect of implicit connection. In temporal modeling, multi-scale temporal feature of joint motion is under-extracted or aggregated rigidly. In this paper, we propose a novel Spatial-Temporal Adaptive Graph Convolutional Network (STA-GCN) where two main modules are introduced, respectively. Spatial Feature Learning Module (SFLM) infers context-sensitive joint implicit connection and adaptively aggregates spatial feature mined from implicit and explicit connection. Temporal Feature Learning Module (TFLM) extracts and adaptively aggregates multi-scale temporal feature of joint motion. It is worth mentioning that we first pre-train the model using hand-crafted affective feature and counterpart gait. Experimental results demonstrate our STA-GCN outperforms state-of-the-art methods in two tasks.