Spatial-Temporal Graph Convolutional Network for Skeleton-Based Gait Recognition

Abstract

Gait is one of the attractive biometrics used for discriminating between individuals. Recent success on pose estimator and graph convolutional network has inspired the research of skeleton-based gait recognition, where a skeleton sequence is represented as a graph for modeling in both spatial and temporal domains. However, the receptive field of temporal domain is limited due to simple connections with only the same joint on the inter-frame. This paper employs a temporal extended module (TEM) to extend temporal connections with multiple neighboring joints, and thus to extract additional features from the extended graph. Moreover, to better understand the contribution of neighboring joints on feature aggregation, we also exploit the performance under different number of neighbor subsets. Extensive experimental results on CASIA-B dataset show that our model could enhance the performance, and it achieves the mean accuracy with nearly 6% higher than the skeleton-based state-of-art methods.