RepGCN: A Novel Graph Convolution-Based Model for Gait Recognition with Accompanying Behaviors

Currently, two challenges exist in the field of gait recognition: (1) there is a lack of gait datasets that include common accompanying behaviors during walking, and (2) it’s necessary to improve feature representation in skeleton sequence data for model-based approaches. To address these concerns, we focused on the study of accompanying behavior-based walking conditions and multiple views, and utilizes depth cameras to collect gait data. We presented the CDUT Gait dataset to investigate the impact of various accompanying behaviors on gait recognition performance. And we proposed a RepGCN, a novel graph convolution networks model with innovative residual strategy in the spatial module, as well as new spatio-temporal feature extraction modules. Experiments demonstrate that RepGCN achieves state-of-the-art performance on CDUT Gait with minimal model parameters compared to existing model-based approaches. The combination of depth cameras and RepGCN has potential applications in access control, smart home, and anti-terrorism areas.