FuseGait: A Gait Recognition Network Integrated FuseFormer and Spatio Temporal Convolutions

Gait recognition (GR), a biometric identification method require no user cooperation, and operate unnoticed from a distance, making it an increasingly attractive area of research for both academia and industry. While the recent methods have achieved excellent results in controlled indoor settings, their performance deteriorates significantly in outdoor environments. Previous approaches have attempted to enhance performance by increasing network depth, but these methods eventually face limitations. Since gait is a behavioral biometric, temporal features play a crucial role in identifying individuals. To address these challenges, we propose a GR framework, FuseGait, which integrates long-term and short-term temporal features with multi-dimensional spatial features. Our FuseGait utilizes P3D-based residual network to establish localized spatiotemporal receptive fields and employs a multi-scale visual encoder to extract comprehensive spatial information. Furthermore, we utilize the FuseFormer block to capture the long-term temporal dependencies within silhouette sequences and fully integrate them with the convolutional branch. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on both indoor dataset (CASIA-B) and outdoor dataset (Gait3D).