DGait: Robust gait recognition using dynamic ST-GCN with global aware attention

Gait recognition, a promising behavioral soft biometric technology, has a significant research area in security and computer vision. Nowadays, joint position-based approaches are of significant interest in gait recognition. ST-GCN, the spatio-temporal graph convolutional network, is employed on the joint stream to identify the gait feature from the spatial–temporal graph, prone to provide attention to dynamic information. Many methods utilize multi-scale operations to integrate long-range relationships among joints. However, these approaches fail to assign equal significance to all joints, leading to an incomplete perception of long-range joint connections. Furthermore, considering the joint stream solely may fail to extract the discriminative features produced by motion and bone structures. This paper presents a multi-stream dynamic spatio-temporal graph convolution (DSTGCN) approach with attention, denoted as DGait. It leverages bone and joint data from the spatial frames and joint-motion data from successive frames to early fusion of informative skeleton features. An improved HOP-extraction approach is introduced to provide equal importance to the relationship between further and closer joints while avoiding redundant dependencies. To address the limitations of ST-GCN, Global Aware Attention (GAA) is incorporated into the ST-GCN units, enhancing the capability for dynamically correlating the spatial–temporal joints. The suggested model exhibits remarkable accuracy on widely used CASIA-B, OUMVLP-Pose, and GREW datasets. The CASIA-B reveals an average accuracy of 96.94 %, 93.56 %, and 90.78 % for the normal walking, carrying-bag, and clothing conditions, respectively. The OUMVLP-Pose and GREW exhibit an average and rank-1 accuracy of 92.7 % and 72.6 %, respectively.