Human gait recognition using dense residual network and hybrid attention technique with back-flow mechanism

Abstract

Gait recognition is a promising biometric technique for person identification, either as a standalone method or in combination with other modalities. A major challenge lies in extracting robust gait features from silhouettes that remain invariant to variation in clothing, carried objects, and camera viewpoints. Recent advances using attention-based convolutional neural networks (CNNs) have improved gait recognition performance; however, many existing methods struggle to preserve semantic information across network layers due to information loss during the stages of downsampling. To address this issue, a novel residual dense back-flow attention network (RDBA-Net) is proposed, which integrates dual-branch hybrid self-attention network (DHSAN) modules with densely connected residual dense blocks (RDBs), and the output features are concatenated in a back-flow direction. This design enables effective learning of discriminative gait features by leveraging attention cues at both spatial-level and temporal-level from silhouette sequences. Furthermore, back-flow mechanism enhances feature learning in earlier layers by reusing refined semantic information from deeper layers. Experimental evaluations on two benchmark datasets, CASIA B and OU-MVLP, demonstrate that RDBA-Net, achieves notable improvements in accuracy compared to existing state-of-the-art methods, with gains up to 91.6% on CASIA B and 89.2% on OU-MVLP under challenging conditions.