TAG: A Temporal Attentive Gait Network for Cross-View Gait Recognition

Recognizing a person from a distance using gait (i.e., walking pattern) is a challenging yet interesting biometric problem. Despite recent advancements in deep learning-based gait recognition (GR) research, learning discriminative gait temporal representation is still challenging because of delicate silhouette differences in the spatial domain. Aiming to address this issue, we propose a novel attention-based GR framework, namely, temporal attentive gait (TAG), which aims to refine the gait feature representation from the temporal dimension’s perspective in a comprehensive fashion. Our proposed TAG mainly consists of three modules, namely, short-term temporal feature learning (ST-TFL), hybrid multikernel temporal attention (H-MKTA), and multikernel temporal self-attention (MK-TSA), respectively. First, ST-TFL aims to capture local temporal contextual clues, facilitating the learning of short-period temporal motion patterns. Second, H-MKTA learns locally and globally distributed gait temporal information by adaptively capturing the multiscale temporal evolutions inside the gait sequential data. To refine the temporal attentive features learned by ST-TFL and H-MKTA, our MK-TSA learns global dependencies between gait temporal frames to recalibrate temporal weights using a self-attention mechanism. To further enhance the discriminative power of the gait feature representation, a multilevel (ML) framework is adopted, combining gait features from different levels of the backbone. Experiments conducted on three benchmark gait datasets, CASIA-B, Gait3D, and CCPG, demonstrate the strong potential of our TAG in learning effective gait representation under complex scenarios for recognition.