GAE-Net: A gait-assisted enhancement network for video-based person re-identification

While video-based person re-identification has garnered significant attention and achieved substantial progress in recent years, existing methods predominantly depend on appearance information, making them vulnerable to changes in illumination and color. In contrast, the gait information concealed within walking postures exhibits robustness against appearance changes, while gait sequences provide additional temporal cues to enhance the temporal information in videos. Although some existing studies have utilized gait features to improve the robustness of Re-ID systems, the gap between gait and RGB data has not been adequately addressed. To bridge this gap, we propose a Gait-Assisted Enhancement Network (GAE-Net) designed to concurrently learn both appearance features and complementary gait features from RGB video sequences. Specifically, GAE-Net consists of two parts: Dynamic Two-stream Aggregation Network (DTA-Net) and Knowledge Distillation Fusion (KD-Fusion) framework. Firstly, DTA-Net applies two branches to extract appearance features and gait features, respectively. Moreover, the Dynamic Feature Aggregation (DFA) module is proposed to fuse gait and appearance features. Additionally, we propose Local Perception Complementary Distillation (LPCD) for Logit knowledge distillation. By leveraging LPCD, robust dark knowledge from the multimodal model (DTA-Net) can be effectively transferred to enhance the robustness of the single-modal model (Re-ID network). Through collaborative efforts between DTA-Net and LPCD, DGA-Net can acquire more comprehensive spatiotemporal representations. Extensive experiments on MARS and LS-VID demonstrate that our proposed method significantly outperforms other state-of-the-art methods.