TDGait: Modeling Temporal Dynamics for Gait Recognition

Gait recognition has emerged as a promising biometric identification method due to its non-cooperative and nonintrusive nature. However, existing methods often exhibit limitations in effectively modeling the temporal dynamics of human gait, which are crucial for robust identification in real-world scenarios. This paper introduces TDGait, a novel network specifically designed to address these limitations by explicitly modeling temporal dynamics of gait patterns. The core innovations of TDGait are twofold: 1) a Temporal Difference Module (TDM) is proposed to efficiently capture frame-to-frame variations, focusing on dynamic changes rather than static appearance. 2) a Residual Dynamic Block (RDB) is introduced to seamlessly integrate TDM without impairing spatial feature learning capacity. This block preserves spatial learning via a residual pathway while enhancing dynamic information through the embedded TDM. TDGait hierarchically combines spatial feature extraction with this explicit temporal dynamics modeling using TDM within RDBs. Comprehensive experiments on three benchmark datasets demonstrate state-of-the-art performance, achieving Rank-1 accuracies of 83.38% on SUSTech1K, 72.5% on Gait3D, and 79.1% on GREW. Extensive ablation studies validate the effectiveness and efficiency of the TDM and RDB components in capturing temporal dynamics. This work provides a new perspective on modeling the temporal dynamics of human gait, contributing to more robust biometric systems resilient to real-world variations.