TMPGait: Hybrid Architecture for Multi-Scale Spatiotemporal Feature Learning in Gait Recognition

Gait Recognition, as a key technology in the field of biometric identification, has significant application value in identity authentication scenarios in complex environments. To address the technical challenges of pedestrian gait being vulnerable to environmental interference in real-world scenarios, this paper proposes a gait recognition network TMPGait, designed for multi-scale spatiotemporal association enhancement. The model integrates the local feature extraction capability of CNN with the global semantic modeling strengths of ViT. Specifically, an Adaptive Spatiotemporal Mixture Extractor (ASME) is developed to capture discriminative spatiotemporal fine-grained correlations. Furthermore, the Temporal Context Co-Mining for Saliency Module (TCCS) is introduced to refine spatiotemporal dynamic representations and precisely model gait movement patterns. Experimental validation on out-the-lab datasets Gait3D and GREW, demonstrates that the proposed framework exhibits superior pedestrian identity recognition performance in complex environments.