Diffusion prior–guided hierarchical state–space gait recognition

Abstract

Silhouette-based gait recognition remains sensitive to view changes, clothing variation, carried objects, and local occlusion because these factors disrupt both long-range motion patterns and local body-part relations. This paper proposes a diffusion-prior-guided hierarchical state-space framework for robust and efficient gait recognition. The framework first uses a hierarchical state-space encoder (H-SSM) to model global, upper-body, and lower-body gait dynamics with selective bidirectional scanning and adaptive hierarchical fusion. During teacher training, a diffusion prior guidance module (DPG) injects intermediate features from a frozen SD-v1.5 U-Net into H-SSM tokens through cross-attention and gated residual fusion, providing soft structural compensation without pixel-level restoration. A structured knowledge distillation strategy (SKD) then transfers the diffusion-enhanced hierarchy-part relation graphs to a lightweight student, so the diffusion branch is removed during inference. Experiments on CASIA-B and OU-MVLP, together with controlled occlusion analysis, show that the proposed student achieves competitive accuracy and robustness while retaining low inference cost. On CASIA-B, it obtains 94.9% average Rank-1 accuracy and 89.5% accuracy under clothing variation, with 4.8M trainable parameters and 3.1 GFLOPs. These results demonstrate an effective training-stage structural enhancement and inference-stage lightweight deployment strategy for benchmark covariate and controlled occlusion settings.