Abstract
<title>Abstract</title> <p>Gait recognition is a biometric technique that identifies individuals by analyzing walking patterns. Silhouette and skeleton are the two dominant input modalities, providing complementary information: silhouettes capture body shape variations, while skeletons describe joint-level structural cues. However, existing methods still face three challenges: insufficient feature extraction for capturing fine-grained body-part motions, weak structural priors for cross-modal alignment under skeleton noise, and limited global temporal modeling for distinguishing key motion fragments from redundant frames.To address these issues, we propose HierGait, a hierarchical multi-modal gait recognition framework with structural–temporal co-optimization. HierGait adopts a collaborative strategy of local temporal enhancement, structure-guided alignment fusion, and global context recalibration to learn robust gait representations. Specifically, (1) the Adaptive Multi-Scale Temporal Branch (AMSTB) captures local motion patterns at multiple temporal granularities and adaptively aggregates informative segments to enhance fine-grained temporal sensitivity; (2) the Multi-Channel Spatial–Joint Structure Fusion (MC-SJSF) module introduces a Gaussian correlation prior based on joint vertical positions to guide strip–joint alignment through cross-attention, reducing skeleton noise interference; and (3) the Global Temporal Context Fusion Module (GTCFM) compresses long sequences into sensitive and robust anchor frames and performs dual-branch modeling with gated fusion to effectively exploit global temporal information. Extensive experiments on CASIA-B demonstrate that HierGait achieves competitive performance, reaching Rank-1 accuracies of 98.8%, 96.3%, and 93.2% under NM, BG, and CL conditions, respectively, showing strong robustness against clothing variations.</p>