Long-Term Adaptive and Occlusion- Aware Gait Recognition: A Spatiotemporal Continual Learning Framework

Gait recognition has emerged as an effective biometric approach for non-contact and long-distance human identification. However, real-world deployment remains difficult due to partial occlusions and long-term gait variations caused by aging, injuries, clothing changes, or environmental conditions. Existing deep learning approaches generally assume clear visibility and stable walking patterns, resulting in degraded performance under unconstrained environments. This paper proposes an Occlusion-Aware and Long-Term Adaptive (OALTA) framework that combines an Occlusion Detection and Refinement Module (ODRM) with a Memory-Augmented Continual Learning (MACL) strategy. The ODRM identifies reliable silhouette regions using regional confidence estimation and temporal attention, while MACL continuously updates gait representations without catastrophic forgetting. A hybrid 3D-CNN and Temporal Transformer architecture is employed to capture discriminative spatiotemporal gait features. Experiments conducted on CASIA-B, OU-MVLP, and a simulated Long-Term Gait dataset demonstrate superior robustness under occlusion and temporal gait changes. The proposed framework achieves more than 8% improvement in Rank-1 accuracy under severe occlusion conditions while maintaining stable long-term recognition performance.