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

An efficient biometric method for long-distance and non-contact human identification is gait recognition. However, partial occlusions and long-term gait variances brought on by aging, accidents, changing clothes, or environmental factors continue to make real-world deployment challenging. Current deep learning techniques typically rely on consistent walking patterns and clear visibility, which leads to poor performance in unrestricted situations. This study presents an Occlusion-Aware and Long-Term Adaptive (OALTA) architecture that combines a Memory-Augmented Continual Learning (MACL) approach with an Occlusion Detection and Refinement Module (ODRM).While MACL continually updates gait representations without catastrophic forgetting, the ODRM uses temporal attention and regional confidence estimation to identify trustworthy silhouette regions. To capture discriminative spatiotemporal gait features, a hybrid 3D-CNN and Temporal Transformer architecture is used. Superior robustness under occlusion and temporal gait alterations is demonstrated by experiments on CASIA-B, OU-MVLP, and a simulated Long-Term Gait dataset. Under extreme occlusion, the suggested architecture improves Rank-1 accuracy by more than 8% while preser performance.ving steady long-term recognition.