Learning Occlusion-Robust Pedestrian Representations via Uncertainty-Guided Feature Pruning

Abstract

Occlusion and background interference remain major challenges for pedestrian re-identification in urban traffic environments. Inspired by uncertainty-aware CLIP-based frameworks, this paper introduces an uncertainty-guided feature selection mechanism that adjusts the contribution of local visual regions and semantic cues according to their estimated reliability. The proposed method is evaluated on two autonomous driving datasets with both real-world and synthetic occlusion patterns, covering occlusion ratios from 20% to 60%. Comparisons are conducted against attention-based and part-based ReID methods, including PCB, OSNet, and transformer-based attention models. The proposed approach achieves mAP improvements ranging from 4.5% to 6.2% under severe occlusion conditions, while maintaining comparable performance in fully visible settings.