Uncertainty-Aware Pedestrian Re-Identification in Autonomous Driving Perception Pipelines

Abstract

Pedestrian re-identification models trained on one city often degrade when deployed in  another due to changes in camera placement, lighting, and pedestrian appearance patterns.  Building on CLIP-based uncertainty modal modeling, this work proposes an uncertainty-weighted  adaptation strategy that transfers vision–language embeddings across cities while down weighting unreliable pseudo-labels. The method combines (i) entropy-based sample filtering, (ii)  uncertainty-aware class prototype refinement, and (iii) consistency regularization between image  and text embeddings. Experiments are conducted on cross-city splits constructed from autonomous  driving data, totaling 320,000 images and 38,000 identities across three urban domains.  Compared with OSNet, TransReID, and CLIP-derived ReID baselines under standard  unsupervised domain adaptation settings, the proposed method improves target-domain mAP by  4.0%–6.3% and rank-1 accuracy by 3.2%–5.1%, with the largest gains observed in nighttime and  rain subsets.