Temporal Reasoning for Pedestrian Re-Identification under VisionLanguage Uncertainty

Reliable identity association is a prerequisite for deploying pedestrian re-identification in autonomous driving systems. Drawing on CLIP-based uncertainty modal modeling, this work investigates the role of uncertainty estimation in vision–language similarity measurement. Uncertainty information is incorporated into the matching stage to suppress ambiguous identity associations. Experiments are conducted on large-scale autonomous driving benchmarks comprising more than 150,000 query–gallery pairs. Compared with conventional similarity-based matching and CLIP-based ReID baselines, the proposed approach reduces false positive rates by 21%–29% at fixed recall levels, while preserving stable rank-1 accuracy. These results indicate that uncertainty-aware modeling improves decision reliability in safety-critical perception pipelines.