Open Set Pedestrian Re Identification through Uncertainty Calibrated Vision Language Similarity

Single-frame pedestrian re-identification is often unstable in traffic scenes due to motion blur and partial visibility, whereas short video clips provide complementary cues. Building on uncertainty-aware CLIP-style modal representations, this work proposes an uncertainty-guided temporal aggregation module that assigns lower weights to unreliable frames and aligns clip-level visual features with text-derived semantic prompts. The model is evaluated on video-based autonomous driving benchmarks containing 52,000 tracklets and 1.1 million frames, with clip lengths ranging from 8 to 32 frames. Comparisons are conducted against video ReID baselines such as temporal pooling, recurrent aggregation, and transformer-based sequence models, as well as frame-level TransReID and CLIP-based methods. The proposed approach improves clip-level mAP by 3.6%–5.4% and rank-1 accuracy by 2.8%–4.1%, with additional gains in fast-motion segments where frame uncertainty is high.