Pedestrian Re-Identification as a Joint Perception–Prediction Problem under Vision–Language Uncertainty
Computer Science Bulletin · 30.01.2026 · 9 (1) · 238-244
DOI: 10.71465/csb217 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer Science Bulletin, online erschienen am 30.01.2026. Verfasst von Alexey V. Ivanov und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Pedestrian re-identification and trajectory prediction are closely connected tasks in autonomous driving perception pipelines. Based on CLIP-derived uncertainty-aware representations, this paper proposes a joint learning framework in which vision–language features are shared across identity recognition and motion prediction tasks. The model is evaluated on a multi-task dataset containing more than 80,000 pedestrian trajectories and 1.2 million annotated frames. Re identification performance is measured using rank-1 accuracy and mAP, while trajectory prediction is evaluated using ADE and FDE metrics. Compared with single-task and deterministic multi-task baselines, the proposed approach improves re-identification mAP by 3.8% and reduces trajectory prediction error by 5.9%–7.1%.
Zitieren
APA
Ivanov, A. V., Petrova, N. S., & Kuznetsov, D. A. (2026). Pedestrian Re-Identification as a Joint Perception–Prediction Problem under Vision–Language Uncertainty. Computer Science Bulletin, 9(1), 238-244. https://doi.org/10.71465/csb217
Vancouver
Ivanov AV, Petrova NS, Kuznetsov DA. Pedestrian Re-Identification as a Joint Perception–Prediction Problem under Vision–Language Uncertainty. Computer Science Bulletin. 2026;9(1):238-244. doi:10.71465/csb217
BibTeX
@article{Ivanov2026,
author = {Ivanov, Alexey V. and Petrova, Natalia S. and Kuznetsov, Dmitry A.},
title = {Pedestrian Re-Identification as a Joint Perception–Prediction Problem under Vision–Language Uncertainty},
journal = {Computer Science Bulletin},
year = {2026},
volume = {9},
number = {1},
pages = {238--244},
doi = {10.71465/csb217}
}
EUVIMED