Mohamed Jaouad Lahgazi and
Jean-Philippe Tarel
Abstract
Automated monitoring of Vulnerable Road Users (VRUs) using fixed infrastructure cameras is essential for traffic safety analytics in Intelligent Transport Systems. General person detection is well-studied, yet pre-trained models often fail to distinguish between the main VRU categories such as pedestrians, cyclists, motorcyclists, and personal mobility device (PMD) users. These categories require different safety interventions due to their distinct speed profiles, infrastructure requirements, and injury risks. We present a track-level classification pipeline that combines real-time person detection, multi-object tracking, and a Swin Transformer classifier fine-tuned on a custom four-class VRU dataset. A voting scheme, based on confidence gating, aggregates frame-level predictions into a single per-track label, making the system robust to temporary occlusion and detection noise. We also introduce an evaluation protocol in which the pipeline is run once on the evaluation video, tracks are labeled by a human annotator, and accuracy is measured by direct track identifier, eliminating the spatial re-matching artifacts that affect conventional pipeline evaluation. The pipeline is evaluated on 2,965 annotated tracks drawn from three urban surveillance recordings, achieving an overall accuracy of 99.70% across all four VRU classes. The pipeline, pretrained weights, and evaluation scripts are released publicly to support reproducible research.
Reference
@InProceedings{ jpt-itsc26,
author = {Lahgazi, M.-J. and Tarel, J.-P.},
title = {Fine-Grained Classification of Vulnerable Road Users: A Robust Track-Level Approach for Urban Surveillance},
booktitle = {IEEE Conference on Intelligent Transportation Systems (ITSC'26)},
date = {September 15-18},
address = {Naples, Italy},
pages = {-},
year = {2026},
url = {http://perso.lcpc.fr/tarel.jean-philippe/publis/itsc26.html}
}