RM3DMOT: Roadside Monocular 3D Multi-Object Tracking With Motion-Appearance Optimization
IEEE Open Journal of Intelligent Transportation Systems · 2026 · 7 · 1674-1686
DOI: 10.1109/ojits.2026.3706528 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Open Journal of Intelligent Transportation Systems, 2026. Verfasst von Ping Wang und 5 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Intergovernmental International Science and Technology Innovation Cooperation of National Key Research and Development Program of China.
Abstract (Originalsprache)
To enhance autonomous driving, vehicle-infrastructure cooperative systems are proposed to extend the vehicles’ perception range and rectify erroneous detection results via roadside perception. As a critical component of the system, multi-object tracking (MOT) provides sequential information of traffic participants, modeling their future motion tendencies. Although 3D MOT has been extensively studied for onboard sensors, roadside monocular 3D MOT remains underexplored. To address this, we propose RM3DMOT, a novel Roadside Monocular 3D Multi-Object Tracking algorithmthat systematically integrates motion and appearance optimization tailored to roadside challenges. In the motion prediction module, the unscented Kalman filter (UKF) is enhanced by incorporating nonlinear motion characteristics of multi-category participants. For the appearance feature extractor, a camera style adaptation mechanism—utilizing generative adversarial networks—is designed to ensure feature precision. Specifically, we customize UKF with category-specific motion models to handle nonlinear trajectories, and we introduce a GAN-based style adaptation to improve appearance feature robustness across diverse cameras. To enable effective motion-appearance association, we design a learning-based association modulethat leverages Multilayer Perceptron (MLP) to adaptively fuse motion and appearance similarities for robust trajectory-detection matching. While learning-based association is an established concept, its integration into a roadside monocular 3D MOT framework—with tailored inputs (HAIoU, appearance cosine, class compatibility) and class-balanced focal loss—constitutes a practical contribution to addressing the unique challenges of roadside scenes. A series of validation experiments are conducted on the V2X-Seq dataset and RCooper dataset. On the V2X-Seq dataset, RM3DMOT achieves state-of-the-art performance ( ${72.06}\%$ MOTA, ${83.95}\%$ IDF1 for Vehicle, and ${59.57}\%$ MOTA, ${78.97}\%$ IDF1 for Non-vehicle). On the RCooper dataset, RM3DMOT demonstrates superior performance in terms of accuracy and robustness.
Förderung
- Intergovernmental International Science and Technology Innovation Cooperation of National Key Research and Development Program of China
Zitieren
APA
Wang, P., Yao, Y., Li, Z., Wang, C., Tian, C., & Wang, X. (2026). RM3DMOT: Roadside Monocular 3D Multi-Object Tracking With Motion-Appearance Optimization. IEEE Open Journal of Intelligent Transportation Systems, 7, 1674-1686. https://doi.org/10.1109/ojits.2026.3706528
Vancouver
Wang P, Yao Y, Li Z, Wang C, Tian C, Wang X. RM3DMOT: Roadside Monocular 3D Multi-Object Tracking With Motion-Appearance Optimization. IEEE Open Journal of Intelligent Transportation Systems. 2026;7:1674-1686. doi:10.1109/ojits.2026.3706528
BibTeX
@article{Wang2026,
author = {Wang, Ping and Yao, Yuyang and Li, Zuxing and Wang, Chao and Tian, Can and Wang, Xinhong},
title = {RM3DMOT: Roadside Monocular 3D Multi-Object Tracking With Motion-Appearance Optimization},
journal = {IEEE Open Journal of Intelligent Transportation Systems},
year = {2026},
volume = {7},
pages = {1674--1686},
doi = {10.1109/ojits.2026.3706528}
}
EUVIMED