Distortion-Aware Multi-Object Tracking via Virtual Plane Projection in Overhead Fisheye Cameras


Panithi Vanasirikul (OxygenAI), Piyanon Charoenpoonpanich (OxygenAI), Ekapol Chuangsuwanich (Chulalongkorn University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Despite progress in multi-object tracking (MOT) for perspective cameras, applying these methods to overhead fisheye cameras remains challenging. We argue that existing MOT approaches often fail due to two key challenges: the linear motion assumption of Kalman filters conflicting with fisheye distortions, and highly variable appearances caused by extremely varying viewpoints and non-uniform geometric distortion. We propose DAFTrack, a motion-based tracker designed for overhead fisheye cameras, with three main contributions. First, we track in a distortion-aware space by projecting detections as torso locations onto a virtual plane. Second, we use the Unscented Kalman Filter (UKF) for nonlinear updates, enabling accurate state estimation. Third, we leverage the camera height to enhance distance-based matching, improving robustness across different camera setups. Experimental results show that DAFTrack outperforms baselines in overhead fisheye MOT datasets, achieving a +6.1 HOTA improvement on CEPDOF and +3.3 HOTA on WEPDTOF.

Citation

@inproceedings{Vanasirikul_2025_BMVC,
author    = {Panithi Vanasirikul and Piyanon Charoenpoonpanich and Ekapol Chuangsuwanich},
title     = {Distortion-Aware Multi-Object Tracking via Virtual Plane Projection in Overhead Fisheye Cameras},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_880/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection