Motion Tracking with Rotated Bounding Boxes on Overhead Fisheye Imagery
Jordan Lam (Zhejiang University)
Jordan Lam (Zhejiang University)
The 35th British Machine Vision Conference
Abstract
Although recent progress in multiple object tracking (MOT) has been notable, effectively tracking rotating bounding boxes and views from an overhead angle remains a considerable challenge. Previous methods typically ignore the rotation parameter such as using the centre point, or focus more on the appearance cues. This paper introduces a simple motion-based tracker that is effective for fisheye imagery, addressing these specific challenges. Our proposed method focuses on motion estimation and detection associations. Our approach is composed of (1) transforming rotated bounding box detections into 2D Gaussian distributions, (2) distribution distances that can replicate Intersection over Union (IoU) to associate detections, and (3) a dynamic buffer during association to alleviate irregular movement in overhead views. In our proposed method, we have experimented with three different distribution distances which have been shown to replicate the IoU behaviour during association. Through these distribution distances, we can effectively track rotated bounding boxes and be applied on a linear Kalman Filter. Experimental results show that our method achieves promising performance on multi-object tracking on overhead fisheye surveillance datasets and demonstrates comparable results on the MOT datasets.Citation
@inproceedings{Lam_2024_BMVC,
author = {Jordan Lam},
title = {Motion Tracking with Rotated Bounding Boxes on Overhead Fisheye Imagery},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year = {2024},
url = {https://papers.bmvc2024.org/0257.pdf}
}
Copyright © 2024
The British Machine Vision Conference is organised by