EscapeTrack: Multi-object tracking with estimated camera parameters

Abstract

Multi-object tracking (MOT) remains a challenging task in dynamic environments. While most 2D tracking methods focus solely on the image plane, they often neglect the Ground Plane Assumption (GPA) — the principle that targets typically move on a consistent ground plane. This is because camera parameters are difficult to obtain and are not very reliable in scenarios involving camera motion or where the GPA does not apply. To address this issue, we propose EscapeTrack, a novel MOT algorithm that robustly handles imprecise camera parameters. Unlike conventional homography projection methods prone to calibration errors, EscapeTrack innovatively models target coordinates on the ground plane as latent variables within a Kalman filter framework. By constructing an observation model that projects these latent states onto the image plane, our method achieves superior tracking accuracy even with significant parameter noise. Extensive evaluations demonstrate state-of-the-art performance on MOT17, MOT20, DanceTrack, SportsMOT, and BDD100K benchmarks. Notably, EscapeTrack excels in scenarios with camera motion or GPA violations, by inherently treating such cases as camera parameter estimation errors. This robustness enables practical deployment in real-world systems where precise calibration is infeasible, advancing intelligent tracking in complex dynamic environments. The source code will be available at https://github.com/corfyi/EscapeTrack.