One Homography is All You Need: IMM-based Joint Homography and Multiple Object State Estimation

The introduction of the IMM Joint Homography State Estimation (IMM-JHSE) algorithm represents a significant advancement in multiple object tracking (MOT). By relying solely on an initial homography estimate for 3D information, IMM-JHSE circumvents the limitations of traditional camera motion compensation techniques that often distort predicted track positions. This innovative method combines static and dynamic camera motion models through an IMM filter, enhancing robustness against motion deviations. Performance evaluations on datasets such as DanceTrack, MOT17, and MOT20 reveal that IMM-JHSE outperforms nearly all existing 2D MOT methods, achieving notable increases in HOTA scores. The algorithm's ability to integrate bounding box motion models with ground-plane-based metrics further solidifies its competitive edge, making it a promising tool for future applications in computer vision.