Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking

Jinkun Cao, Xinshuo Weng, Rawal Khirodkar, Jiangmiao Pang, Kris M. Kitani

Keywords: multi-object tracking

Abstract: Recent advances in object detection and re-identification have greatly improved the performance of Multi-Object Tracking (MOT) methods, but progress in motion modeling has been limited. The motion model is a key component of many MOT methods and is commonly used to predict an object's future position. However, mainstream motion models in MOT naively assume that object motion is linear. They rely on detections on each frame as the observation value to supervise motion models. However, in practice, the observations can be noisy and even missing, especially in crowded scenes, which greatly degrade the performance of existing MOT methods. In this work, we show that a simple filtering-based motion model can still obtain state-of-the-art tracking performance if proper care is given to missing observations and noisy estimates. We emphasize the role of observations when recovering tracks from being lost and reducing the error accumulated by the assumption of linear motion when the target is lost. In contrast to the popular motion-based method SORT, which is estimation-centric, we name our method Observation-Centric SORT (OC-SORT). It remains simple, online, and real-time but improves robustness over occlusion and non-linear motion. It achieves state-of-the-art on multiple MOT benchmarks, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2203.14360/code)