Abstract:

Modern methods solve multi-object tracking from two perspectives: motion modeling and appearance matching. As a classic paradigm, motion-based tracking by Kalman filters suffers from complicated motion patterns and the problem becomes more difficult when we only have noisy bounding boxes. To improve Kalman filter-based multi-object tracking in scenarios with complex motion, occlusion, and crossover, we propose Observation-Centric SORT with some simple modifications. The proposed modifications relieve the error in Kalman filter parameters after occlusion, thus better tracking objects in non-linear motion. Integrated with modern object detectors, such a simple tracking method can achieve state-of-the-art performance on multiple tracking datasets without requiring appearance information for object association. The proposed OC-SORT does not require any training and is purely motion-based. It runs in real-time on a single CPU.

Committee:

Prof. Kris Kitani (advisor)

Prof. Deva Ramanan

Prof. David Held

Zhengyi Luo (RI PhD Student)