OC_SORT
OC-SORT is a simple, online, and real-time multi-object tracking software designed to maintain robustness in crowded scenes and during non-linear motion. Accepted at CVPR 2023, it addresses limitations in traditional Kalman filter approaches by focusing on observation-centric motion modeling. The system employs an Observation-Centric Re-Update mechanism to improve tracking consistency when objects are occluded or move unpredictably. It is highly flexible, supporting integration with various detectors and matching modules, including those utilizing appearance similarity. The software is available in both Python and C++ implementations, with the latter offering deployment optimizations. Performance metrics indicate high accuracy on major benchmarks such as MOT17, MOT20, KITTI, and DanceTrack. When detection boxes are provided, the association phase runs at 700 frames per second on a standard CPU, while full inference achieves approximately 28 frames per second on an RTX 2080Ti GPU. This tool is suitable for res