OC_SORT

Why we included this project

Tracking a person or vehicle across crowded, occlusion-prone video is where OC-SORT fits. It rebuilds the classic SORT tracker around observation-centric momentum and a re-update step, which cuts down on identity switches when targets disappear briefly or move along non-linear paths. The tracker is purely motion-based and lightweight, so you can drop in detections from your own detector and later bolt on appearance similarity if motion alone isn't enough. It also ships practical deployment paths: ONNX, TensorRT, and ncnn exports, a C++ port, and a refactored Python implementation that reportedly runs 1.5 to 2 times faster. The reported results across MOT17, MOT20, DanceTrack, and KITTI give a concrete picture of where it helps and where it doesn't.