Trackers gives you clean, modular re-implementations of leading multi-object tracking algorithms released under the permissive Apache 2.0 license. You combine them with any detection model you already use.
BAM-SORT: A state-of-the-art multi-object tracking algorithm using bottom edge cost matrix and activated trajectory matching for robust tracking in crowded and occlusion-heavy environments. Excelling in benchmarks like MOT17, MOT20, and DanceTrack.
Person multi-object tracking using YOLOv8n and a custom SORT tracker, evaluated on MOT17-04.
Real-time pedestrian tracking using YOLOv8 for detection and DeepSORT for multi-object tracking on the MOT17 dataset.
Track video objects with YOLOv8 and StrongSORT, using appearance embeddings and MOT metrics for reliable multi-object tracking
Real-time multi-object tracking with YOLOv8 detection and BoT-SORT + CLIP ReID — evaluated on MOT17 (43.3% HOTA).
Advanced multi-object tracking pipeline — YOLOv8m + StrongSORT with OSNet appearance embeddings on MOT17. HOTA 41.6 | MOTA 38.1 | IDF1 50.8. Benchmarked against ByteTrack baseline with TrackEval.
Adaptive motion-feature fusion for robust multi-object tracking · 面向复杂交通场景的自适应运动特征融合多目标跟踪
Comparative analysis of DeepOCSORT, StrongSORT, and ByteTrack for robust tracking under heavy occlusion.
Person detection and tracking, benchmarked on MOT17 — HOTA 41.99, IDF1 50.32 with private detections. SSD MobileNet v3 and YOLOv8 behind one interface.
Real-time multi-object tracking that schedules a slow detector against a frame budget. ByteTrack-style tracker, YOLOX on ONNX Runtime, optical flow and ReID, with every trade-off measured on MOT17.
Object detection + multi-object tracking on CPU-only edge-class compute: C++ ONNX Runtime, Docker, reproducible benchmarks
An almost reproduction of my first published work back in 2013. Laptop-friendly multi-person human detection and tracking using YOLO, OpenCV, webcam streaming, and MOT17 video demos.
ByteTrack-based multi-object tracking with Hungarian matching and modified Kalman filter variants
Measurement code and data for 'Camera-Motion Compensation Is Not the Bottleneck: A Measurement Study of Shared Warps in Tracking-by-Detection'
A real-time multi-person tracker with a learned matcher and a memory of everyone it has seen. On MOT17, it beats SORT, ByteTrack, and DeepSORT and runs at 33 fps on a laptop RTX 4050. I trained every model on that same laptop.
Multi-object tracking by detection: YOLO + Kalman SORT baseline and ByteTrack, evaluated on MOT17 with HOTA, MOTA and IDF1.
Manual track annotation in CVAT vs MOT17 ground truth: IDF1, ID switches, fragmentation, amodal occlusion analysis, annotator guidelines
Reproduction of BoostTrack++ (arXiv 2408.13003). Achieved +1.2 HOTA, +0.7 MOTA, +1.6 IDF1, -24 IDSWs on MOT 17