HybridTrack is a novel 3D multi-object tracking (MOT) framework that combines the strengths of traditional Kalman filtering with the adaptability of deep learning. Designed for traffic and autonomous driving scenarios, it delivers state-of-the-art accuracy and real-time performance—no manual tuning or scenario-specific designs required.
- HybridTrack has been accepted to RAL 2025!
📌 A schematic of the HybridTrack architecture will be added here.
- 3D Object Tracking using LiDAR
- Learnable Kalman Filter (LKF)
- Real-time performance (112 FPS)
- High tracking accuracy without handcrafted noise or motion models
- Generalizes across different driving scenarios
| Method | HOTA | FPS | Modality | Model Weights |
|---|---|---|---|---|
| HybridTrack (Ours) | 82.08% | 112 | 3D (LiDAR) | Download (.pth) |
See the paper for detailed comparison across metrics like MOTA, IDF1, and association accuracy.
HybridTrack is evaluated on the KITTI Tracking Benchmark.
- Prepare your data: Follow the Data Preparation Guide for step-by-step instructions on downloading, organizing, and linking the KITTI dataset, detections, and annotations.
- Install requirements:
pip install -r requirements.txt
- Configure and run:
- For training, see Training Guide
- For tracking, see Tracking Guide
This repository builds upon the following open-source projects:
We thank the authors for their valuable contributions to the community.
If you use HybridTrack in your research, please consider citing:
@article{di2025hybridtrack,
title={HybridTrack: A Hybrid Approach for Robust Multi-Object Tracking},
author={Di Bella, Leandro and Lyu, Yangxintong and Cornelis, Bruno and Munteanu, Adrian},
journal={arXiv preprint arXiv:2501.01275},
year={2025}
}
