🔥 RegenTrack: Distance-Adaptive Regeneration Pool Matching for Drone-Based Crowd Tracking

[Paper] | [Pretrained_weight]

Official implementation of "RegenTrack: Distance-Adaptive Regeneration Pool Matching for Drone-Based Crowd Tracking" (Accepted by IEEE TCSVT).

We propose RegenTrack, a tracking framework that integrates distance-adaptive fusion with Regeneration Pool matching to address challenges in drone-based crowd tracking，achieving better performance on DRONECROWD, DRONEBIRD, and CROHD dataset.

---

📢 News

• [2025.9.6] Upload code
• [2026.3.8] Upload pretrained model
• [2026.2.23] Accept by TCSVT!

🧠 Overview

🔹 Framework

🎨 Visualizations

Compared with Drone crowd tracking methods

Tracking results on DroneBird and HT21

⚙️ Environment

python >= 3.8
pytorch >= 2.0.1
faiss-cpu
salesforce-lavis
opencv-python
scipy

Install dependencies:

pip install -r requirements.txt

📂 Dataset

Download datasets:

• Download DroneCrowd dataset from GitHub, or Google-Drive
• Download DroneBird dataset from Roboflow
• Download HT21 dataset from MOTChallenge

Place data in the following directory:

#example for dronecrowd
/data/dronecrowd/
├── train_data/
│   ├── 00011/
│   │   └── origin/
│   │       ├── img011001.jpg
│   │       ├── img011002.jpg
│   │       └── ...
│   ├── 00012/
│   │   └── origin/
│   │       ├── img012001.jpg
│   │       └── ...
│   └── ...
├── test_data/
│   ├── 00001/
│   │   └── origin/
│   │       └── ...
│   └── ...
└── val_data/
├── 00001/
│   └── origin/
│       └── ...
└── ...

🧩 Model

Download pretrained model:

• Pretrained_weight -baidu 提取码为：6757

Place it in:

/RegenTrack/
├── detector/
│   ├── pretrained/
│   │   
├── tracker/
│   ├── pretrained/
│   └── ...

⚡ Quick Start

Clone repo:

git clone https://github.com/Zebrabeast/RegenTrack.git
cd detector

🔹 Test

python pet_dets_output.py  --root_dir dataset  --output_dir dec_result  --resume weight

#organize the data in term fromat like example for dronecrowd above
cd tracker
python regen_track.py --dataset_path dataset  --output_path track_result  --weight_path weight

🔹 Visualization

python visualization/draw_one_track.py 

📊 Evaluation

Run evaluation:

cd eval
python eval_dronecrowd.py 

Results on UAV-view dataset test set

Dataset	T-mAP	T-AP0.10	T-AP0.15	T-AP0.20
DroneCrowd	56.27	58.53	56.35	53.92
DroneBird	62.30	62.70	62.46	61.75

Comparison on DroneCrowd test set

Method	T-mAP	T-AP0.10	T-AP0.15	T-AP0.20
RegenTrack	56.27	58.53	56.35	53.92
SparseTrack	33.37	34.51	33.31	32.28

Notes:

• All inference experiments are performed on 2 NVIDIA GeForce RTX 4090 GPU.
• Both methods use the same detector as SparseTrack.
• RegenTrack: +23.0 T-mAP improvement over SparseTrack (56.27 vs. 33.37).

---

🏋️ Training

Train the model:

python train.py --dataset xxx --batch_size xx --lr xx

Tips

• Adjust learning rate for better convergence
• Use data augmentation to improve generalization
• Increase batch size if GPU memory allows

📖 Citation

If you find the code helpful in your research or work, please cite the following paper(s).

@article{lei2026regentrack,
    author={Lei, Yi and Zhou, Kang and Yuan, Jingling and Zhu, Huilin and Wang, Jinqiao and Zhong, Xian},
    journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
    title={RegenTrack: Distance-Adaptive Regeneration Pool Matching for Drone-Based Crowd Tracking}, 
    year={2026},
    volume={},
    number={},
    pages={1-1},
    doi={10.1109/TCSVT.2026.3671963}
}

🙏 Acknowledgement

This project is built upon the following works: PET, MPM, diffusion. We are grateful to the original authors for their outstanding work.

📬 Contact

For any questions:

• Email: week_fine@whut.edu.cn
• Github Issues