FastTrackTr

FastTrackTr: Fast Multi-Object Tracking with Transformer Trajectory Modeling

📄 IEEE Paper | 📑 arXiv | 🎬 Demo Videos

---

News

• [2026/06] Code released with bug fixes.
• [2026] Paper published on IEEE Xplore.

Overview

FastTrackTr is a fast and efficient multi-object tracking framework built on transformer architectures. It leverages trajectory modeling with RT-DETR as the detection backbone for real-time multi-object tracking on benchmarks such as DanceTrack, SportsMOT, and MOT17.

Installation

Requirements

• Python >= 3.10
• PyTorch >= 2.0
• CUDA >= 11.8

Setup

# Clone the repository
git clone https://github.com/liaopan-lp/fasttracktr.git
cd fasttracktr

# Install dependencies
pip install torch torchvision
pip install einops scipy pyyaml tqdm wandb

Data Preparation

Prepare the datasets (DanceTrack, SportsMOT, MOT17, etc.) and organize them under a datasets/ directory (or specify the path via --data-root):

datasets/
├── DanceTrack/
│   ├── train/
│   ├── val/
│   └── test/
├── SportsMOT/
│   ├── train/
│   ├── val/
│   └── test/
└── MOT17/
    ├── train/
    └── test/

Training

Train on DanceTrack

python main.py --mode train \
    --config-path ./configs/r50_rt_detr_rtmot_dancetrack.yaml \
    --data-root ./datasets/ \
    --device cuda

Resume Training

python main.py --mode train \
    --config-path ./configs/r50_rt_detr_rtmot_dancetrack.yaml \
    --resume-model ./output/checkpoint_13.pth \
    --resume-optimizer True \
    --resume-scheduler True \
    --resume-states True

Evaluation

python main.py --mode eval \
    --config-path ./configs/r50_rt_detr_rtmot_dancetrack.yaml \
    --inference-model ./output/checkpoint_13.pth \
    --inference-dataset DanceTrack \
    --inference-split val

Submission / Inference

python main.py --mode submit \
    --config-path ./configs/r50_rt_detr_rtmot_dancetrack.yaml \
    --inference-model ./output/checkpoint_13.pth \
    --inference-dataset DanceTrack \
    --inference-split test

Project Structure

fasttracktr/
├── main.py                 # Entry point
├── train_engine.py         # Training logic
├── eval_engine.py          # Evaluation logic
├── submit_engine.py        # Submission / inference logic
├── configs/                # YAML configuration files
│   ├── r50_rt_detr_rtmot_dancetrack.yaml
│   ├── r18_fasttracktr_dancetrack.yaml
│   └── ...
├── models/                 # Model implementations
│   ├── fasttracktr.py      # FastTrackTr main model
│   ├── criterion.py        # Loss functions
│   ├── rt_detr_cross/      # RT-DETR with cross-attention
│   └── ...
├── data/                   # Dataset and data loading
├── structures/             # Data structures (Instances, Args, etc.)
├── utils/                  # Utility functions
├── log/                    # Logging utilities
└── TrackEval/              # Evaluation toolkit (HOTA, CLEAR, Identity)

Citation

If you find this work useful, please cite our paper:

@ARTICLE{fasttracktr2026,
  author={Liao, Pan and Yang, Feng and Wu, Di and Yu, Jinwen and Li, Xingxin and Zhang, Dingwen},
  journal={IEEE Transactions on Industrial Informatics}, 
  title={FastTrackTr: Real-Time Multiobject Tracking With Transformers for Real World}, 
  year={2026},
  volume={22},
  number={3},
  pages={1817-1827},
  keywords={Transformers;Decoding;Real-time systems;Accuracy;Target tracking;Computer architecture;Computational modeling;Feature extraction;Training;Object recognition;Multiobject tracking (MOT);real-time;transformers},
  doi={10.1109/TII.2025.3631698}}

License

This project is released under the Apache 2.0 License.

Acknowledgements

• RT-DETR
• Deformable DETR
• TrackEval