MSTFormer: Motion Inspired Spatial-temporal Transformer with Dynamic-aware Attention for long-term Vessel Trajectory Prediction

This is a method for long-term ship trajectory prediction. It mainly improves the trajectory prediction performance by data augmentation, dynamic-aware attention, and knowledge-inspired loss function.

Requirements

geopy==2.2.0
ipdb==0.13.9
matplotlib==3.5.1
numpy==1.22.4
numpy_ext==0.9.8
pandas==1.4.3
scipy==1.8.1

Data

A simple dataset is provided for testing the code, or you can generate your own by rewriting data_loader.py.This project provides some commented out code that we use to process the data for reference.

Run

After configuring the environment, just run main_MSTFomer.py directly. Also, you can test different data by changing the parameters inside. The file where the logs are saved can be changed by changing the path in log.py.

Citation

If you find this repository useful in your research, please consider citing the following paper:

@misc{qiang2023mstformer,
      title={MSTFormer: Motion Inspired Spatial-temporal Transformer with Dynamic-aware Attention for long-term Vessel Trajectory Prediction}, 
      author={Huimin Qiang and Zhiyuan Guo and Shiyuan Xie and Xiaodong Peng},
      year={2023},
      eprint={2303.11540},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Contact

If you have any questions, feel free to contact Huimin Qiang through Email (qianghuimin21@mails.ucas.ac.cn) or Github issues. Pull requests are highly welcomed!

Acknowledgments

Thanks to NOAA for providing the raw data (ttps://coast.noaa.gov/htdata/CMSP/AISDataHandler/2021/), and thanks to everyone for their interest in this work!

simple316 / MSTFormer