Graph Self-Attention Network for Learning Spatial-Temporal Interaction Representation in Autonomous Driving

GSAN

Introduction

Code for paper GSAN: Graph Self-Attention Network for Learning Spatial-Temporal Interaction Representation in Autonomous Driving, which was published on IEEE Internet of Things Journal. And the link is https://ieeexplore.ieee.org/document/9474961.

To reference the code, please cite this publication:

  @article{ye2021gsan,
    title={GSAN: Graph Self-Attention Network for Learning Spatial-Temporal Interaction Representation in Autonomous Driving},
    author={Ye, Luyao and Wang, Zezhong and Chen, Xinhong and Wang, Jianping and Wu, Kui and Lu, Kejie},
    journal={IEEE Internet of Things Journal},
    year={2021},
    publisher={IEEE}
  }

Datasets

• For lane-changing prediction task, we choose the open-source High-way Drone (HighD) Dataset.
• For trajectory prediction task, we choose NGSIM I-80 and US-101 Dataset.
• Datasets(NGSIM us-101, i-80 and HighD) are not included in the repo, please download by yourself from the official website.

Quick Start

1. Install/Update python dependency library

   pip install -r requirements.txt
   

2. Build the directory

   python buildfolder.py
   

Task1: Lane-changing classification

1. Get the data

   • HighD: https://www.highd-dataset.com/

2. Run all cells in highD_data_process.ipynb

Task2: Trajectory prediction

1. Get the data

   • NGSIM: https://data.transportation.gov/Automobiles/Next-Generation-Simulation-NGSIM-Vehicle-Trajector/8ect-6jqj

   • Follow this introduction to pre-process the data and get following files:

     • TestSet.mat
     • TrainSet.mat
     • ValSet.mat

   • Put these 3 files into data/ folder.

2. Format the data to fit GSAN model

   python datapreprocessing.py