BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving

Official implementation of BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving Accepted at AAAI 2024

Overview

This repository contains the official implementation of BAT, a behavior-aware trajectory prediction framework for autonomous driving.

Paper: BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving Venue: AAAI 2024

Important Notice

[!WARNING] Due to the loss of an earlier internal code version, this repository is a carefully reconstructed implementation of the original project. The current version has been rebuilt as faithfully as possible based on our available codebase and experimental pipeline, with reconstruction grounded on STDAN. In our reproduced experiments, the overall model performance is slightly better than the results reported in the original paper.

We sincerely apologize for any inconvenience this may cause.

If you would like to quickly reproduce stronger results, please also refer to our latest work: HLTP.

👉 HLTP Project on GitHub 👈

Installation

Environment

The codebase has been tested on the following setup:

Ubuntu 20.04
CUDA 11.7
Python 3.8
PyTorch

1. Clone the repository

git clone https://github.com/Petrichor625/BATraj-Behavior-aware-Model.git
cd BATraj-Behavior-aware-Model

2. Install dependencies

If you already have a Python environment prepared, install the required packages with:

pip install -r requirements.txt

3. Create the Conda environment

We also provide a Conda environment file in the environments/ directory:

cd environments
conda env create -f environment.yml

If the above command does not work properly on your system, you may manually create the environment and install the main dependencies:

conda create -n behavior_aware python=3.8
conda activate behavior_aware
conda install pytorch==1.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

Configuration

Before training or evaluation, please update the relevant paths in the source code.

1. Checkpoint path

In config_new.py, set:

args['path']

to the directory where model checkpoints will be stored.

2. Dataset path

Please update the dataset path accordingly, for example:

lo.NgsimDataset('../NGSIM/TrainSet.mat')

Make sure the path matches the actual location of the dataset on your machine.

Training

If you want to resume training from pretrained checkpoints, specify the checkpoint paths in train5f_behavior_new.py:

generator.load_state_dict(t.load('../epoch8_g.tar'))
gdEncoder.load_state_dict(t.load('../epoch8_gd.tar'))

If you want to train the model from scratch, simply comment out these lines.

Evaluation

To evaluate a trained model, edit evaluate5f_behavior_new.py and specify the checkpoint epoch number:

if __name__ == '__main__':
    names = ['XX']  # e.g. ['8']
    evaluate = Evaluate()

    for epoch in names:
        evaluate.main(name=epoch, val=False)

Replace XX with the corresponding checkpoint epoch number you want to evaluate.

Reproducibility Notes

Since this repository is a reconstructed version rather than the exact original code used in the paper, minor implementation and performance differences may exist. Nevertheless, we have made every effort to preserve the original methodology and reproduce the reported behavior as closely as possible.

Citation

If you find this repository useful in your research, please cite:

@inproceedings{liao2024bat,
  title={Bat: Behavior-aware human-like trajectory prediction for autonomous driving},
  author={Liao, Haicheng and Li, Zhenning and Shen, Huanming and Zeng, Wenxuan and Liao, Dongping and Li, Guofa and Xu, Chengzhong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={9},
  pages={10332--10340},
  year={2024}
}