CRAT-Pred: Vehicle Trajectory Prediction with Crystal Graph Convolutional Neural Networks and Multi-Head Self-Attention

Official repository of the paper:
CRAT-Pred: Vehicle Trajectory Prediction with Crystal Graph Convolutional Neural Networks and Multi-Head Self-Attention
Julian Schmidt, Julian Jordan, Franz Gritschneder and Klaus Dietmayer
Accepted at 2022 IEEE International Conference on Robotics and Automation (ICRA)

Citation

If you use our source code, please cite:

@InProceedings{schmidt2022cratpred,
  author={Julian Schmidt and Julian Jordan and Franz Gritschneder and Klaus Dietmayer},
  booktitle={2022 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={CRAT-Pred: Vehicle Trajectory Prediction with Crystal Graph Convolutional Neural Networks and Multi-Head Self-Attention}, 
  year={2022},
  pages={7799--7805},}

License

CRAT-Pred is licensed under Creative Commons Attribution-NonCommercial 4.0 International License.

Check LICENSE for more information.

Installation

Install Anaconda

We recommend using Anaconda. The installation is described on the following page:
https://docs.anaconda.com/anaconda/install/linux/

Install Required Packages

conda env create -f environment.yml

Activate Environment

conda activate crat-pred

Install Argoverse API

pip install git+https://github.com/argoai/argoverse-api.git

Setup Argoverse Dataset

Download and Extract Dataset

bash fetch_dataset.sh

Preprocess the Dataset

Online and offline preprocessing is implemented. If you want to train your model offline on the preprocessed dataset, run:

python3 preprocess.py

You can also skip this step and run the preprocessing online during training.

Train Model

python3 train.py

or

python3 train.py --use_preprocessed=True

Checkpoints are saved in the lightning_logs/ folder. For accessing metrics and losses via Tensorboard, first start the server:

tensorboard --logdir lightning_logs/

Navigating to http://localhost:6006/ opens Tensorboard.

Test Model on Validation Set

python3 test.py --weight=/path/to/checkpoint.ckpt

Generate Predictions on Test Set

python3 test.py --weight=/path/to/checkpoint.ckpt --split=test