Improving Efficiency and Generalisability of Motion Prediction with Deep Multi-Agent Learning and Multi-Head Attention

This project is focused on developing a data-driven prediction framework for highly automated vehicles (AVs), leveraging multiple inputs to provide a multimodal, probabilistic estimate of the future intentions and trajectories of surrounding vehicles in freeway operations. Our proposed framework utilises deep multi-agent learning to effectively capture social interactions between vehicles, without relying on map information. Please refer to our recent publication for more details. This algorithm has demonstrated superior prediction performance, with reduced error in real-world traffic data from highways. Evaluation using the NGSIM (US-101 and I-80) and HighD datasets indicates strong prediction performance for long-term trajectory forecasting of multiple surrounding vehicles. Furthermore, the framework exhibits enhanced prediction accuracy and generalisability compared to state-of-the-art methods.

Djamel Eddine Benrachou

Centre for Accident Research & Road Safety - Queensland, QUT

---

News 🚀🚀

• [12/2023] Work on Transactions on Intelligent Transportation Systems (T-ITS, Q1-JCR): "Improving Efficiency and Generalisability of Motion Predictions With Deep Multi-Agent Learning and Multi-Head Attention"
• [09/2022] Work on Transactions on Intelligent Transportation Systems (T-ITS, Q1-JCR): "Use of Social Interaction and Intention to Improve Motion Prediction Within Automated Vehicle Framework: A Review"

---

This is the official repository and PyTorch implementations of different works presented in:

Our papers:

• Improving Efficiency and Generalisability of Motion Predictions with Multi-Agent Learning and Multi-Head Attention in IEEE Transactions on Intelligent Transportation Systems
• Use of Social Interaction and Intention to Improve Motion Prediction Within Automated Vehicle Framework: A Review in IEEE Transactions on Intelligent Transportation Systems

Setup:

The code can be executed in a Conda environment—please follow the tutorial below:

conda create -n MyEnvi python=3.7
source activate MyEnvi

conda install pytorch==1.12.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch
conda install tensorboard=1.14.0
conda install numpy=1.16 scipy=1.4 h5py=2.10 future

The code can also be written in the following environment:

python 3.7.11
pytorch 1.10.0
cuda 11.3 or above
cudnn 8.2.0

If available, check requirements.txt

conda create --name GATLSTM_env python=3.7 \
conda install -n GATLSTM_env ipykernel --update-deps --force-reinstall

python3 -m pip install --upgrade pip \
python3 -m pip install --upgrade Pillow \

pip install \
    prodict \
    torch \
    pyyaml \
    torchvision \
    tensorboard \
    glob2 \
    matplotlib \
    sklearn \
    gitpython \
    thop \
    fvcore \
    torchstat \
    torchsummary \
    ipykernel \
    sns

Datasets

NGSIM

From NGSIM website:

• Register at NGSIM
• Download [NGSIM-US-101-LosAngeles-CA]US-101-LosAngeles-CA.zip and [NGSIM-I-80-Emeryville-CA] I-80-Emeryville-CA.zip
• Unzip and extract vehicle-trajectory-data into raw/us-101 and ./raw/i-80

From googledrive:

• Download i-80: https://drive.google.com/open?id=19ovxiJLCnS1ar1dYvUIrfdz_UnGejK-k
• Download us-101: https://drive.google.com/open?id=14dMKew22_5evfOoSGEBYqHP3L92e7nyJ

HighD

• https://www.highd.com/

Pre-processed NGSIM and HighD can be downloaded from here

Data Preparation: The raw data of Next Generation Simulation (NGSIM) is downloadable at https://ops.fhwa.dot.gov/trafficanalysistools/ngsim.htm

1. Download the data: Once the data downloaded the raw data into ./raw
2. Preprocess the data : Run

preprocess_data.m to pre-preprocess the raw data (note: the pre-processed data will be uploaded on cloud).

This repository is for testing the trained models for motion prediction on highways. No end-to-end trainer is needed here. 3.To use the pretrained models at ./ .py and evaluate the models performance run! 🎉 🎉

python3 evaluate.py --name NAME_MODEL --batch_size BATCH_SIZE \
    --test_set ./datasets/HIGHD/test.mat 

4. Running (end-to-end training)

python3 train.py --name NAME_MODEL --batch_size BATCH_SIZE --pretrain_epochs 5 --train_epochs 10 \
    --train_set ./datasets/HIGHD/train.mat \
    --val_set ./datasets/HIGHD/val.mat

Citation

Please cite our papers if you used our code. Thanks.

@article{benrachou2022use,
  title={Use of social interaction and intention to improve motion prediction within automated vehicle framework: A review},
  author={Benrachou, Djamel Eddine and Glaser, Sebastien and Elhenawy, Mohammed and Rakotonirainy, Andry},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2022},
  publisher={IEEE}
}

@article{benrachou2023improving,
  title={Improving Efficiency and Generalisability of Motion Predictions With Deep Multi-Agent Learning and Multi-Head Attention},
  author={Benrachou, Djamel Eddine and Glaser, Sebastien and Elhenawy, Mohammed and Rakotonirainy, Andry},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2023},
  publisher={IEEE}
}

License

LICENSE