Basic Info

This is the unofficial implementation of the paper "Multi-Agent Trajectory Prediction with Heterogeneous Edge-Enhanced Graph Attention Network", Arxiv ID: 2106.07161.

The code is based on repo of the author. Since the Interaction Dataset has updated to v1.2 which is incompatible with the author's code, we re-code the dataset related part, and also made some simplification on the code.

Requirements

To install the required python libraries, use pip to install: pip install -r requirements.txt

Run Code

Preprocess

This code is based on INTERACTION dataset v1.2. The dataset folder arrangement is as follows:
.
├── maps
│  ├── {Scene_name_n}.osm_xy
│  ├── {Scene_name_n}.osm
│  ├── ...
│  └── ...
├── train
│  ├── {Scene_name_n}_train.csv
│  ├── ...
│  └── ...
├── val
│  ├── {Scene_name_n}_val.csv
│  ├── ...
│  └── ...
└── test
  ├── {Scene_name_n}_obs.csv
  ├── ...
  └── ...

To preprocess dataset, run: python data_preprocess.py --root /path/to/dataset --split {train/val/test}, the processed data would be store as '.pyg' and '.pt', folder structure are as follows:
.
├── maps
│  ├── {Scene_name_n}.osm_xy
│  ├── {Scene_name_n}.osm
│  ├── ...
│  └── ...
├── maps_png
│  ├── {Scene_name_n}_map.png
│  ├── ...
│  └── ...
├── train
│  ├── {Scene_name_n}_train.csv
│  ├── ...
│  └── ...
├── val
│  ├── {Scene_name_n}_val.csv
│  ├── ...
│  └── ...
├── test
│  ├── {Scene_name_n}obs.csv
│  ├── ...
│  └── ...
└── processed
  ├── train
  │  ├── {Scene_name_n}{case_id}.pyg
  │  ├── {Scene_name_n}map.pt
  │  ├── ...
  │  └── ...
  ├── val
  │  ├── {Scene_name_n}{case_id}.pyg
  │  ├── {Scene_name_n}map.pt
  │  ├── ...
  │  └── ...
  └── test
     ├── {Scene_name_n}{case_id}.pyg
     ├── {Scene_name_n}_map.pt
     ├── ...
     └── ...

Please notice that the '.png' map are provided by the original repo.

Run

Run python trainval.py --batch_size {bs} to train the model. The trained model will be saved to './models' as '.tar' file. Run python trainval.py --eval --metrics ALL --model {path/to/model} to evaluate the trained model with specific or all metrics.

Results

Since the original repo didn't give the result, we don't have a reference of the model. And now we don't get result as the paper show. Results after 60 epoches training:

ADE	FDE	ApFDE	LogCosh	AFDE	challenge_ADE
0.2291	0.7646	0.6114	0.0842	0.4969	0.2333

Hope someone would find the mistake of our code and issue us, and let's get the model work together!

Citation

If you have found this work to be useful, please consider citing the original paper:

@article{mo2022multi,
  title={Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network},
  author={Mo, Xiaoyu and Huang, Zhiyu and Xing, Yang and Lv, Chen},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2022},
  publisher={IEEE}
}

Thanks for code of the authors! https://github.com/Xiaoyu006/MATP-with-HEAT.git