BiTraP: Bi-directional Pedestrian Trajectory Prediction with Multi-modal Goal Estimation

Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan and Xiaoxiao Du

This repo contains the code for our paper:BiTraP: Bi-directional Pedestrian Trajectory Prediction with Multi-modal Goal Estimation.

Our BiTraP-NP network architecture:

Our BiTraP-GMM decoder architecture:

Installation

Dependencies

Our code was implemented using python and pytorch and tested on a desktop computer with Intel Xeon 2.10GHz CPU, NVIDIA TITAN X GPU and 128 GB memory.

NVIDIA driver >= 418
Python >= 3.6
pytorch == 1.4.1 with GPU support (CUDA 10.1 & cuDNN 7)

Run following command to add bitrap path to the PYTHONPATH

cd bidireaction-trajectory-prediction export PYTHONPATH=$PWD:PYTHONPATH

One can also use docker with docker/Dockerfile.

Training

Users can train the BiTraP models on JAAD, PIE or ETH-UCY dataset easily by runing the following command:

python tools/train.py --config_file **DIR_TO_THE_YML_FILE**

To train on JAAD or PIE dataset, the users need to download the dataset from their original webpage and follow their directions to extract pedestrian trajectory files ('.pkl'). The trajectories should locate at the DATASET.TRAJECTORY_PATH in the configuration file.

To train on ETH-UCY dataset, the users can download the trajectory files ('.pkl') from the Trajectron++ repo and put them at the DATASET.TRAJECTORY_PATH in the configuration file.

To train/inferece on CPU or GPU, simply add DEVICE='cpu' or DEVICE='cuda'. By default we use GPU for both training and inferencing.

Inference

The checkpoints of our models trained on JAAD, PIE and ETH-UCY can be downloaded here.

Bounding box trajectory prediction on JAAD and PIE

We predict the bounding box coordinate trajectory for first-person (ego-centric) view JAAD and PIE datasets. Test on PIE dataset:

python tools/test.py --config_file configs/bitrap_np_PIE.yml CKPT_DIR **DIR_TO_CKPT**
python tools/test.py --config_file configs/bitrap_gmm_PIE.yml CKPT_DIR **DIR_TO_CKPT**

Test on JAAD dataset:

python tools/test.py --config_file configs/bitrap_np_JAAD.yml CKPT_DIR **DIR_TO_CKPT**
python tools/test.py --config_file configs/bitrap_gmm_JAAD.yml CKPT_DIR **DIR_TO_CKPT**

Point trajectory prediction on ETH-UCY

We predict the point coordinate trajectory for bird's-eye view ETH-UCY datasets.

python tools/test.py --config_file configs/bitrap_np_ETH.yml DATASET.NAME **NAME_OF_DATASET** CKPT_DIR **DIR_TO_CKPT**
python tools/test.py --config_file configs/bitrap_gmm_ETH.yml DATASET.NAME **NAME_OF_DATASET** CKPT_DIR **DIR_TO_CKPT**

Citation

If you found the repo is useful, please feel free to cite our papers:

@article{yao2020bitrap,
  title={BiTraP: Bi-directional Pedestrian Trajectory Prediction with Multi-modal Goal Estimation},
  author={Yao, Yu and Atkins, Ella and Johnson-Roberson, Matthew and Vasudevan, Ram and Du, Xiaoxiao},
  journal={arXiv preprint arXiv:2007.14558},
  year={2020}
}

umautobots / bidireaction-trajectory-prediction