This is a code repo for Pedestrian Graph +: A Fast Pedestrian Crossing Prediction Model based on Graph Convolutional Networks

• Pedestrian Graph + Available via a colab notebook .

bilibili

Or on Youtube

If you use any of this code, please cite the following publications:

@ARTICLE{9774877,
  author={Cadena, Pablo Rodrigo Gantier and Qian, Yeqiang and Wang, Chunxiang and Yang, Ming},
  journal={IEEE Transactions on Intelligent Transportation Systems}, 
  title={Pedestrian Graph +: A Fast Pedestrian Crossing Prediction Model Based on Graph Convolutional Networks}, 
  year={2022},
  volume={},
  number={},
  pages={1-12},
  doi={10.1109/TITS.2022.3173537}}

@inproceedings{cadena2019pedestrian,
  title={Pedestrian graph: Pedestrian crossing prediction based on 2d pose estimation and graph convolutional networks},
  author={Cadena, Pablo Rodrigo Gantier and Yang, Ming and Qian, Yeqiang and Wang, Chunxiang},
  booktitle={2019 IEEE Intelligent Transportation Systems Conference (ITSC)},
  pages={2000--2005},
  year={2019},
  organization={IEEE}
}

install:
pytorch 1.8.0 or above
pytorch lightning 1.5.10 or above

Also you can use docker

sudo docker pull nvcr.io/nvidia/l4t-base:r32.4.3

sudo run sh ./run_docker.sh

our code was tested on the jetson nano 4Gb

• Download the linked material below Sample dataset for training and testing:
  

pre-processed data baidu (data).
pre-processed data google dive (data).

PIE data baidu (PIE).
PIE data google dive (PIE).

JAAD data baidu (JAAD).
JAAD data google dive (JAAD).

All three folders must be inside the Pedestrian_graph_plus folder

test JAAD

python3 final_jaad_test.py --ckpt ./weigths/jaad-23-IVSFT/best.pth

The following inference is made on a 4Gb jetson nano:

This inference shows that Pedestrian Graph + is able to run on low-resource hardware, being efficient while maintaining high accuracy.

Inference time on jetson nano is 24ms, on the GTX 1060 (laptop) the inference time is 3 ms.

test JAAD with 2D human keypoints

python3 final_jaad_test.py --ckpt ./weigths/jaad-23-IVSFT-h2d/best.pth

test PIE

python3 final_pie_test.py --ckpt ./weigths/jaad-23-IVSFT-h2d/best.pth

To train

python3 pl_jaad_muster23_forecast.py --logdir ./weigths/jaad-23-IVSFT/

MIT license