TRANS: Benchmark for Pedestrian Stop & Go Forecasting

This repository contains the code of TRANS, the benchmark introduced in our paper for explicitly studying the stop and go behaviors of pedestrians in urban traffic. TRANS is built on top of several existing autonomous driving datasets annotated with walking behaviors of pedestrians (see Table I), so that it includes transition samples from diverse traffic scenarios with a unified interface

Table of Contents

• Installation
• Data preparation
• Interface
• Statistics
• Further work
• Citation

Installation

# To clone the repository using HTTPS
git clone https://github.com/vita-epfl/pedestrian-transition-dataset.git

The project is written and tested using python 3.8. The interface also require several external libraries. All required packages can be found in requirements.txt and to install dependencies run:

pip install -r requirements.txt

Please ensure all expected modules are properly installed before using the interface.

Data preparation

Supporting datasets

Currently we augment three existing self-driving datasets datasets:

• Joint Attention in Autonomous Driving(JAAD) Dataset
• Pedestrian Intention Estimation(PIE) Dataset
• Trajectory Inference using Targeted Action priors Network(TITAN) Dataset

Download

JAAD

• Download the videos and annotations from official page.
• Use the scripts provided to extract images from videos.
• Use JAAD's interface to generate complete annotations in the form of python dictionary. More precisely use generate_database() to obtain the all JAAD annotations in dictionary form as a pkl file. Please rename the .pkl file as JAAD_DATA.pkl and place it in DATA/annotations/JAAD/anns
• Train / val / test split: link.

PIE

• Download the videos and annotations from offical page.
• Use the scripts provided to extract images from videos.
• Use PIE's interface to generate complete annotations in the form of python dictionary. In detail, please use generate_database() to obtain the all JAAD annotations in dictionary form as a pkl file. Rename the .pkl file as PIE_DATA.pkl and place it in DATA/annotations/PIE/anns
• Train / val / test split: link

TITAN

To obtain the dataset, please refer to this page and contact them directly.

After download the data, please place the images, annotations and video split ids in DATA. The expected result:

├── DATA/
│   ├── annotations/ 
│   │   ├── JAAD/ 
│   │   │   ├── anns/  
│   │   │   │   ├── JAAD_DATA.pkl # see Data-preparation-download-JAAD
│   │   │   └── splits/
│   │   │   │   ├── all_videos/
│   │   │   │   │   ├── train.txt
│   │   │   │   │   └── val.txt
│   │   │   │   │   └── test.txt
│   │   │   │   └── default/
│   │   │   │   │   ├── train.txt
│   │   │   │   │   └── val.txt
│   │   │   │   │   └── test.txt
│   │   │   │   └── high_visibility/
│   │   │   │   │   ├── train.txt
│   │   │   │   │   └── val.txt
│   │   │   │   │   └── test.txt
│   │   └── PIE/
│   │   │   ├── anns/ 
│   │   │   │   ├── PIE_DATA.pkl # see Data-preparation-download-PIE
│   │   └── TITAN/
│   │   │   ├── anns/
│   │   │   │   ├── clip_x.csv
│   │   │   └── splits/
│   │   │   │   │   ├── train_set.txt
│   │   │   │   │   └── val_set.txt
│   │   │   │   │   └── test_set.txt
│   └── images/
│   │   ├── JAAD/
│   │   │   ├── video_xxxx/ # 346 videos
│   │   └── PIE/
│   │   │   ├── set_0x/ # 6 sets
│   │   │   │   ├── video_xxxx/ 
│   │   └── TITAN/
│   │   │   ├── clip_xxx/  # 786 clips
└── (+ files and folders containing the raw data)

Note : The benchmark works with arbitrary subsets of supported datasets when we give only the paths to desired ones.

Interface

At the heart of TRANS is the TransDataset class. TransDataset generated transition samples from original annotations of JAAD, PIE and TITAN. Using attributes of TransDataset, the user can conveniently extract the frame sequences related to the stop & go of pedestrians:
The extracted samples each has an unique id specifying the source dataset, transition type, data split and sample index. We provide customized dataloader for loading samples according to the user's need. For basic usage please check the example in example.ipynb .

Statistics

All unique pedestrians in the original datasets can be categorized into Walk, Stand, Stop and Go. Overall we have 1,008 go events annd 1,138 stop events in TRANS. For more detailed analysis, please refer to the paper. You can explore the statistics of different datastes using the verbose feature.

Further work

TRANS serves as the primary data source for the project "Pedestrian Stop and Go Forecasting" in VITA. The interface is still under development. In future we may expect more relevant datasets to be integrated in this benchmark. Please leave remarks if you encounter any problems using the interface or have suggestions for improving the usability.

Citation

If you use this project in your research, please cite the corresponding paper:

@article{guo2022pedestrian,
  title={Pedestrian Stop and Go Forecasting with Hybrid Feature Fusion},
  author={Guo, Dongxu and Mordan, Taylor and Alahi, Alexandre},
  journal={arXiv preprint arXiv:2203.02489},
  year={2022}
}