News

20230910: All PSI data including videos, CV annotations, and Cognitive Annotation (PSI1.0 & PSI 2.0) are public for download and future exploration! ~~[Google Drive]~~ [PSI Homepage] 🎇⚡
20230607 - ‼️ We are hosting a competition about pedestrian behavior prediction, please check details and participate via [IEEE ITSS PSI Student Competition]
20230607 - The new [PSI 2.0 benchmark (Training & Validation & Test)] is released with more video samples and diverse multimodal interpretable annotations for pedestrian intent and driving decision!

❗ This repo is deprecated. Please check our new PSI Dataset and the [IEEE ITSS PSI Student Competition in Pedestrian Behavior Prediction]. ❗

IUPUI-CSRC Pedestrian Situated Intent (PSI) Dataset

This repository contains IUPUI-CSRC Pedestrian Situated Intent (PSI) Dataset pre-processing and baseline.

For more situated intent data and work, please see Situated Intent!

Download dataset and extract

Download the dataset from link, then extract via

unzip Dataset.zip

Output:

Archive:  Dataset.zip
creating: PSI_Intention/Dataset/ 
inflating: PSI_Intention/Dataset/VideoWithIndicator.zip  
inflating: PSI_Intention/Dataset/RawVideos.zip  
inflating: PSI_Intention/Dataset/README.txt  
inflating: PSI_Intention/Dataset/IntentAnnotations.xlsx
inflating: PSI_Intention/Dataset/XmlFiles.zip

Extract videos and spatial annotations:

unzip ./PSI_Intention/Dataset/RawVideos.zip -d ./PSI_Intention/Dataset
unzip ./PSI_Intention/Dataset/XmlFiles.zip -d ./PSI_Intention/Dataset

Video to frames

python split_clips_to_frames.py

The splited frames are organized as, e.g.,

frames{
    video_0001{
        000.jpg,
        001.jpg,
        ...
    }
}

CV_annotations and NLP_annotations re-organize

python reorganize_annotations.py

Note: video_0060 and video_0093 are removed due to the missing of spatial segmentation annotations.

Create database with frames labeled

python pedestrian_intention_database_processing.py

Output:

database_*.pkl: The annotaions of reasoning and intention do not exactly match, i.e., the last several frames only have intention annotations without reasoning, because the reasoning is only for the previous time period before the last annotated time point, while the intention annotation lasts till the end of the video.
database_*_overlap.pkl: By removing the last frames only with intention labels, the length of annotated reasoning and intention frames are Equal now.

Train/Val/Test split

train: [1 ~ 75]
val: [76 ~ 80]
test: [81 ~ 110]

Note: Due to the missing of spatial segmentation annotations, video_0060 and video_0093 are removed. Besides, video_0003 and video_0028 are ignored as the annotated frame sequences are too short.

In our PSI paper experiments, the observed tracks length is 15, while predicting the 16-th frame intention. The overlap rate is set as 0.8 for both train and test stages.

Citing

@article{chen2021psi,
title   = {PSI: A Pedestrian Behavior Dataset for Socially Intelligent Autonomous Car},
author  = {Chen, Tina and Tian, Renran and Chen, Yaobin and Domeyer, Joshua and Toyoda, Heishiro and Sherony, Rini and Jing, Taotao and Ding, Zhengming},
journal = {arXiv preprint arXiv:2112.02604},
year    = {2021} }

PSI-Intention2022/PSI-Intention