psi-benchmark/PSI-Benchmark

Pedestrian Situated Intent (PSI) Bencharmark

This Repository contains the scripts and instructions about preparing the Pedestrian Situated Intent (PSI) 2.0 dataset.

Dataset Summary

HomePage (http://pedestriandataset.situated-intent.net/)

Prediction of pedestrian behavior is critical for fully autonomous vehicles to drive in busy city streets safely and efficiently. The future autonomous cars need to fit into mixed conditions with not only technical but also social capabilities. It is important to estimate the temporal-dynamic intent changes of the pedestrians, provide explanations of the interaction scenes, and support algorithms with social intelligence.

The IUPUI-CSRC Pedestrian Situated Intent (PSI-1.0) benchmark dataset has two innovative labels besides comprehensive computer vision annotations. The first novel label is the dynamic intent changes for the pedestrians to cross in front of the ego-vehicle, achieved from 24 drivers with diverse backgrounds. The second one is the text-based explanations of the driver reasoning process when estimating pedestrian intents and predicting their behaviors during the interaction period.

PSI-2.0 covers 196 scenes including 110 scenes from PSI-1.0 and contains the labels from 74 new subjects. Additionally PSI-2.0 includes bounding box annotations for traffic objects and agents which can be linked with text descriptions and reasoning explanations for building vision-language models.

These innovative labels can enable computer vision tasks like pedestrian intent/behavior prediction, vehicle-pedestrian interaction segmentation, and video-to-language mapping for explainable algorithms. The dataset also contains driving dynamics and driving decision-making reasoning explanations.

Supported Tasks

• Pedestrian crossing intent prediction and reasoning generation
• Pedestrian trajectory prediction and reasoning generation
• Driving decision prediction and reasoning generation
• Drivier-pedestrian interaction prediction and cognitive analysis
• Uncertainty of estimation in crowd-sourcing in the autonomous driving domain
• Visual-language matching and causal reasoning
• General classification, object detection, segmentation, and more tasks.

Dataset Preparation

Download the PSI 2.0 Dataset videos and annotations from Hugging Face or [PSI Homepage]. Move *.zip files to the dataset ROOT_PATH, and unzip them by

    cd ROOT_PATH # e.g., root/Dataset
    unzip '*.zip' -d .
    rm *.zip

The extracted folder contains all videos (Train/Val):

• ROOT_PATH/PSI_Videos/videos.

The extracted folder contains all annotations of the PSI 2.0 Dataset (Train/Val)

• ROOT_PATH/PSI2.0_TrainVal/annotations/cognitive_annotation_key_frame
• ROOT_PATH/PSI2.0_TrainVal/annotations/cv_annotation

and the train/val/test splits:

• ROOT_PATH/PSI2.0_TrainVal/splits/PSI2_split.json.

Bias, Privacy, Social Impact, and Limitations Considerations

• Bias and Representativeness: Our group of 74 annotators includes 44 males and 30 females, ranging in age from 19 to 77. The age distribution is balanced across three groups — 19 to 30, 31 to 54, and 55 and older. All annotators hold valid U.S. driver’s licenses. Their driving experience varies, with approximately 25 percent driving around 5,000 miles per year, 40 percent driving about 10,000 miles per year, and the remainder exceeding 15,000 miles annually. This diversity provides a broad range of perspectives for pedestrian intent estimation and reasoning. However, we acknowledge that the dataset is U.S.-centric, as the driving scenes were collected from multiple cities across the United States, and may not generalize to global contexts.
• Privacy and Consent: The dataset does not contain any personally identifiable information. All data was collected from public environments and anonymized where necessary. Users are responsible for ensuring compliance with local data protection regulations when using or sharing the dataset. All video and annotation data were reviewed and approved by the Indiana University legal department under an IRB-approved protocol. Users must agree to use the dataset solely for research purposes. While the current release does not include face or license plate blurring, we acknowledge this limitation and plan to implement de-identification and anonymization protocols in future updates.
• Social Impact and Ethical Considerations: The dataset was collected with careful consideration of data ownership and consent. However, users should remain mindful of potential ethical concerns when deploying models trained on this data, particularly in sensitive domains such as surveillance, hiring, or healthcare. We recognize the risk of overfitting models to subjective human reasoning. To help mitigate this, PSI includes annotations from a diverse group of annotators and preserves disagreements across labels. This approach supports the development of models that can reason under uncertainty rather than relying solely on consensus.
• Limitations: This dataset provides valuable insights into driver-pedestrian interactions and reasoning for autonomous driving research. However, users should be aware of several limitations. (1) The dataset is smaller than some large-scale benchmarks, which may affect generalization in data-intensive models. (2) Geographic scope — all videos were collected in U.S. cities, which may limit applicability to regions with different traffic rules, infrastructure, or pedestrian behavior. (3) Demographic representation — the dataset may not fully capture the diversity of pedestrian demographics, such as age, mobility, or clothing styles, which could introduce bias in behavior prediction or reasoning. (4) Annotation subjectivity — reasoning or intention annotations may reflect annotator bias or interpretation variability, especially in ambiguous scenarios. (5) Temporal relevance — the videos may not reflect recent advancements in camera technology, changes in traffic behavior, vehicle systems, or pedestrian norms, which could affect the dataset’s relevance for current autonomous driving systems.

Researchers and developers are encouraged to consider these limitations when using the dataset and to complement it with additional data sources where possible.

Citation

@article{chen2021psi,
title={Psi: A pedestrian behavior dataset for socially intelligent autonomous car},
author={Chen, Tina and Jing, Taotao and Tian, Renran and Chen, Yaobin and Domeyer, Joshua and Toyoda, Heishiro and
Sherony, Rini and Ding, Zhengming},
journal={arXiv preprint arXiv:2112.02604},
year={2021} }

@inproceedings{jing2022inaction,
title={Inaction: Interpretable action decision making for autonomous driving},
author={Jing, Taotao and Xia, Haifeng and Tian, Renran and Ding, Haoran and Luo, Xiao and Domeyer, Joshua and Sherony,
Rini and Ding, Zhengming},
booktitle={European Conference on Computer Vision},
pages={370--387},
year={2022},
organization={Springer} }