PeDesCar: A Large-Scale Dataset for Simulated Car-Pedestrian Collisions to Advance Safety Research
Year: 2025
Type: conference-paper
Abstract: Car-pedestrian collisions are a daily occurrence worldwide, yet there is a notable absence of public datasets in this domain.Research in this area is crucial, as it directly impacts pedestrian safety and serves as a basis for validating autonomous driving systems.Although finite element simulations are used, they are computationally intensive and yield insufficient data for deep learning applications.In this work, we present the PeDesCar dataset for safe autonomous driving, which spans around 15 days of simulated time and encompasses over 1 million collision events, each constrained within a temporal window of up to 2 seconds per event.The dataset is generated using MuJoCo as a physics simulator, proving its effectiveness in sim2real robotics research.We use PeDesCar to train and assess stateof-the-art models in human motion prediction and validate the realism of the simulation against realistic highfidelity finite element simulations.Our results validate that PeDesCar is sufficient for preliminary carpedestrian collision research.The visualization code and dataset are accessible on the project website https: //github.com/QualityMinds/PeDesCar.
Source: Computer Science Research Notes
Authors: Edgar Medina, Leyong Loh
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4412762993