Researcher(s)
- Duy Duc Tran, Computer Science, University of Delaware
Faculty Mentor(s)
- Weisong Shi, Computer Science, University of Delaware
- Yuankai He, Computer Science, University of Delaware
Abstract
Autonomous vehicle (AV) safety remains a critical concern due to widespread skepticism from both citizens and governments. Ensuring AVs can safely interact with traffic elements, particularly vulnerable pedestrians, is essential. This study introduces a novel testing method using an indoor testbed to model pedestrian interactions with AVs, addressing key challenges such as cost efficiency, scalability, realistic sensor data testing, and ethical considerations. Building on the ICAT platform, we developed a robotic pedestrian model using a Raspberry Pi 4B, with a custom object detection model trained on YOLOv5 and deployed on a Jetson Xavier device using TensorRT and PyCUDA for real-time detection. Although the autonomous navigation of the pedestrian robots is still under development, preliminary results demonstrate the feasibility and potential effectiveness of this approach. Our study offers a scalable and ethical alternative for AV testing without real human subjects, and we outline the next steps to enhance AV safety testing further.