Pedestrian Safety Analysis using Computer Vision

Abstract

The goal of this research project was to explore the capabilities of computer vision for pedestrian safety analysis. Computer vision, an AI application in image processing, tracks the movement of cars, bikes, and pedestrians, offering superior information about speed, trajectory, and count data for various transportation modes. The University of Idaho acquired two computer vision sensors from the startup Numina. This project funded the installation and one year of data access. In collaboration with the City of Moscow, we identified a test location, but the sensors failed to provide the necessary data. Consequently, we pivoted to the open-source computer vision package YOLOv8. Our research then focused on YOLOv8’s capabilities for pedestrian safety analysis. The first task tested detection accuracy, and the second compared different model sizes, or “brain sizes,” of YOLOv8, which range from smaller, faster models to larger, more accurate ones. Accuracy tests compared average detection confidence across various zones and times of day, revealing that cars in high daylight had the highest confidence levels, while objects closer to the camera and oriented perpendicularly were detected more accurately. In contrast, objects at skewed angles and farther distances had lower confidence levels. The model size comparison showed that larger models, despite requiring more time and storage, produced significantly higher-quality detections.