Intelligent collision avoidance system for urban scenarios using deep learning

Abstract

Pedestrian collisions represent a significant challenge in road safety research, attributable to the high vulnerability of pedestrians and the considerable implications for traffic flow and public health systems. Such incidents frequently result in severe injuries or fatalities and can cause notable disruptions to urban transportation networks. In densely populated and dynamically evolving urban environments, a comprehensive understanding of road infrastructure and pedestrian behavior is essential to enhance safety outcomes. In response to these concerns, this study proposes a novel, efficient, and highly accurate automated collision avoidance system, built upon an optimized YOLOv9 deep learning architecture. The system is designed to proactively detect and mitigate potential pedestrian collision scenarios by analyzing critical behavioral and environmental indicators, including pedestrian exposure, jaywalking behaviors, intersection density, and high-risk turning zones. The effectiveness of the proposed system is validated through extensive experimentation on two benchmark datasets, City Persons and KITTI. The system achieved notable performance, with precision rates of 98.79% and 96.62%, and mean Average Precision at IoU threshold 0.5 (mAP@0.5) of 96.21% and 94.83%, respectively. These results highlight the model’s ability to provide both high detection accuracy and real-time inference capabilities, underscoring its potential as a robust and scalable solution for enhancing pedestrian safety in complex urban settings.