Localized AI-Based Urban Hazard Detection for Defensive Driving Support

Abstract

This research investigates the feasibility and effectiveness of implementing localized artificial intelligence-based system for enhancing defensive driving through automated urban hazard detection. For this purpose, a lightweight You Only Look Once version 11 nano (YOLOv11-nano) model containing 2.6 million parameters was fine-tuned on a localized dataset consisting of 18,132 images collected over 300 days in Chiang Mai, Thailand. These results reveal a mean Average Precision (mAP) of 0.96 when using an intersection over union (IoU) threshold of 0.5. Even if the precision rate is quite high in ideal conditions, there is an observed decrease of 13.1 percent in the Average Detection Confidence due to monsoon rain and night-time conditions. Thus, a safety margin has been established for the application of Advanced Driver Assistance Systems. The reliability test conducted across seven localized object categories reveals consistent detection performance for vulnerable objects, such as motorcycles and pedestrians. Furthermore, a cost-benefit analysis based on localized accident cost data confirmed the economic viability of implementation in commercial transporting services.