Vehicle Crash Detection Using You Only Look Once Version 8

As new energy vehicles and advanced driving features proliferate, post-accident handling is often constrained by delays between discovery and reporting. Targeting multisource video from roadside surveillance and unmanned aerial vehicles, this study proposes a vehicle-collision detection and rapid alerting scheme. We curate nearly ten thousand images and annotate them with Labelme, then train a two-level model for collision severity—moderate and severe—based on You Only Look Once, Version 8 (YOLOv8). The confidence and intersection over union (IoU) thresholds are adjustable, and the system supports real-time inference on images, prerecorded videos, and live camera streams. Experiments across highways, urban expressways, and tunnels yield high mean average precision (mAP), while precision–recall (PR) curves indicate low false-positive rates alongside real-time performance. The system can dispatch alerts to traffic police and hospitals, reducing response time and improving public-safety efficiency.