AcciSense: An Automated Accident Detection and Real-Time Alert System Using Deep Learning

Abstract

Prompt hospital care in the “Golden Hour” increases survival after road traffic injuries. However, the current reporting and/or alarm mechanism for accident are by human intervention, monitoring video passively or with time delay response (time is very important to save people who have accident). AcciSense: Cloud-Based Real-Time Accident Detection and Notification through Vision 12 This paper introduces AcciSense, a vision-based on-device accident detection and alert system aimed at reducing response time.

The YOLOv8 deep learning model is used by the proposed system to achieve a high accurate detection of vehicular crashes in live video feeds. When detected, an event-driven video buffer retrieves a brief evidence clip of the seconds before and after a collision. This visual information - complete with geo- tagged and time-stamped metadata - is sent automatically to first responders over the cloud via a messaging interface. Experiments show that the system is effective with a mean Average Precision (mAP50) of 0.90 and 100% alert precision if using a confidence threshold at at level 0.926, guaranteeing its real-time performance without any false alarm. The experimental results demonstrate the feasibility of our framework in enabling real-time accident situation detection and emergency response coordination for the smart transportation.