Published Paper Details:

Accident detection is an essential application in intelligent transportation systems for the safety of drivers and passengers. In recent year, deep learning based objects detection models have shown significant improvements for detecting objects in real-time. YOLO (You Only Look Once) is one such model that has gained popularity due to its real-time performance and high accuracies. In this paper, we propose an accidents detection system using YOLOv3, a state-of- the- art version of YOLO. The proposed system is designed to detect three types of accidents, namely vehicle rollover, rear-end collision, and head-on collision. The system uses a pre-trained YOLOv3 model which is trained on the COCO dataset, which is then fine-tuned on a custom dataset of accident images. The proposed system achieves an average precision of 0.94 for vehicle rollover detection, 0.93 for rear-end collision detection, and 0.92 for head-on collision detection. The system also shows promising results in terms of real-time performance, with an avg processing time of 0.03 seconds per frame on an NVIDIA GeForce GTX 1080 Ti GPU. The proposed system can be also integrated into intelligent transportation systems which provides real-time accident detection and alerting, improving the safety of drivers and passengers on the road.