Use of a deep learning system for autonomous accident identification in tunnels with inadequate CCTV surveillance

Abstract

This research uses the Object Detection and Tracking System (ODTS) in tandem with the Faster Regional Convolution Neural Network, a popular deep learning network. Automatic identification and monitoring of unanticipated occurrences, such as (1) Wrong-Way Driving (WWD), (2) Stopping, (3) People Getting Out of Vehicles in Tunnels, and (4) Fires, will be implemented via the use of a (Faster R-CNN) for Object Detection and a Conventional Object Tracking method. Bounding Box (BBox) findings from Object Detection are accepted as input by ODTS, and ID numbers are assigned to each moving and detected object based on a comparison of the Boombox from the current and prior video frames. Conventional object detection frameworks often fail to accomplish the ability to follow a moving item over time, but our system does just that. Average Precision (AP) values of 0.8479, 0.7161, and 0.9085 were achieved for target objects of cars, people, and fires, respectively, when a deep learning model in ODTS was trained using a datasets of event photos in tunnels. The ODTS-based Tunnel CCTV Accident Detection System was then evaluated with four movies including each accident, using the trained deep learning model.