A Fault Detection Method and System for Highway Tunnel dome light Based on Improved YOLO with Locatization Loss Function

Abstract

Sufficient light intensity in the tunnel plays an extremely important role in ensuring the safety of driving in the tunnel. Tunnel dome light is the basic facility to ensure tunnel lighting. The traditional fault detection method is manual inspection, and the discovery of the problem is not timely. In this paper, a tunnel dome light fault detection method and system based on video monitoring is proposed. Constructing a tunnel dome light detection data set. The original positioning loss function in YOLOv5 is changed from CIOD_Loss function to SCALoss function, which is composed of side overlap (SO), corner distance (CD) and aspect ratio (AR) loss. So that the network can generate more penalties for low overlap positioning frames, and the network model has better positioning performance and faster convergence speed, it is more suitable for the dense and small target such as tunnel headlight. After improving the model, the recognition accuracy is improved by 13.1 %, and the positioning loss is also reduced. So that it can quickly and accurately locate the target to be detected. Finally, the fault lamp detection model is constructed to locate the specific location of the fault lamp. The experiment shows that the model has good performance, and it can effectively detect the state of tunnel dome light in real time and detect abnormal working conditions.