Design and Experimental Evaluation of Nighttime Traffic-Sign Detection and Classification Based on Low-Light Enhancement

Abstract

This paper focuses on solving the problem of traffic-sign detection and classification at nighttime when lighting is often lower. Images taken by the camera at night are underexposed and fail to show details. Low-light enhancement method can improve the brightness of the image. By studying four different low-light enhancement methods and applying four methods to the detection and classification of traffic signs, the performance of low-light image enhancement is compared. The evaluation results show that the nighttime traffic-sign detection and classification method, which combines a robust end-to-end convolutional neural network (CNN) and Zero-Reference Deep Curve Estimation (Zero-DCE) low-light enhancement, has high precision and a higher recall rate compared to other methods.