Applications of Deep Learning to Road Sign Detection in DVR Images

Abstract

This paper applies deep learning to road sign detection based on the images obtained from DVRs. Three convolutional neural network based state-of-the-art real-time object detection systems are explored for road sign detection, where the road sign plates are vertical or horizontal. First, collect the images of street scenes including road signs as an image library, which is divided into three parts for training, validation and testing. Secondly, according to the exposure values (EVs) of images, divide the images into three categories, high exposures labeled as EV -H, middle exposures labeled as EV-M, and low exposures labeled as EV-L. Thirdly, the road signs in each image are labeled to perform supervised learning. Fourthly, apply three object detection systems to road sign detection, You Only Look Once (yOLO) version 2 (v2), YOLO version 3 (v3), and Single-Step Multi-Box Detection (SSD), where an improved method based on YOLO v2 is proposed by dividing the images into three kinds of meshes and recalculating the values of the anchor boxes by k-means algorithm. Finally, analyze the mean average precisions and the frame rates of all images to conclude the best models for the three conditions EV-L, EV-M and EV-H.