Research on Detection and Recognition of Traffic Signs Based on Convolutional Neural Networks

Introduction

With the continuous development of society and the continuous improvement of people's living standards, motor vehicles have become an inaccessible means of transportation in daily life. According to the statistics of the public security department, as of 2017, the number of motor vehicles in China has reached 310 million, and the number of drivers is as high as 385 million, of which cars account for the majority, about 217 million, and the number of car drivers is up to 342 million. As the main body of motor vehicles, the proportion of automobiles has increased year by year, from 54.93% to 70.17%. With the increase in the number of cars, people's travel has brought a lot of conveniences, but also with a lot of trouble, leading to traffic congestion and traffic accidents occur frequently(Abadi, 2016). According to statistics, there are more than one million traffic accidents in China every year, causing a large number of casualties and causing immeasurable losses to the victims and their families. Because of this background, the intelligent transportation system (ITS) and advanced auxiliary driving system (AADs) are emerging in time, and have been widely concerned by researchers at home and abroad.

In the actual driving environment, the driver must maintain a high degree of attention, correct judgment, and keen observation, and be able to make a timely and accurate operation for various situations to avoid unnecessary loss of life and property. However, restricted by psychological and physiological factors, it is difficult for drivers to maintain a high degree of attention and keen observation for a long time, resulting in fatigue driving and driving hidden dangers. With the continuous development of the computer level, it is possible to solve the problems that human beings can't complete themselves through computers. The computer can simulate the process of human detection and recognition of things, and apply it to the advanced auxiliary driving system, which can solve the contradiction between the driver's limitations and the conditions required for safe driving of motor vehicles. It helps to reduce the incidence of traffic accidents and reduce unnecessary loss of life and property.

A traffic sign is an important part of road traffic, which contains important information about road conditions, and is the main body of the road traffic sign detection and recognition (TSDR) system. TSDR system mainly includes two technical links: the first is the detection of road traffic signs, including the regional positioning of traffic signs, feature extraction, classifier classification, and related post-processing. The second is the identification and output of traffic signs, which can extract and identify the information on the traffic signs and provide them to drivers for decision-making. Due to the complexity of the actual driving environment, the image is often distorted to a certain extent due to factors such as illumination transformation, vehicle speed transformation, and obstacle occlusion. In recent years, with the continuous improvement of computer level, image processing(Cao et al., 2017; Feng et al., 2018), computer vision(He et al., 2015a; He et al., 2015b; Krizhevsky et al., 2017), machine learning(Ren et al., 2017), artificial intelligence, and other technologies have been greatly developed. Deep learning has made great achievements in image detection and recognition. Target detection based on deep learning can be divided into two categories: target detection method based on region nomination, It includes R-CNN(Saxena, 2016), SPP-Net, Fast R-CNN(Shi & Bao, 2018),The target detection algorithm based on regression mainly includes YOLO and SSD.

The main structure of this paper is as follows: The first chapter describes the research background and significance of traffic sign detection and recognition. The factors affecting the accuracy of traffic sign detection and recognition are summarized. The main research contents of this paper are introduced to improve the detection accuracy of small-scale traffic signs.

The second chapter mainly introduces the application and development of deep learning in target detection and analyzes the advantages and disadvantages of R-CNN, SPP-Net, Fast R-CNN, Faster R-CNN and RPN. Based on Faster R-CNN, a network structure combined with part of the underlying feature map is proposed to detect small-scale traffic signs.