Traffic Sign Detection Based On Deep Learning

Traffic sign detection has an important place in the field of assisted driving as well as unmanned driving.In this field,there is a blossoming of technical approaches to detect the obtained sign images.Among them,traffic sign detection using deep learning detection algorithms has become a key research area in traffic sign detection today.In road traffic,due to the complex road traffic conditions and the possible physical discomfort or mental inattentiveness of drivers,various problems often occur when drivers identify traffic signs with their naked eyes alone,so the need for traffic sign detection by other means is gradually increasing,while in intelligent transportation and autonomous driving,traffic sign detection is very important and is a road driving It is an important guarantee for road driving safety.With the increasing maturity of the development of convolutional neural networks,the application of target detection to traffic sign detection has become an increasingly emerging research area.Since the object of traffic sign detection is often the picture taken by the vehicle-mounted camera,it is necessary to select the picture of the scene under real traffic conditions and to consider the stretching deformation of the picture that occurs when the vehicle is driving at variable speed.In the actual traffic sign detection,since Faster R-CNN has high localization accuracy but slow detection speed,it is suitable for low-speed driving in complex traffic scenes;YOLOv3 algorithm has high real-time performance,but poor localization accuracy,and is suitable for detection in high-speed driving situations.However,since the algorithm designers started to detect for all categories of targets,this paper improves both different algorithms for these two categories to make them better for traffic signs and thus achieve higher detection accuracy,and the related research conducted in this paper is as follows.(1)Production of data set.The main body of the dataset used in this paper is the Chinese traffic sign detection dataset,in which the real situation pictures and the stretched and deformed pictures are selected,and the pictures with Gaussian noise and the pictures taken by vehicle cameras from the laboratory are added,including the pictures of traffic scenes in foggy and rainy days,and then uniformly labeled to complete the production of the traffic sign dataset.(2)Improvement of Faster R-CNN algorithm.Since the initial Faster R-CNN algorithm targets detection objects are often animals or people with large target size,which are targets with a relatively large proportion in the picture.In this paper,the target detection object is traffic signs,and in many cases,traffic signs occupy a relatively small proportion of the picture,so the default value of anchor in the original RPN network is improved.And the original VGG-16 model is changed to ResNet101 model to further improve the detection accuracy of small targets in complex situations,and finally the relevant experiments are conducted to compare the experimental results,and the results show that the detection accuracy has been improved to some extent.(3)Improvement of YOLOv3 algorithm.the detection speed of YOLOv3 algorithm has been improved compared with Faster R-CNN algorithm,but the detection effect of small targets,as well as clustered small targets,is not very good.this paper improves the detection effect of small targets for traffic signs,by adjusting the convolutional network layer structure of Darknet-53,so that YOLOv3 By adjusting the convolutional network layer structure of Darknet-53,YOLOv3 is able to detect some small traffic signs with good results,and finally,relevant experiments are conducted to show that the detection accuracy is improved.