Research On Traffic Sign Detection Method Based On Deep Learning

In recent years,the number of motor vehicles on urban roads has continued to rise under the favorable background of China’s rapid economic development,steady rise in national income,and highly developed urban transportation network.Thus,people’s transportation comfort has been greatly improved,and the quality of life has undergone a qualitative leap.However,the influx of motor vehicles into urban roads has also brought a huge burden to urban traffic safety,including a series of social problems such as traffic accidents and property losses.With the advancement of science and technology and the rise of artificial intelligence technology,Intelligent Transportation System is expected to ease the pressure on urban traffic,by reducing the incidence of road traffic accidents,and the loss of people’s property.As an important part of Intelligent Transportation System,traffic sign detection has been attracting more and more researchers’ attention.Due to the influence of factors such as complex background environment,extreme weather factors,and shooting equipment,the detection approaches of traffic signs can hardly achieve satisfactory accuracy in real scenes.Considering the speed requirements of traffic sign detection and the efficiency advantages of single-stage object detectors,this thesis focuses on traffic sign detection based on YOLO series algorithms on the TT100 K and CCTSDB datasets published in China’s real road scene.This thesis attempts to improve the accuracy of traffic sign detection and make an effort for the development of Intelligent Transportation System.The main work of this thesis includes the following aspects:(1)Through the analysis of TT100 K dataset,it is revealed that there is a serious sample imbalance problem in this dataset.This thesis uses data enhancement method to expand the dataset.(2)This thesis analyzes the problems lied in the YOLO V3 object detection algorithm.Then,this thesis integrates the attention mechanism and dilated convolution to improve the YOLO V3 model.The detection m AP of the improved YOLO V3 model achieves 84.33% on the TT100 K dataset,which is 4.42% higher than the baseline model.On the CCTSDB dataset,the detection accuracy has also been significantly improved.(3)In order to further improve the detection accuracy,we further study the traffic sign detection by improving the YOLO V5 object detection algorithm.This thesis integrates the balanced feature pyramid and GC block to complete the optimization of the model.The detection precision and recall of the optimized model are greatly improved on the TT100 K dataset,and the proposed method achieves a m AP of 89.7%,which is 1.9% higher than the baseline method.(4)For applying the proposed traffic sign detection method to practical scenarios,this thesis implements the optimized YOLO V5 model as a web-based traffic sign detection system.The system is based on B/S architecture.Users can realize real-time detection of traffic signs in pictures and videos through local browsers.The system has simple web pages,convenient operation and friendly human-machine interface.