Research On Traffic Sign Recognition Based On Deep Learning In The Nature Scenes

With the rapid development of the society and economy,the population concentrated in the big and middle type cities has brought enormous opportunities and challenges to the cities,including the problems of traffic congestion,and the concept of intelligent transportation system came into reality.And the intelligent transportation system,including the technologies behind self-driving vehicles,requires the real-time digitization of various traffic conditions on the roads.How to eliminate interference from the environmental factors as much as possible,and identify the traffic signs correctly and efficiently are key technical issues that must be solved.The early traditional methods of the traffic sign recognition divided the area by features such as the color and the shape of the images,and then use the methods like the template matching,the SVM classifier and neural networks to identify.With the development of CNNs,there have been two major types of the traffic sign recognition methods based on deep learning,including the two-stage detection method represented by the RCNN series,and the single-stage detection method represented by the YOLO series.This thesis conducts research and experiments on the current mainstream deep learingbased the traffic sign recognition methods.Firstly,the image data is preprocessed based on prior knowledge,and the data volume is reduced by 49.14% without losing any important information.With the traffic sign recognition experiment of the faster RCNN method,its shortcomings is illustrated.We improved the method and achieved a 28.31% improvement on average recall in our dataset.With the experiment of the YOLOv3 method,an optimization is applied,and we achieved a 24.15% improvement on average recall in our dataset.