A Traffic Sign Detection And Recognition System Based On Deep Learning

As a subsystem of ITS,traffic sign recognition(TSR)plays a positive role in assisting driving,regulating traffic order,improving traffic efficiency and reducing the incidence of human traffic accidents with the rapid development of artificial intelligence.In the actual detection and recognition,traffic sign detection and recognition in natural scenes are easily affected by light and dark,scene complexity,foreign object occlusion,distance and other factors,and the detection and recognition of traffic signs belongs to small object detection in complex scenes.The pixel and size in an image are often very small,which is more difficult than ordinary object detection.At present,although the mainstream deep learning target detection algorithm is constantly improving for common target detection and recognition,the performance of specific targets such as traffic signs still needs to be improved.Based on the above research significance and existing problems,this paper studies the mainstream deep learning target detection algorithm,aiming at the characteristics of traffic sign target and the shortcomings of the current algorithm,improves the performance of the original algorithm,improves the accuracy and speed of traffic sign detection,and further studies the design and implementation of traffic sign detection and recognition system on the embedded platform.The main work of this paper is as follows:1.Aiming at the problem of small target size and complex scene in traffic sign detection and recognition in natural scene,the traffic sign detection algorithm based on improved darknet-62 is studied to improve the detection accuracy.By deepening the convolution layer of feature extraction network Darknet to 62 layers,the shallow and deep feature information of traffic signs can be more fully extracted.A five scale prediction network based on feature pyramid network FPN is proposed,and improves the image quality by fusing shallow high-resolution detail features and deep abstract semantic information.The loss function of the algorithm is improved based on GIo U index Replacing the common loss function,which makes the network model training effect more accurate;Improve the central positioning point of the anchors of the output feature map,and specifically enhance the accuracy of the location at the corner of the grid cell of the feature map,or the small traffic sign location detection in adjacent area.2.Aiming at the problems in(1),taking the accuracy and speed as the target,the yolov3 algorithm with a balance of detection accuracy and running speed is selected to study the traffic sign detection algorithm based on improved yolov3.Combining BN layer and volume layer without affecting the performance of the model,reducing the number and parameters of network model and improving the forward reasoning speed of algorithm.Improve the original loss function and the anchors to improve the overall detection accuracy and the detection effect of the original yolov3 algorithm for small targets.Fit the size of the target traffic sign data set by K-means,recluster theanchors size,improve the accuracy of the positioning of the target prediction frame and speed up the detection speed.After the improvement,the algorithm has achieved good results in the balance and improvement of precision and speed.3.According to the system function,the whole structure of the embedded system is designed,the system hardware and software platform environment is built,the serverside training model is transplanted,and the program code is written to realize the voice broadcast function.Through the system function real-time test based on dynamic video,the embedded traffic sign detection and recognition system in this paper can achieve the expected function and maintain the embedded level good running speed on the platform.