The Research On Traffic Sign Detection And Recognition Algorithm Based On Deep Learning

The task of detection and recognition of traffic signs is an important technical component of autonomous driving and assisted driving.There are still many problems in the current research on the detection and recognition of traffic signs in real scenes.First of all,when a vehicle is driving on a road at a relatively high speed,it must be able to detect and make judgments in a timely manner in a long-distance scene which means the traffic sign is very small.The higher driving speed also requires the identification system to have a fast processing speed.At the same time,the road conditions in the real scene are more complicated,and traffic signs will be accompanied by insufficient light,deformation and damage,extreme weather and other impacts,which also brings major challenges to the detection and recognition tasks of traffic signs.In order to solve the above problems,this article is based on the deep learning convolutional neural network theory,combined with the characteristics of the domestic open source traffic sign data set,and carried out related network model improvement work.The main contents are as follows:(1)First,the current typical open source traffic sign data set was analyzed.We find the target area ratio in TT100k data set is very small,target area ratio in CCTSDB data set is common.Then we make the data set mixed and enhanced,including the tilt processing based on geometric shapes,extreme weather processing and HSV-based color gamut transformation processing to increase the sample richness of the data set,which is conducive to the network learning of sample characteristics in different environments.(2)To solve the problem of insufficient detection accuracy of small and medium traffic sign targets,this paper proposes an improved detection and recognition network for small and medium traffic signs based on YOLOv3.First of all,this article draws on the idea of ResNext’s multi-branch structure,by replacing the residual structure of the backbone network with this structure,to achieve the purpose of improving the feature extraction ability of the backbone network.Since the shallow feature layer of YOLOv3 strengthened extraction network(FPN)focuses on detecting small and medium-sized targets,this paper connects the four-fold downsampling layer of the original backbone network to FPN to achieve the purpose of strengthening the extraction of semantic information of small and medium-sized targets.At the same time,this paper draws on the characteristics of fusing multi-scale receptive field information with the hole convolution in the RFB network,and performs the 3*3 convolution kernel replacement improvement on the RFB network,and connects the improved RFB to the quadruple downsampling layer to complete the purpose of enhancing the extraction of small target details.Experiments show that compared with the original YOLOv3 network,the improved network has a 4.32%increase in detection accuracy on the TT100K data set.(3)To solve the problem that the current mainstream network detection speed is slow and the project cannot be implemented.We propose a lightweight traffic sign detection and recognition network based on improved YOLOv4.First of all,in order to lighten the processing of the backbone network,this paper introduces the MobileNetv2 network inverted residual structure and the deep separable convolution idea:the inverted residual structure can avoid the loss of feature information in the process of extracting features from the network.Depth separable convolution uses point-by-point convolution and channel-by-channel convolution without reducing too much accuracy,while greatly reducing the amount of network parameters.So we replace the MobileNetv2 network with the backbone network.At the same time,in order to lighten the processing and strengthen the extraction network(PANet),we use deep separable convolution to replace the 3*3 convolutional layer in the PANet structure.Experiments show that the improved network parameters are reduced by 82.85%,and the detection speed is increased from 9.4FPS to 25.02FPS.