Research On Traffic Sign Detection Algorithm Based On Deep Learning

With the great rejuvenation of the Chinese nation,China’s economy,science and technology are in the forefront of the world,the number of cars is also increasing,traffic safety issues have attracted increasing attention,traffic signs detection has become a hot topic.Although the detection algorithm based on deep learning is fast,it has the problems of complex network structure and low detection precision.In order to solve the above problems,this paper uses deep learning technology to carry out the following research:(1)In order to better suit the national conditions of our country,an improved TT00 K data set is established by studying and analyzing the domestic TT100 K traffic signs data set.There are some problems in the original data set,such as large background area of image,small size of image and large memory.In the original dataset,there are some problems that the amount of data of some classes is too low.This paper adjusts the data annotation and increases the amount of data of these classes.The original data set has the problem of single image form and low learning complexity of neural network.In this paper,some data are processed to increase the learning complexity.(2)In this paper,the popular deep learning target detection algorithm is analyzed,and YOLOv4 and YOLOv4-Tiny model are selected as the basic network model considering the detection accuracy and speed of the algorithm.YOLOv4 has high detection accuracy,but the network model is not lightweight and difficult to migrate to mobile devices or autopilot assistance systems.Ghost Net is a lightweight feature extraction network,which can generate redundant feature graphs with fewer parameters through lower computational complexity.In this paper,an improved convolution network algorithm based on YOLOv4 model is proposed by deeply studying the YOLOv4 model.The improved algorithm is to use Ghost Net lightweight trunk network to replace the CSPDarknet53 network structure,and finally add the NMS algorithm and CIOU loss function into the improved algorithm.Experimental results show that the improved algorithm not only has higher accuracy than the original algorithm,but also has only one sixth of the parameters of the original algorithm model,and the real-time detection is strong.(3)YOLOv4-Tiny is a lightweight target detection algorithm,which can be transplanted to embedded devices,but the detection accuracy of the algorithm is too low to meet the needs of the traffic sign detection task.Attention mechanism mainly includes channel Attention mechanism and spatial Attention mechanism.Adding Attention mechanism module to algorithm model can make the model pay more attention to semantic information.In this paper,an improved convolution network algorithm based on YOLOv4-Tiny model is proposed by deeply studying YOLOv4-Tiny model.Due to the lack of SPP pooling structure in YOLOv4-Tiny network,the receptive field is greatly limited,and the lack of down-sampling operation leads to insufficient feature fusion.In addition,in the FPN structure of the original algorithm,the detection accuracy of the original algorithm is insufficient due to the lack of semantic information of small targets.In order to solve this problem,attention mechanism is added to the backbone network and feature fusion module of the original algorithm model to enrich the semantic information of the target and improve the detection accuracy.Experimental results show that the improved algorithm can improve the detection accuracy compared with the original algorithm,and meet the requirement of the traffic sign detection task.