Research On Traffic Sign Detection And Recognition Based On Machine Learning

Traffic sign detection and recognition is an important technology for assisted driving systems.With the development of deep learning,many scholars have conducted a large number of excellent researches on the public traffic sign data sets at home and abroad,but the detection result for small targets is not ideal.Therefore,this paper takes the traffic signs in the panorama image as the research object,and proposes several improvement strategies for the problems of low detection efficiency and poor real-time performance of small-size traffic signs in high-resolution images.The main research contents of this article are:(1)Construct a new traffic sign data set N-TT100 K.On the basis of TT100K(Tsinghua-Tencent 100K),by collecting traffic signs around the campus and crawling pictures on the Internet,data expansion and data enhancement methods are used to achieve data balance processing for the imbalance of data categories;then use labelme to label the data,The experimental results show that the new data set has a better detection effect.(2)Aiming at the poor detection effect caused by the small proportion of small-size traffic signs in the image and the image blurring,a lightweight Super Resolution Net(SRN)was designed.Resolve the reconstructed image and improve the resolution of the small-size traffic sign area.From the experimental results,it can be seen that SRN increases the semantic information of small-size traffic signs,and the m AP of the Efficient Det model on the N-TT100 K data set reaches 90.73%.(3)In view of the need to set Anchor-related hyperparameters before training in the Efficient Det network,and a large amount of calculations,resulting in waste of memory resources and poor generalization,the N-EDet network was inspired by the idea of FCOS.The prediction part uses the element-wise method to predict positive and negative samples,and adds a centrality prediction branch,which suppresses the positive samples at the boundary points of the real frame through centrality,improves the detection speed and reduces the false detection rate.Finally,all optimization strategies are integrated together to obtain the TS-EDet model,which is tested on the N-TT100 K.Experimental results show that the average accuracy of the algorithm has reached 92.57%,and the detection speed FPS reaches55.5.Compared with models such as Tiny-YOLOv3,TS-EDet has a better performance in the detection and recognition of small-size traffic signs.