Traffic Sign Recognition Technology Based On Deep Learning

Nowadays the domestic market economy is booming day by day,intelligent transportation system（ITS）,which aims to alleviate traffic congestion,ensure traffic safety and save energy,came into being.Intelligent image processing and target recognition technology for fast and accurate identification of traffic lights is an important content of ITS.How to ensure the high accuracy and rapidity of traffic signal recognition algorithm has always been a research hotspot.This paper studies the traffic signal recognition algorithm based on deep learning in complex environment.In order to make the recognition algorithm more suitable for my country’s traffic signs,this paper expands the images of traffic signs by shooting traffic signs under various traffic conditions on the basis of the TT-100 K data set.The Label Img image annotation tool is used to label the traffic sign images one by one.Based on the color and shape of the traffic sign,a rough classification data set and a fine classification data set of traffic signs are established respectively.Aiming at the two representative algorithms Faster R-CNN and YOLOv4 in the field of target recognition,the two algorithms are trained into rough classification and fine classification of traffic signal data respectively.Through the analysis of map positioning accuracy and positioning speed per second,the performance of the two algorithms is evaluated.Though the recognition precision of YOLOv4 is a little poorer than that of Faster R-CNN,the recognition speed of YOLOv4 is extremely faster than that of Faster R-CNN.YOLOv4 recognition algorithm is better than Faster R-CNN algorithm.so YOLOv4 algorithm is used as The basic algorithm of traffic sign recognition technology.In order to further improve the comprehensive performance of YOLOv4 algorithm in traffic sign recognition,the algorithm is optimized and improved.Using methods such as rotating the image,reducing image saturation,enhancing or reducing image brightness,and image blurring,the traffic sign data set is enhanced to enhance the robustness of the algorithm;Mobile Netv3 is used as the backbone network of the improved algorithm,by introducing depth The convolution and channel attention mechanisms can be separated to extract the deep features of the image,decrease model complexity and increase network identification speed;By optimizing the pre frame generation method based on kmeans cluster analysis,the generation size of pre frame is adjusted in light of the features of traffic signal to accelerate model conjunction;Soft-NMS is used to optimize the design of the prediction box screening mechanism to obtain a more accurate prediction box;The loss function EIo ULoss is used to replace CIo ULoss,learn from focal loss to shorten the convergence time of loss function and optimize the regression effect.Simulation outcome manifest that the modified YOLOv4 markedly elevates the recognition velocity of the algorithm by reducing the recognition precision slightly.Compared with other modified lightweight algorithms,the lightweight effect of the new model is more obvious and the overall manifestation of the model is better.Finally,a visual traffic sign recognition system is designed based on the GUI toolkit Py Qt5 in the standard Python library.By calling the traffic sign recognition model,the traffic signs in the image can be accurately recognized.