Research On Traffic Sign Detection And Recognition Based On Machine Vision

With the rapid development of social economy and the improvement of people’s living standards,the output of cars is increasing year by year,road traffic congestion and traffic accidents occur frequently,and the application of intelligent transportation system emerges.As an important part of intelligent transportation system,traffic sign detection and recognition are widely concerned by scholars at home and abroad.However,the actual road environment is changeable and complex,which brings challenges to the detection and recognition of traffic signs.Therefore,the research on the detection and recognition of traffic signs has certain practical significance.Considering all kinds of interference factors and the real-time requirements of the algorithm,the main work of this paper is as follows:Aiming at the problem of traffic sign detection which the objects with larger category and smaller size,this paper proposes a coarse classification detection strategy based on the existing YOLOv4-tiny single-stage algorithm,and then increased the attention to small targets by adding a layer of feature pyramid network structure.Through policy adjustment and algorithm improvement,the improved YOLOv4-tiny achieved 86.27% detection effect on CCTSDB dataset,which was 6.16% higher than the original network.It shows that the improved YOLOv4-tiny network model has high detection accuracy and good generalization ability and it has good performance in the detection of three main kinds of traffic signs in China.Aiming at the problems of low recognition accuracy and poor real-time performance of traffic signs,two different recognition algorithms based on convolution neural network were proposed to classify traffic signs.In the first method,residual connection,inception module and dense connection were introduced to improve the structure of traditional CNN.At the end of the network,3 × 3 convolution was used to replace the full connection layer.The accuracy of the improved CNN in the large traffic sign database GTSRB reached 98.71%,and the recognition time of each image was about 0.158 ms.It shows that the algorithm has high accuracy and good real-time performance.The second method concatenates the improved semantic segmentation network UNet with the classification network LeNet.Through the improved UNet,the regions of interest of traffic signs were extracted accurately,so that LeNet can directly learn the features of interest in the image.The accuracy rate of this method on the self-made data set reaches 98.96%.Although the time-consuming was still insufficient compared with the first method,this method relies less on data sets and shows good classification ability on a small number of data sets.Comparing two different traffic sign recognition algorithms,the traffic sign detection and recognition methods were integrated,and the algorithm testing process was completed on the actual road by connecting the computer with the camera.It shows that the proposed method can not only deal with the problem of low detection accuracy of traffic signs in the case of unbalanced data,but also recognize traffic signs in a short time and show good robustness,which has a certain guiding significance in practical application.