Research On Urban Road Traffic Sign Detection And Step-by-step Recognition System In Complex Scenes

With the rapid development of artificial intelligence,intelligent driving has gradually become a reality.As its foundation,traffic sign detection and recognition has become extremely important.The urban road environment is complex,and the traffic signs also have problems such as deformation,fading,illumination,occlusion,background interference,weather interference and so on.The traditional methods are insufficient in reliability and real-time performance,and can not meet the actual life application.To improve the speed and complexity of traffic sign detection,The main work of this paper is:1)Create a dataset.Establish scene data set for image scene recognition;The improved mosaic data enhancement method is used to obtain the detection of traffic signs;According to the label file in the detection data set,the label image is cut to obtain the recognition data set,which is expanded by zooming,translation,clipping and so on.2)Design a scene adaptive image preprocessing algorithm.Firstly,the algorithm will adaptively scale the input image to meet the needs of the network and retain the image information to the greatest extent.Aiming at the fuzzy characteristics of traffic signs in different scenes(rain,snow,fog,night and normal),Alexnet neural network is trained to recognize the scene of the image,and the corresponding image preprocessing is carried out according to the scene category,so as to finally get the image to be detected.The experimental results show that the accuracy of picture scene recognition is 99.5%,the features of traffic signs are prominent in the processed image.3)Proposes a traffic sign detection algorithm based on improved YOLO-V5.The improved YOLO-V5 uses DIK-means algorithm to cluster to get the number of the highest priority check boxes;In view of the large difference between the number of background frames and target frames and the mismatch of classification difficulty in the training process,CIoU Loss,Focal Loss and binary cross entropy are used to calculate the regression loss,confidence loss and classification loss respectively;The decouple head structure is used as the probe,combined with the simple algorithm integrating a priori mechanism to complete the traffic sign prediction and speed up the regression speed of the algorithm.The algorithm improves MAP to 86.11%and FPS to 156.39.4)Design a step-by-step traffic sign recognition algorithm based on improved AlexNet.The improved AlexNet takes Adam as the optimization algorithm and BN as the normalization function;Firstly,the image correction algorithm based on the contour of the region of interest is used to correct the input traffic sign image,and the results are trained into three models.Finally,the step-by-step recognition of warning signs,prohibition signs and index signs is completed.Combined with the detection results of traffic signs,the step-by-step recognition of traffic signs is completed,and the effect of MAP is 85.97%and FPS is 148.47.