| Traffic signs play a vital role in guiding safe driving,alleviating urban congestion,and reducing traffic accidents.In the intelligent driving system,the detection and recognition of traffic signs has always been the focus and difficulty of research.With the maturity of deep learning technology,convolutional neural networks are increasingly used in the field of traffic sign detection,and have achieved good results in traffic sign data sets with good environments.However,in real road scenes,traffic signs are easily affected by complex environments,such as weather,occlusion,and light.In addition,smart cars often capture panoramic images,and traffic signs account for a small proportion of the panoramic images and have a wide variety of them,making it difficult to detect.At the same time,the graphics processor will encounter resource constraints when processing high-resolution panoramic images,which makes it difficult for the convolutional neural network model to achieve the effect of real-time detection,and even the model is difficult to train.In order to solve these problems.In this paper,a small target oversampling training data generation method,image segmentation and geometric perspective detection preprocessing method,and an improved lightweight neural network are used to propose a lightweight panoramic image traffic sign detection method based on deep learning.the specific work is as follows:1.If the resolution of the original panoramic traffic image is too large,it will encounter the problem of insufficient GPU resources.This paper proposes a random cropping algorithm based on traffic signs to preprocess the high-resolution images of the training set to obtain image blocks that match the input of the neural network model.The image block enriches the number of small targets on the basis of retaining the original information of the traffic signs,making the trained model more sensitive to small targets.2.The missed detection rate of Tiny YOLOv3 is higher in panoramic traffic scenes with dense small targets.Aiming at this point,this paper proposes the Improved-Tiny-YOLOv3 algorithm.Sufficient shallow features are beneficial to small target detection.In this paper,the shallow features and deep features are merged through feature reorganization to achieve the third scale that is more conducive to small target detection.And CIoU(Complete Intersection Over Union,CIoU),which is more in line with the target box regression mechanism,is used to improve the bounding box position error term in the model loss function,thereby improving the network performance.3.In the detection stage,this paper designs an image detection preprocessing method based on the priori information of the geometric perspective in the panoramic image,uses the idea of divide and conquer,uses high-resolution images to block traffic signs in dense areas,and uses down-sampled images in other areas.Image segmentation balances the contradiction between image block redundancy and image information retention to a certain extent.In addition,this paper designs a complete detection process to ensure the smooth development of traffic sign detection.Based on the above methods,this paper conducts comparative experiments on the TT100 K data set.The results show that the proposed method not only performs well in terms of detection accuracy,but also greatly improves the detection efficiency.Besides,it has good detection effect for different kinds of traffic signs with strong robustness. |
