Research On Detection And Recognition Method Of Lane Arrow Signs In UAV Images

The detection and recognition of traffic signs is an important part of the intelligent driving assistance system.By combining lane line detection technology and major map software,the road surface information can be fed back to the vehicle driver in time,which greatly reduces the frequency of traffic accidents.However,the increase in vehicles has also accelerated the wear and tear of road traffic signs and increased the burden of road maintenance.However,the current road sign sites survey methods have problems such as high cost and low work efficiency,which not only delay the road maintenance and update schedule,but also affect the road traffic safety.Therefore,this paper takes UAV image as the object,and takes advantage of the advantages of UAV image acquisition speed,wide range and high resolution,proposes a detection and recognition method of UAV image lane arrow signs based on deep learning,which provides conditions for large-scale monitoring of arrow signs.The main work and research content of this paper are as follows:(1)Datasets production.There is currently no public lane arrow sign data set.In order to carry out related research,this article uses UAV images to produce two sets of high-quality UAV image lane arrow sign data sets according to the requirements of different algorithms.The data set contains arrow road signs in different scenarios,which enhances the universality of the data set while reducing the imbalance of the data.(2)An improved Mask R-CNN arrow road sign detection and recognition method is proposed.Aiming at the shortcomings of traditional methods such as low accuracy,high time complexity and poor robustness in road sign detection,this paper proposes an improved Mask R-CNN arrow road sign detection and recognition method.First,perform image enhancement and denoising algorithm processing on the images in the dataset,and then input the processed images into the backbone network “residual network 101(Res Net101)+ feature pyramid networks(FPN)” for feature extraction.This paper also adjusts the aspect ratio of the anchor frame in region proposal network(RPN),and replaces the candidate frame selection method non-maximum suppression(NMS)with relaxation soft non-maximum suppression(Soft-NMS),which improves the accuracy of model detection and recognition and is more suitable for UAV images.The experimental results show that the mean average precision(m AP)value of the method proposed in this paper is 98.33%,which is higher than the other two comparison methods,indicating that this method has a good detection and recognition effect on arrow road signs in small-area images,but it is effective for large images without cropping.The detection and recognition of arrow road signs is poor.(3)An improved YOLOv4 lane arrow signs detection and recognition method is proposed.Aiming at the shortcomings of the method in(2)in arrow road sign detection and recognition,this paper proposes an improved YOLOv4 lane arrow signs detection and recognition method.First,on the basis of the data set used in the above method,a large number of original UAV image images without cropping were added to increase the detection and recognition ability of large images;secondly,the K-means clustering algorithm was used to retrieve 9 anchor frames,and passed Introduce the latest optimizers adaptive moment estimation projection(Adam P)and stochastic gradient descent projection(SGDP)to improve the speed and accuracy of road sign detection and recognition.The experimental results show that the m AP of this method is 95.63%,and the F1-measure is 94.08%,which is higher than other comparison methods.It proves that the proposed method can effectively detect and recognize lane arrow signs in complex scene images,and verifies the effectiveness of the proposed method.