Research On Structured Road Lane Detection Based On Monocular Vision And Lane Departure Detection Based On SAE Algorithm

With the rapidly increase of motor vehicle ownership and the construction of road infrastructure,road transport has brought great convenience to people,but the resulting road traffic accidents have become one of the global security issues.China's car ownership accounts is about 2%-3%of the world,but the death toll accounted for about 22%of the world,leading to China is one of the serious traffic accidents country in the world.Throughout the reasons for the formation of road traffic accidents,the reason that the vehicles do not remain in the right lane or depart from the current lane is known as the second cause of road traffic accidents caused by the driver's factor.Statistics show that if the lane departure warning is provided to the driver before is of potential traffic accident occurred,we can avoid 90%of traffic accidents.In this paper,related technology of image preprocessing,lane detection,lane tracking and lane departure detection are researched.Firstly,structured road image sequences are acquire based on monocular vision,image preprocessing technology(Weighted average method,FIR filter,Otsu's method,and the dynamic region of interest is set based on the vanishing point detection)is used to preprocess image sequences and get the dynamic region of interest which highlighted lane lines.Secondly,Hough transform of polar angle constraint is proposed to detect lane marks based on the segmented line model,and the matching algorithm is used to track the lane marks,the recognition rate is 96.69%,and the real-time processing time is about 20ms per frame,which not only test the expressway lane marks robustly but also meet the real-time requirement.In addition,this paper also studied the identification and fitting of lane marks based on artificial fish swarm algorithm.Finally,a lane departure detection algorithm is proposed.Based on the results of lane detection and matching,the vehicle lateral offset(LO),lane slope and intercept are used as feature input of our lane departure detection classifier which is constructed by Stacked Sparse Autoencoder(SAE)neural networks.Experiments are used to determine the optimal parameter settings and the hidden layer structure in order to obtain the highest recognition rate classifier for lane departure detection.At last,a large number of experiments were carried out,and compared SAE classifier with the other six classifiers(NFT,LO model,S.function,SAE-DN,SVM-LS,SVM-NL),the superiority and effectiveness of the SAE classifier were verified.