Lane Detection Algorithm Based On Deep Learning

With the increasing prevalence of automobiles,the traffic accidents has increasingly risen and seriously affected people's lives and properties.Autonomous driving can greatly reduce traffic accidents,and lane detection is an important component of intelligent driving system.The traditional method of lane detection is mainly based on edge feature or image segmentation,which is easily disturbed by illumination changes,driving vehicles and road damage,resulting in decline of robustness of the algorithm.Under the bad weather and complex environment,it can't achieve the higher accuracy.Deep learning method uses the network model to automatically learn the target features and has a high generalization ability,therefore can improve the accuracy of the lane detection in real scenes.This paper's research of lane detection is based on convolution neural network,and the main work is as follows:1.The methods of image preprocessing are investigated.The vanishing point detection is used to determine the region of interest of the road image.Removing the task-independent region,such as the sky,can reduce the search space and speed of the algorithm.Based on inverse perspective transformation,the road image is transformed into a top view,in which the lanes are parallel and equally wide,to restore the true characteristics of the lane,therefore can improve the algorithm detection performance.2.Two methods of lane detection based on deep neural network are designed.The first method extracts gray feature of lanes to obtain the candidate location,and then employ convolution neural network to classify lanes and non-lanes;the second method performs orientation and classification of lanes based on the R-FCN network.Six test datasets were set up,the experiment results show that the recall rate and the accuracy rate of lane detection based on R-FCN network are higher than the CNN method,with the average results of 97.1833%and 93.8317%.3.The connection,fitting and evaluation criteria of lane detection are studied.Aiming at the fitting of straight and non-straight lanes,a method based on angle estimation is proposed to connect lane candidate areas.A method based on gradient features and least square method is designed for lane fitting.Experimental results show that the proposed methods can adapt to different scenarios with an average error rate of 2.85%,an accuracy of 98.7715%and a horizontal offset of 3.1531 pixels.4.Vehicles,pedestrians and other targets may block the lane,leading to the degradation of lane detection performance.In order to realize the simultaneous detection of lane,vehicles and pedestrians,and therefore to improve the detection performance of those tasks,this paper transplants the inverse perspective transformation into R-FCN multitask network and realizes the end-to-end lane detection and further achieves the simultaneous detection of vehicles and lane.