Deep Learning Based Lane Detection Technology

Autonomous driving can reduce traffic accidents and improve traveling safety and convenience.One of the most challenging tasks for autonomous driving is traffic scene understanding,which includes computer vision tasks,such as lane line detection and semantic segmentation.Lane line detection can guide vehicles to run in the correct area,provide a basis for lane maintenance.In addition,it can make driving more safely by providing early warning when the vehicle departs from the lane.In view of the complexity and diversity of driving scenarios,how to quickly and robustly detect lane lines in complex and diverse driving scenarios is difficult in lane detection.Lane lines are composed of solid lines,dashed lines,lane convergence lines,and separation lines.Its color includes yellow and white.Thus,Lane line features are relatively short.There are no complicated features to use when lane line detection.so it is difficult to develop a single model to detect all these lane lines.In addition,the characteristics of the lane line will be unclear due to road wear and obstruction of surrounding buildings and vehicles.Furthermore,light,weather,and road conditions will greatly affect the detection of lane lines.Lane line detection has high requirements for real-time performance,so shortening the delay is also very important.In view of the above problems,this thesis motivated by the following two aspects:On one hand,due to the influence of buildings,tree shadows,and other vehicle occlusions,the characteristics of single-frame lane lines in urban scenes are often fuzzy.Aiming at this problem,an urban environment lane line detection algorithm is proposed according to the characteristics of lane lines,including:(1)Design based on Convolutional Neural Networks(CNN)Encoder-decoder structure to extract features from adjacent frames;(2)The features obtained by CNN are input into the recurrent neural network(RNN),making full use of the strong correlation between adjacent frames of lane lines.In particular,it improves the success rate of lane line detection in complex scenes.(3)Introducing deep separable convolutions,reducing parameters and increasing training speed.It turns out that the algorithm is helpful to solve the problem of poor lane detection ability in complex scenes.On the other hand,vehicle driving is mainly divided into straight driving and lane changing,and the success rate of lane detection of vehicles driving straight is particularly important for lane keeping.lane lines cannot be detected every frame in complex scenes.In response to this problem,a method for tracking lane lines using a Kalman filter is proposed.When the number of lane lines continuously detected is greater than a threshold,it can be considered that the lane lines have been successfully detected,and then lane departure detection is performed without leaving the lane.In this case,the predicted value of the Kalman filter is used instead of the detected value.When leaving the lane,the lane line is re-detected.When the vehicle does not deviate from the lane,this method can quickly and accurately track the lane line.When the parameters of the lane line change continuously for several frames,CRNN is used to detect the lane line.