Research On Lane Detection Algorithm Based On Image Segmentation And Inverse Perspective Transformation

The lane detection is not only a vital part of the Advanced Driving Assistance System,but also the foundation of intelligent driving.It provides important guarantees for the safe driving of vehicles.Due to the complex and changeable road conditions,many traditional lane detection algorithms are susceptible to light intensity,road environment interference,lane line quality,etc.The accuracy and robustness of the algorithm detection need to be further improved.In recent years,the rapid development of deep learning technology has brought new technical support for lane detection,but the domestic research has started late.Many algorithms today are based on foreign data sets and cannot be well matched with domestic urban roads.In addition,there are also problems in that the model is difficult to train and has poor real-time performance.In this paper,different lane detection ideas are adopted,which are divided into two modules of road image binary segmentation and post-processing,to study and optimize traditional algorithms and algorithms based on deep learning.The main research contents of the paper are as follows:1)Research the traditional road image binary segmentation algorithm.Improve the image processing part,including image graying,filter denoising and edge detection,etc.,put forward a method of combining multiple color space channel information,normalize processing and fuse image information,and obtain more complete binary information of lane line segmentation map.Using the CUDA parallel computing architecture optimization algorithm and GPU parallel computing to process images,the system operation is accelerated by 3.26 times.2)Research on road image binary segmentation algorithm based on deep learning semantic segmentation technology.A new V-Net network is designed based on the UNet network structure,and V-Net is improved by referring to the design ideas of VGG,Res Net and other network structures,including network feature extraction methods,residual design,Skip connection and other structures,and design loss functions and parameter optimization methods,select the domestic lane line data set for model training.Test the effect of V-Net network binary segmentation.Compared with U-Net and Deeplab V3 + networks,it performs better in accuracy and real-time.3)Research the lane line post-processing algorithm based on inverse perspective transformation.Perform camera calibration research and inverse perspective transform processing of binary segmented images,use sliding window method to scan lanes,use quadratic polynomial based on least squares to fit lane line curve,and use Kalman filter algorithm to track lane line.Test the lane line detection algorithm,in the highway and urban road scenes,it can correctly fit the lane line straight line and curve,adapt to complex lighting and road environment,and meet the requirements of real-time detection and robustness.