Road Segmentation And Lane Detection Based On Normal Map

With the rapid development of science and technology and increasing improvement of the national economy,the number of automobile in the world has increased sharply.The road traffic accidents and traffic congestion caused by it have become more and more serious.The existing traffic pattern cannot meet people’s needs.Intelligence and safety have become an inevitable trend in the future development of the automobile industry.In order to improve the traffic situation,automatic driving technology has emerged as the times require.As the basis and key technology of automatic driving,lane-mark detection is of great concern to lane departure warning,lane maintenance and path planning.Therefore,the research on lane detection has great application value and significance.Aiming at the problem that the existing lane detection algorithms have weak robustness in the scene of dense traffic and complex illumination,a method for road segmentation and lane detection based on a normal map is proposed.The lane detection algorithm of this thesis can be divided into three stages:road segmentation,lane feature enhancement and identification of the lane.In the first stage,firstly,the depth information is calculated by using the left and right lane images obtained by the binocular camera.Then,normal map inferred from the stereo image pair is utilized to segment the road based on the knowledge that the pixels of pavement have the same normal vector as they are in the same plane.In the second stage,edge enhancement,threshold segmentation and sub-area denoising are performed in the extracted road area to enhance lane features and reduce noise interference.In the third stage,we combine the Hough transform and the vanishing point to determine the position of the lane line in the initial image frame.For the detection of lanes in the subsequent frames,the inter-frame relationship is used to determine the initial region of interest(ROID and the lane edge points are searched in the region of interest.Then,the least squares method is used to fit the lane edge points,and the position of the subsequent ROIs are linearly predicted to track the lane lines.Compared with existing lane detection algorithms,the main contributions of this thesis are as follows:1.The road area is extracted by the road segmentation algorithm based on the normal map,which can effectively eliminate the interference of vehicles and buildings,and improve the robustness of lane detection in the case of dense traffic.2.By comparing and analyzing the current widely used threshold segmentation algorithms through experiments,this thesis improves the local threshold segmentation method,and determines the "optimal" configuration parameters through a large number of experiments,so that it can better adapt to complex illumination conditions.Then,the sub-region denoising processing is adopted to retain the lane lines more complete and pure,and lays a good foundation for the subsequent lane line detection.3.The combination of Hough transform and vanishing points can accurately obtain the position of the initial lane line.The method of linear prediction of discrete ROI is used to detect subsequent frames,which overcomes the shortcomings of fixed ROI of existing detection algorithms and can detect straight and curved lanes flexibly.In order to verify the robustness of the proposed algorithm.four state-of-the-art lane detection methods are compared on two databases.Experimental results indicate that the proposed algorithm works robustly and accurately under various challenging situations,especially in the scenes of complex illumination and dense traffic,which is more credible than the detection results of existing algorithms.