A new real-time method for finding temporary and permanent road marking and its applications

Abstract

In this paper, a new real-time method for finding temporary and permanent road marking is proposed. The method is based on the geometrized histograms method for segmenting and describing color images. This method is able to deal with both rectilinear and curvilinear marking, as well as with color temporary and permanent road marking. It also makes it possible to distinguish temporary road marking from white permanent road marking. The developed method is stable under illumination and is able to work even for partially disappearing road marking, typical for late winter and early spring. In contrast to many other methods, this method does not require any information about camera parameters and calibration and is able to find road marking in images taken under unknown conditions. The proposed method has been implemented by a program written in C++, operating under Windows and Linux. The program operation has been tested on video records shot on typical Russian roads during different seasons and under diverse whether and illumination conditions. The processing speed is about 20 fps for a standard modern computer. Using parallel computing, this speed is reduced considerably. The results of program operation are presented and discussed. The developed program is a part of the computer vision component of the control system of the AvtoNiva pilotless vehicle. IntroductionThe problem of finding road markings is very important due to applications in advanced driver assistance systems (ADAS) and vision systems of autonomous vehicles. Efficient algorithms for finding road marking implemented in ADAS can prevent many road accidents and save many lives. The first efficient methods for real-time lane (road marking) detection were proposed and implemented by E.D. Dickmanns and his colleagues [1, 2]. These methods were dated back to the late 1980s and employed the so-called 4D approach proposed by E.D. Dickmanns [1]. Many methods based on Hough transform, RANSAC algorithm, and learning approaches, including deep learning of convolutional networks, have been proposed (see [3][4][5][6][7] and the references in these papers) since then. The most recent review of the results obtained in this area can be found in [6]. In spite of the serious progress in investigations on this topic, vital problems connected with lane marking, disappearing in early spring and late winter, wet places and occlusion still take place. It deserves special mentioning that temporary lane markings may have extremely high curvature, which may pose additional problems to the existing methods. Despite of the impressing results reached by application of convolutional networks, there are certain problems that arise in using them, such as overfitting, a large