A MonoSLAM approach to lane departure warning system

Abstract

Lane Departure Warning (LDW) systems are one of the widely researched topics under Advanced Driver Assistance Systems (ADAS), because they are seen as the most viable way to prevent the traffic accidents caused by involuntary lane departures from happening. Various methods and algorithms used for lane tracking to be used in LDW in the literature; however, most of them only track the lanes or the position of the vehicle inside the lane. This article introduces MonoSLAM based method for LDW design, assuming that the camera is moving in a previously unknown scene. While applying this method, a constant lateral velocity model for the vehicle is used, which assumes that the vehicle is exposed to undetermined Gaussian lateral accelerations. As the first output, the localization of the vehicle on the road is achieved. Moreover, the method is applied with a low cost webcam attached on a vehicle. Five control points for each lane is used to track the lanes and these control points are modelled as if they have a constant position. Detection is made with steerable filters exploiting the state covariance from EKF to make detection more robust. In addition to this, off-line experimental results are given for 200 frames. Results of lane slope on image plane compared with ground truth marked manually for performance benchmarking and localization estimation of a scenario similar to loop closure test is given.