A new lane detection system for unmanned vehicles

Abstract

As the intelligent transportation technologies have rapidly been improving, it is plausible to state that intelligent ground vehicles will be on the roads driving through the tra c in the near future. Moreover, intelligent vehicles will have a revolutionary e ect in transportation industry, while eliminating accidents, and reducing the emission of carbon gases. Not only presently multiple intelligent vehicle projects are in progress in several countries but also the research and development departments of automobile manufacturers work on the tra c safety applications as well as autonomous navigation for future autonomous transportation. This study focuses on one of the most important tra c safety applications which is also a crucial part of a perception subsystem of an autonomous ground vehicle, that is, Lane Detection. Main purpose of lane detection is to estimate the position of the road lane marks with a camera in order to calculate the distance to the lane and relative direction of the vehicle. Vision-based lane detection systems try to overcome the challenges of weather, lighting, road conditions to detect the lane marks. Within this thesis, we have proposed a novel lane detection method which produces successful results under challenging conditions. The method has three important phases; lane feature extraction, lane selection and road model tting. The lane feature extraction method in our work is similar to the symmetrical local threshold. However, it is more powerful to detect eroded lane marks and road borders. In order to estimate the lane on the left and right side of the vehicle, Random Sample Consensus (RANSAC) algorithm ts a line with an error variance for lanes on both sides. Using the lane pixels on the estimated lines, a hyperbola-pair model is generated and the parameters of the road and the vehicle is calculated for the navigation subsystem. Results are presented on a reference image database Road Marking database (ROMA) and evaluated with Receiver Operating Characteristic (ROC) and Dice Similarity Coe cient (DSC) curves. Synthetic images are also used to evaluate the performance of the lane detection technique.