A novel lane detection approach fusion by vehicle localization

This paper proposes a novel approach for lane detection fusion by visual and vehicle localization system. To estimate accurate camera´s gesture as well as road parameters, model-driven approach was introduced in our approach, which was the most popular way for lane detection so far. Be different from traditional processing way relying on visual recognition results, our approach develops a fusion method that was required for more reliable results under various environment conditions. Two processing engines are contained in our system. [1] Visual processing focuses on road segment detection on image. It is called near zone, in which road shape seems as linear and lane feature is easy to be extracted. Parameters: such as road width, vehicle lateral displacement, camera´s inclination angle and heading angle, are estimated in this area. [2] Since visual sensor is hard to locate road lane in far zone, which may introduce uncertain noise, vehicle localization with digital map is proposed to assist lane detection for restricting detection zone, where lane might be not seen clearly or block by other objects. Because parameter vector is predicted by our approach and its variance is limited in the rational range, model updating and parameters tracking will be more efficient and reliable by EKF in the fusion step.