Jens Goldbeck

LANE FOLLOWING COMBINING VISION AND DGPS

Abstract A real-time system focusing on lane sensing for autonomous vehicle guidance is described which runs on commercial hardware. For safe and comfortable vehicle control even on winding roads with bumpy surfaces, the ego-state of the vehicle as well as lane geometry needs to be known. This information is needed with high precision and reliability. For the sake of safety, two redundant sensor systems are employed. The first is a video camera that tracks visual lane boundaries such as white or yellow road markers. The second is a high precision location system using DGPS (Differential Global Positioning System) combined with an INS (Inertial Navigation System) for ego-state recognition. The achieved information is related to a high precision digital map.

Lane detection and tracking by video sensors

Describes a video sensor system for vehicle environment sensing. The system comprises a high dynamic range CMOS camera, dedicated image evaluation circuitry and image processing software for capturing vehicle environment data. The main part of the paper is concerned with the description of a real-time lane detection and tracking algorithm which is able to determine the course of the road ahead of the vehicle together with the position of the vehicle relative to the road with high accuracy. The key feature of the presented system is its ability to adapt to nearly all weather conditions and road types without any external adjustments together with its self-assessment capabilities, both of which are crucial for automotive applications.

Camera calibration for lane and obstacle detection

This paper deals with camera calibration for use in guided vehicles. Camera calibration yields parameters which describe the relationship between 2D images and 3D environment. An intrinsic camera calibration is performed and the extrinsic parameters are determined using a weighted least squares technique within a robust, iterative estimator. They allow a correct 3D reconstruction from the given images. This information is indispensable for lane and obstacle detection with the goal of driving an autonomous, unsupervised vehicle. The impacts of uncertainty in the calibration parameters on lane recognition and obstacle detection are quantified. It is shown that reliable results of lane recognition and object detection can only be performed with calibration parameters of high precision.