Calibration and Sensitivity Analysis of a Stereo Vision-Based Driver Assistance System

As safety and efficiency issues of transportation hand-in-hand with the intelligent vehicle concept – have gathered ground in the last few years in the automotive research community and have penetrated into the automotive industry, vision-based applications have become increasingly important in driver assistance systems (Kastrinaki et al., 2003; Bertozzi et al., 2002). The primary targets of the safety and efficiency improvements are intelligent cruise control (e.g. vehicle following), lane keeping and lane departure warning systems, assistance in lane changing, avoidance of collision against vehicles, obstacles and pedestrians, vision enhancement and traffic sign recognition and signalling. Our focus of interest here is stereo machine vision used in the context of lane departure warning and lane keeping assistance. The primary purposes of our vision system are to determine the vehicle's position and orientation within the current lane, and the shape of the visible portion of the actual lane on structured roads and highways. That is, we face a somewhat simplified simultaneous localization and mapping (SLAM) problem here, with the assumption of a structured man-built environment and a limited mapping requirement. The localization is achieved through the 3D reconstruction of the lane's geometry from the acquired images.