Abstract:
Stereo vision is currently used in the car industry as a tool for designing driver- assistance systems. However, limitations inherently due to the discrete nature of disparities observed by a stereo-vision system have not yet been modelled and analysed so far; this thesis aims at closing this gap. Stereo-vision results are used in driver-assistance systems for estimating trajectories or just speed. Besides accuracy limitations in stereo matching, the discrete nature of disparities also de nes limitations to detected trajectories or speed. This thesis proposes and discusses a novel tool for a safety engineer which permits the safety of these driver assistance systems to be estimated. It is based on a model which considers the true error in measured velocities of objects. Outputs from this tool show that the choice of stereo-system parameters, so as to optimally place the disparity change boundaries, is critical to the e ectiveness of such a system. As soon as the possibly colliding object crosses one of these boundaries, the range of possible trajectories for a (possibly colliding) object reduces signi cantly. This factor also means that larger objects (e.g. trucks) are slightly better tracked by stereo vision than smaller ones (e.g. signs or pedestrians). Completely safe stereo-based systems are also shown to issue many precautionary (and ultimately unnecessary) warnings if the stereo parameters are not chosen carefully. Keywords: Stereo analysis, accuracy, driver assistance, object tracking, Kalman lter