Stereo vision in driver assistance systems

The paper is concerned with stereo vision applied to driver assistance. Prerequisites for transfer of recent results in vehicle guidance research to driver assistance functions are discussed. For the sake of a simple, low cost hardware a single processor and a stereo high dynamic range camera are employed achieving maximum dynamic and maximum robustness of 3D scene analysis. Self-assessment of the vision sensor is emphasized as a prominent demand for realization of driver assistance functions in the automotive market. For this, two measures are required besides an estimate itself, namely a confidence and a goodness-of-model-fit measure. The concept of self-assessing image analysis is illustrated using the example of road surface estimation. Preliminary results for obstacle recognition demonstrate good performance.