Probabilistic representation of the uncertainty of stereo-vision and application to obstacle detection

Stereo-vision is extensively used for intelligent vehicles, mainly for obstacle detection, as it provides a large amount of data. Many authors use it as a classical 3D sensor which provides a large tri-dimensional cloud of metric measurements, and apply methods usually designed for other sensors, such as clustering based on a distance. For stereo-vision, the measurement uncertainty is related to the range. For medium to long range, often necessary in the field of intelligent vehicles, this uncertainty has a significant impact, limiting the use of this kind of approaches. On the other hand, some authors consider stereo-vision more like a vision sensor and choose to directly work in the disparity space. This provides the ability to exploit the connectivity of the measurements, but roughly takes into consideration the actual size of the objects. In this paper, we propose a probabilistic representation of the specific uncertainty for stereo-vision, which takes advantage of both aspects - distance and disparity. The model is presented and then applied to obstacle detection, using the occupancy grid framework. For this purpose, a computationally-efficient implementation based on the u-disparity approach is given.