Obstacles detection from disparity properties in a particular stereo vision system configuration

Many approaches were proposed to detect obstacles in front of a car. Most of them start with modelling the road geometry and extract some coherent structures such as road edges or lane-markings in the stereo image pair. Our approach copes with this problem differently: At first, we give a definition of an obstacle and in a second step, we exploit its relationship with the stereo vision system configuration. This allows us to separate image features into those belonging to the obstacles and those due to non-obstacles (e.g. lane markings). Experimental results seem to demonstrate the effectiveness of such a method. Future work will incorporate this approach into a stop & go application.