Robust Obstacle Detection Based on Dense Disparity Maps

Abstract

Obstacle detection is an important component for many autonomous vehicle navigation systems. Several methods have been proposed using various active sensors such as radar, sonar and laser range finders. Vision based techniques have the advantage of relatively low cost and provide a large amount of information about the environment around an intelligent vehicle. This paper deals with the development of an accurate and efficient vision based obstacle detection method that relies on dense disparity estimation between a pair of stereo images. Firstly, the problem of disparity estimation is formulated as that of minimizing a quadratic objective function under various convex constraints arising from prior knowledge. Then, the resulting convex optimization problem is solved via a parallel block iterative algorithm which can be efficiently implemented on parallel computing architectures. Finally, we detect obstacles from the computed depth map by performing an object segmentation based on a surface orientation criterion.