Efficient parallel processing for depth calculation using stereo

Stereo vision generates the depth map of a scene by fusing information in images taken from two or more views. Stereo is a computationally intensive task and an efficient real time stereo application needs either a dedicated hardware or a parallel computer. This paper proposes a parallel stereo algorithm developed at CAIR using both edges and regions as features for matching. The algorithm employs an intelligent stereo matching technique that reduces the search space for correspondence thereby reducing the computational load. A load balancing scheme is also proposed for further improving the efficiency of the algorithm, with the increase in the number of nodal processors. The algorithm is implemented on a PACE parallel processor developed at the DRDO laboratory ANURAG at Hyderabad, India, and its capability is demonstrated through an application example.