Depth map computing from three dimensional voxel data for multiple cameras systems

Common methods to derive depth maps from multiple cameras systems are based on the popular energy minimization or dynamic programming framework; unfortunately, this technique can only handle images from systems with short baselines, i.e. small angular offsets between the optical axes. In this paper, an efficient method based on three dimensional voxels is proposed to generate multiple view depth maps and synthesize the depth scene of arbitrary viewpoint even if the wide-baseline system is used. Space carving methods using two dimensional edges sampling are proposed to reconstruct people in a scene by pre-calibrated multiple cameras; they offer reduced computing costs. And then gray-valued mathematical morphology, which can robustly eliminate banded noise, is applied to smooth the depth maps. At the same time, one method based on the Rodrigues equation is presented for interpolating rotation; it can smoothly calculate the location of virtual cameras from the given camera parameters. In experiments, the efficient results of depth maps synthesized with different number of cameras are provided by using the depth maps. Copyright © 2011 Binary Information Press.