A neural architecture for 3D segmentation

Abstract

An original neural scheme for segmentation of range data is presented, which is part of a more general 3D vision system for robotic applications. The entire process relies on a neural architecture aimed to perform first order image irradiance analysis, that is local estimation of magnitude and orientation of the image irradiance gradient.In the case of dense 3D data, irradiance is replaced by depth information so irradiance analysis of these pseudo-images provides knowledge about the actual curvature of the acquired surfaces. In particular, boundaries and contours due to mutual occlusions can be detected very well while there are no false contours due to rapid changing in brightness or color. To this aim, after a noise reduction step, both magnitude and phase distributions of the gradient are analysed to perform complete contour detection, and all continuous surfaces are segmented.Theoretical foundations of the work are reported, along with the description of the architecture and the first experimental results.