Syntactic depth-segmentation of edge images: a representation scheme for the 3-D scene layout

A representation scheme for sparse depth maps which supports 3-D interpretation of unstructured scenes is developed. The depth map is a result of matching corner features in consecutive images with a relaxation method and incorporating available estimates of the motion parameters. An edge-connectivity digraph maps features into nodes with the corresponding depth attributes. The edge connectivity in the image domain is mapped into the digraph connectivity by a self-recursive edge-tracking that operates on intensity data without resorting to image segmentation. A depth-parametrized grammar has been created to segment the digraph into strings of connected nodes that share a similar depth attribute. The strings are an approximate representation of observable 3-D scene edges. The overall scheme is tested on blurred synthetic images containing occlusions, with good results. The representation is useful for augmenting coarse-resolution terrain databases during navigation in order to detect and avoid obstacles