Object edge contour localisation based on HexBinary feature matching

This paper addresses the issue of localising object edge contours in cluttered backgrounds to support robotics tasks such as grasping and manipulation and also to improve the potential perceptual capabilities of robot vision systems. Our approach is based on coarse-to-fine matching of a new recursively constructed hierarchical, dense, edge-localised descriptor, the HexBinary, based on the HexHog descriptor structure first proposed in [1]. Since Binary String image descriptors [2]-[5] require much lower computational resources, but provide similar or even better matching performance than Histogram of Orientated Gradient (HoG) descriptors, we have replaced the HoG base descriptor fields used in HexHog with Binary Strings generated from first and second order polar derivative approximations. The ALOI [6] dataset is used to evaluate the HexBinary descriptors which we demonstrate to achieve a superior performance to that of HexHoG [1] for pose refinement. The validation of our object contour localisation system shows promising results with correctly labelling ~ 86% of edgel positions and mis-labelling ~ 3%.