Object recognition via attributed graph matching

摘要

A complete machine vision system implemented at Monash University is described, covering all the stages beginning with data acquisition, segmentation, modelling, through to matching. The process provides a near complete scene description, including the identity, location and pose of objects in it. This information is clearly suitable for the support of subsequent intelligent robotic manipulation in that domain. In this paper particular emphasis is made on the segmentation, modelling and matching stages. Attributed graphs are used to describe both objects in the scene and the model database. A sub-graph isomorphic approach is used to generate matching hypotheses. The hypotheses are confIrmed or rejected by examining the residual of an affine transformation between the objects in the scene and proposed models. The system is capable ofhandling multiple objects with occlusion. Preliminary results are presented. 1