Qualitative recognition using Bayesian reasoning》.pdf

Pattern Recognition in Practice IV E.S. Gelsema and L.N. Kanal © 1994 Elsevier Science B.V. All rights reserved. 255 Qualitative recognition using Bayesian reasoning* Jianming Liang \ Henrik I Christensen, Finn V. Jensen Institute of Electronic Systems, Aalborg University, Fr. Bajers Vej 7, DK-9220 Aalborg East 1. INTRODUCTIO N Object recognition has been an active area of research for more than 3 decades . Most of the research has been based on recovery of structural descriptions with associated metric features. The obtained description is typically matched against models of a-priori defined objects; a good description of such an approach has been provided by Crimson [6]. For generic object recognition the above mentioned approach has several problems . Tradi­ tionally the approach is based on low-level features which results in an indexing problem as individual primitives have a limited indexing power. It is necessary to adopt abstract primitives to circumvent thi s problem . Use of explicit geometry is another problem, as it does not necessarily capture the overall geometric structure . To enable categorisation of objects Biederman has proposed a set of volumetric primi­ tives that may enable recognition of man-made objects [2]. The proposed primitives are all variations of generalised cylinders where cross-section, axis, symmetry and sweeping rule are assigned qualitative characteristics to provide a total of 36 primitives, termed GEONS {GEometnc lONS). Several researchers have tried to use these ideas from