Search

Thesis (M.S.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1991., Includes bibliographical references (leaves 57-60)., by Anselm Spoerri., M.S.

This thesis shows how to detect boundaries on the basis of motion information alone. The detection is performed in two stages: (i) the local estimation of motion discontinuities and of the visual flowsfield; (ii) the extraction of complete boundaries belonging to differently moving objects. For the first stage, three new methods are presented: the "Bimodality Tests,'' the "Bi-distribution Test,'' and the "Dynamic Occlusion Method.'' The second stage consists of applying the "Structural Saliency Method,'' by Sha'ashua and Ullman to extract complete and unique boundaries from the output of the first stage. The developed methods can successfully segment complex motion sequences.

This thesis addresses the problem of how to detect boundaries on the basis of motion information alone, and its solution is performed in two stages: (1) the local estimation of motion discontinuities and the computation of the visual flow field; (2) the extraction of complete boundaries belonging to differently moving objects. For the first stage, three new methods are presented that can independently estimate motion boundaries: the Bimodality Tests, the Bi- distribution Test, and the Dynamic Occlusion Method. These methods can estimate motion boundaries in a scene containing several moving objects, without prior knowledge of their shapes or motions, and they require only local computations. The motion boundary estimators have been implemented on the Connection Machine, a large parallel network of simple, locally interconnected processors. Further, it is also shown that the visual flow field can be locally estimated as a by- product of the early estimation of motion boundaries, and a mathematical formulation is provided to show that the proposed computation of visual motion is well-posed. The second stage consists of applying and modifying the Structural Saliency Method by Sha'ashua and Ullman to extract complete and unique boundaries from the output of the first stage, which is often broadly defined and can contain gaps. Results are presented that show that the methods can sucessfully segment complex dynamic images composed of random-dot patterns or natural textures. It is also shown how the methods can be used in stereopsis and surface reconstruction.