Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.597617
Title: Detection of occluding boundaries in spatiotemporal images
Author: Chiu, C. C.
Awarding Body: University of Cambridge
Current Institution: University of Cambridge
Date of Award: 1996
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
Abstract:
The extraction of occluding contours - projection of depth discontinuities - in unstructured scenes is one of the most important unresolved problems in computer vision research. Traditionally, occlusion is studied in static images in the context of T-junction analysis, object segmentation, or stereo. In this thesis, we motivate the use of deliberate but constrained viewer motion to make explicit the dynamic properties of occluding contours; namely, motion parallax and accretion/deletion of image texture. We examine long, densely-sampled image sequences in order to capture the spatiotemporal coherence of moving images, and thus eliminating the correspondence problem. Two methods of occlusion detection are proposed: the first is based on locally finding spatiotemporal junctions; the second is based on extracting long-range spatiotemporal trajectories and finding their intersections. Regarding the first method, we conjecture that occlusion detection is a local operation which can be achieved by spatiotemporal filters of finite support. We compute a coherence measure which quantifies how well the local spatiotemporal structure is aligned. Occlusion events are then signalled by a drop of the coherence value. In order to study the error performance of this operator, we conduct a Monte Carlo simulation using a large number of synthetic, but realistic, spatiotemporal images. The results show that our detector gives reasonable response in many cases, but the detection rate and localisation accuracy are not entirely satisfactory. The second method is concerned with fitting straight lines to spatiotemporal images and finding their intersections.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.597617  DOI: Not available
Share: