City Research Online

Abstract

The visual search process is required when locating an object in some region of space. To perform this search two capabilities must be available: the ability to recognise the object when it comes into view; and a way of selecting these views. Visual search is often complicated by object occlusion and low spatial resolutions of the object. Although the human visual system performs this task effortlessly, the mechanisms of it are not properly understood. Object colour and geometry, however do play an important role. This thesis develops an object search methodology which assumes that a computer vision system captures both wide-angle and zoomed images of the scene containing the object. Since most of the research has focused on object recognition using geometry, this system is purely colour-based. It is not expected that object colour will always give a definitive solution, however database pruning will often occur leading to reduced search times.

The thesis argues that because colour is salient and more resilient than geometry to decreases in spatial resolution, it is more appropriate for visual search when the object occupies a small spatial resolution in an image with a large field of view. It also demonstrates that colour can be used to recognise objects when they occupy most of the field of view; as well as discriminate between database models with similar colour proportions but different region topologies. These conclusions are supported by the results produced by three algorithms, two of which perform colour object search and one that performs colour object recognition.

The first object search algorithm uses image locations containing salient object colours as a method of selecting views. Each of these views are ranked indicating which view most likely contains the object. The second object search algorithm identifies image regions with similar colour and topology as the object. These results are produced in a best-first order. The object recognition algorithm uses an invariant based on region area to identify three corresponding model and image regions. A transformation is calculated to bring the model and object into the same viewpoint where region matches are based on position and colour.

Each of these methods produced good results in complex indoor scenes with fluorescence and/or tungsten filament lighting; also the search speeds were impressive.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics T Technology > TK Electrical engineering. Electronics Nuclear engineering
Departments:	School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses School of Science & Technology > Engineering

Export

Downloads