City Research Online

Abstract

The work described in this thesis is concerned with the 3D interpretation of monocular images. First, the perspective transformation is interpreted in an image which enables the extraction of 3D information. Then, connectivity in the image is utilized in order to infer 3D groupings in the scene. The result of the process is 3D structures, leading to local 3D maps, the scales of which are unknown. The processing of several images from unknown viewpoints allows the relative scales of the various maps to be known. Thus, an unsealed but consistent 3D map is extracted. This map has a 3D symbolic representation and may be integrated in a CAD database.

A probabilistic approach is used for interpreting the image. First, two types of error are defined : errors due to the measurement uncertainty and errors due to accidents such as the proximity of unrelated features, called segmentation errors. Because of measurement errors, relations are not exactly fulfilled. Accounting for such errors is responsible for segmentation errors, and thereby unreliability of the process. The best trade-off for checking these relations is based on the maximum likelihood test. In order to determine this test, a precise statistical model of the data is defined. Moreover, accounting for measurement uncertainty leads to an original process for detecting the vanishing points in the image in a consistent way over the space.

Another central theme of this work is the 3D representation adopted for the structures extracted from the image. The intrinsic parameters of these structures are viewpoint and scale invariant and the geometric relationships and the degree of freedom of these structures are implicit in such a representation. This considerably eases the construction of the 3D structures, and then of the local and global maps.

The method is illustrated by the processing of images of indoor scenes of a power plant.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology > Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads