|Ph.D Student||Croitoru Arie|
|Subject||Improvement of Updating Procedures in Geographic|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Emeritus Yerach Doytsher|
A comprehensive automated accuracy-driven updating process is a challenge yet to be met. This work faces two aspects of this challenge: efficient up-to-date feature extraction, and data accuracy estimation. Up-to-date feature extraction was realized by a two phase process. In the first step (a top-down approach) buildings are hypothesized in image space using a corner-based pose clustering algorithm that was developed in this work. The main advantage of pose clustering is its lower complexity and the ability to employ an occupancy model, through which the detection threshold and the expected false alarm rate could be estimated. In the second step (bottom-up approach), each of the hypothesized buildings is verified by an attempt to extract the polygon describing the building rooftop. As right-angle buildings were addressed in this work, two perpendicular sets of parallel lines are extracted from the image. These line sets form a Right-Angle Skeleton (RAS), which is transformed into a Right-Angle Graph (RAG) by an inhabitation process of right-angel corners. The RAG is then used for guiding a grouping process, in which a sequence of corners is searched. This approach significantly reduces the number of corners that should be considered. The underlying principle of the proposed extraction scheme is that new residential areas exhibit considerable regularity in the shapes of the buildings (regularized areas). This diminishes the need for more general and highly complex building extraction schemes.
Data accuracy estimation was realized in this work by a modified Geostatistical paradigm. This paradigm is based on two tools: an estimator and a structure function. As an alternative to the commonly used Kriging estimator, this work explores the Least Squares Collocation (LSC) technique as an alternative interpolation scheme. The performance of LSC was examined both quantitatively and qualitatively under varying conditions of data content and structure function uncertainty. Additionally, computational tools for predicting the effect of structure function errors on the LSC results were developed and tested. In conjunction with the LSC estimator, two structure function estimation tools were developed in this work. The first tool consists of a quantitative data-driven criterion for the number of bins that should be used in the structure function estimation process. This criterion is based on an iterative white-noise test of the Covariogram spectral density function. A second tool that was developed is the DCT-based non-parametric covariance function. Current non-parametric algorithms employ an FFT-based approach, which requires eliminating complex coefficients. The new approach avoids the complex coefficients problem while ensuring positive definiteness.
This dissertation outline the theoretical as well as the practical aspects of these issues. Each of the proposed tools and algorithms is detailed and its performance is evaluated both on simulated and on real-world data. Conclusions and future work are indicated as well.