Building detection for monitoring of urban changes

Abstract

This thesis addresses the problem of monitoring urban changes by decomposing it to building and change detection. HOG features, in combination with several other discriminative features, such as NDVI, FAST and LBP features, are employed in a search for the development of a robust and accurate building detector. Furthermore, a novel cosine-based distance function is introduced for the computation of distances between the SVM feature vectors in order to suppress the sensitivity of the SVM classifier to the presence of noise and outliers. Moreover, the transformation of SVM scores to probabilities and the definition of a better threshold that differentiates positive from negative feature vectors are proposed. To allow the transition from the object-based building detection to the pixel-based building delineation, a set of novel region refinement processes that includes an unsupervised image segmentation technique and the construction of building candidates by employing the most probable to correspond to buildings image regions based on a novel scoring procedure is proposed. Taking advantage of the ability of the CNNs to automatically generate discriminative features, another approach to the problem of building detection involves the introduction of a Modular-CNN architecture. Two novel layers are proposed and added to the Modular-CNN architecture in order to improve its generalisation power and robustness. The change detection task is approached by a top-down approach that employs the computed building masks in order to identify building changes and a bottom-up approach that initially detects changes prior to their modelling using the computed building masks. In the change detection framework, we propose a novel change amplification algorithm that enhances the differences between the compared images in order to be more easily recognised and extracted. Finally, we propose a new robust change detector based on CNNs with the ability to automatically detect building changes, while discarding all other changes.