Window detection in 2D and 3D scenes for geodetic frameworks

Abstract

The goal of this thesis work is to address the classical problem of windows object detection in 2D and 3D scenes. To solve this problem two novel algorithms Gentle Adaboost and RANdom SAmple Consensus (RANSAC) applications of windows detection are proposed.<br />First, a learning based methodology is proposed for the development of a window detection system in 2D scenes. In particular, a novel Gentle Adaboost driven supervised cascaded decision tree has been learned using Haar-like features and their respective rotated versions. This new approach achieves less than 10% false detections for Single Window (SW) and Window Region Of Interest (WROI) based Evaluation schemes on the available benchmarking building datasets (TSG-20, TSG-60 and ZuBuD).<br />Second, an image-based window detection application is proposed in 3D laser scanner data. The facade segmentation is performed by using a novel RANSAC plane fitting algorithm. A new statistical analysis of distances has been introduced for the selection of window candidates pixels in spherical coordinate laser distance image as well as ortho distance images of segmented 3D facades images. The approach has been evaluated on the CityScanner dataset and reaches a detection accuracy of more than 80%.<br />Finally, these two proposed approaches by supervised learning and by segmentation are integrated in an optical 3D measurement system. The main goal of its development is the automation of the whole measurement process, including the tasks of point identification and measurement, deformation analysis and interpretation.