Back to search

National Technical University of Athens (NTUA)

Aνίχνευση και εντοπισμός μεταβολών κτισμάτων από εικόνες και νέφη σημείων

Abstract

dc:description

The technological development in the fields of image processing, computer vision and digital photogrammetry and remote sensing, provides new tools and automated solutions for applications in urban studies, cadastre, etc, associated with urban development, identification of illegal constructions, 3D modeling, change detection, etc. Apart from the traditional techniques that utilize purely aerial or satellite imagery, recent approaches exploit 3D point clouds captured from aerial laser scanning systems (LIDAR) or extracted from Dense Image Matching (DIM) methods. Building extraction, change detection and roof segmentation from images and point clouds are topics of high research interest as they are useful for urban city monitoring. These topics still remain challenging due to the increasing demands for accuracy, rapidity and economy. The present Phd thesis provides an integrated, extensive and detailed methodology for building extraction change detection and roof segmentation using images and point clouds and applying sophisticated methods. Several complex urban regions of various types of buildings, pixel resolutions and types of data are examined. To verify the utility and functionality of the proposed methodology, the corresponding 3D models of the buildings are extracted.Initially, the differences between the LIDAR and DIM point clouds were highlighted in terms of robustness and efficiency through direct point cloud comparison as well as by using the corresponding orthoimages. The results shown that the LIDAR point clouds not only describe with fidelity the objects of the scene but also present a global homogenous behavior either at urban or sub-urban areas. However, LIDAR point clouds suffer in some cases from noise and local under-sampling indicating local lack of spatial information at building’s boundaries. On the other side, the DIM is a low cost and flexible solution with promising results. The DIM point clouds and the corresponding orthoimages achieved satisfying results at sub-urban areas. However, weak in urban areas with intense relief or depth discontinuities indicating excessive interpolations and radiometric lesions at building’s boundaries due to occlusions or possible mismatches. The aforementioned issues, both for LIDAR and DIM point clouds, affect the building extraction, change detection and roof segmentation results.Dependent on the data source employed, building extraction, change detection and roof segmentation techniques can be classified into three groups: i) the ones that use radiometric information (airborne or satellite imagery data), ii) the ones that exploit height information (LIDAR or DIM point clouds), and iii) those that combine both of data sources. However, the limitations of using information from multi-modal sources (e.g., LIDAR and imagery data) are the additional cost of acquisition and processing and the co-registration related issues. For this reason, the present doctoral dissertation focuses on the combination of data that are extracted from one sensor, i.e., either from camera devices (aerial images combined with the corresponding DIM point clouds) or from laser scanning systems (LIDAR point clouds combined with the corresponding proper geometric/morphological features). The proposed methodology achieves similar building extraction results with other approaches that utilize multi-modal sources. Also, the building extraction results extracted from the aerial images combined with the corresponding DIM point clouds were similar with those extracted from the LIDAR point clouds combined with the corresponding proper geometric/morphological features. Usually, the approaches of building extraction and change detection are discriminated into the ones that apply supervised machine learning and those that use model-based methods. In this research, both approaches are performed. More specifically, a deep machine learning scheme through CNNs (Convolutional Neural Networks) was developed as well as a new model-based scan line smooth filtering method was proposed. The extracted building detection results from the supervised machine learning and model-based methods can be considered as similar. This was verified by comparing the CNN with other model-based methods at the ISPRS (International Society for Photogrammetry and Remote Sensing) benchmark dataset. Also, in this research, the various factors that indicate the appropriateness of each of the two approaches are discussed. Additionally, the CNN was compared with other shallow and non-linear machine learning schemes such SVMs (Support Vector Machines) και ANN (Artificial Neural Network). The results indicate the clear predominance of the CNN. However, the CNN weak in computational complexity compared to the SVM classifiers. The use of a better CPU and more RAM capacity as well as the implementation of methods that accelerate the learning phase can significantly speed the processing. Concerning the roof segmentation from point clouds, the mostly used data driven plane detection techniques are region growing, RANSAC and Hough methods. The SHT (3D Standard Hough Transform) and its variation, the RHT (3D Randomized Hough Transform), are the most popular Hough methods. In this research, a new accumulator design and novel extensions of the RHT are proposed associated with: 1) additional constraint criteria during the random selection of the 3 points, 2) an additional normal tolerance regarding the detected plane, 3) automatic selection of sub-regions surround adaptive center points, and 4) automatic descent tuning process of parameters. Thus, five variations of the RHT were developed namely iRHT, ΕRHT, eΕRHT, APRHT-C and APRHT-R. A sensitivity analysis of each variation was carried out through extensive experiments using simulated data as well as real life LIDAR and DIM point clouds of several types of buildings. The iRHT led to under-segmented, spurious and over-segmented planes even for typical cases of buildings. The eERHT, APRHT-C and APRHT-R achieved better results compared to the ERHT at complex buildings and building structures of special architecture. The eERHT, APRHT-C and APRHT-R achieved similar results, however, the computational time of the eERHT was significantly higher compared to the two APRHTs. Nevertheless, the ERHT achieves satisfying results for typical cases of buildings after a post processing strategy absorbing spurious and over-segmented planes. The comparison with RANSAC indicated a clear predominance of the ERHT. However, under-segmented planes and local distortions at the boundaries of small planes occurred, which is challenge for further improvement of the refinement process. Connectivity criteria or strict tuning of the APRHT-C and APRHT-R may overcome the under-segmentation problem. The 3D models of the buildings extracted via the ERHT and APRHT-C were qualitative evaluated illustrating their potential for 3D modeling in LoD 1 and LoD 2.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2019

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Maltezos, Evangelos
  • Μαλτέζος, Ευάγγελος

Subjects

dc:subject × 20

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/46588

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Maltezos, Evangelos; Μαλτέζος, Ευάγγελος. Aνίχνευση και εντοπισμός μεταβολών κτισμάτων από εικόνες και νέφη σημείων. National Technical University of Athens (NTUA), 2019. http://hdl.handle.net/10442/hedi/46588