Duration of the Thesis: 6 months
Completion: April 2011
Tutor: Prof. Dr.-Ing. Norbert Haala
Examiner: Prof. Dr.-Ing. Norbert Haala
In recent years, 3D city reconstruction is one of the active researches in the field of photogrammetry. However, in this study, we focus on the extraction of building roofs, streets and vegetation. In order to reach this goal, the dense and high accurate 3D point clouds should be acquired. Image matching heralds the renaissance in the modern digital photogrammetry due to generating dense and accurate 3D point clouds that is comparable with the LIDAR data. Furthermore, it is more cost efficient and no need of further processing. 3D point clouds of this research were provided from pixel-wise image matching by Semi-Global Matching (SGM). As a result of image matching process, X-Y-Z image, valid image and disparity image are produced. In X-Y-Z image, each pixel corresponds to 3D coordinates of the visible objects and it is raw structured 3D point clouds. In valid image, valid or invalid image pixels are the results of successful or unsuccessful image matching process. In disparity image, pixel values correspond to parallax to stereo model. Furthermore, intensity values from the RGB or grayscale image as an additional information assists us in better interpretation of the aerial images. Moreover Digital Terrain Model (DTM) is utilized to distinguish ground and non-ground 3D point clouds.
A fundamental step to process and interpret the 3D point clouds is the segmentation. In this work, it is performed on the basis of surface growing-based approach. It is firstly carried out in object space based on similarity measures like direction of locally estimated surface normals and considering distances of the 3D point clouds to the best fitted plane. It starts from optimal seed points to create meaningful, disjoint and connected segments with homogenous property. Least squares plan fitting is applied to compute surface normals. Moreover, in order to improve the results of the estimated surface normals, robust estimation has been applied. In addition, we determined nearest neighboring search algorithm to find closest points to the query point to compute surface normals. Secondly, segmentation is followed in image space by using segmented regions from the object space as a seed image pixel and color or gray value from the RGB or grayscale image as additional information.
Vegetation is extracted by means of spectral information from the RGB image and the median filter and morphologic dilation are applied to eliminate the noises and fill the gaps between the extracted vegetation. Afterward, building roofs and streets are extracted according to the segmentation procedure in object space and image space.
Ultimately, results of segmentation in object space and image space with delineation of building roof boundaries are stored in the separate images with *.png extension. Moreover, extracted and segmented 3D point clouds that include building roofs, streets and vegetation are stored in the VRML format to display them in 3D.
The aim of this thesis is implementation of test-bed program to segment and classify 3D dense point clouds from aerial image matching and features of interests are building roofs, streets and vegetation. The implemented program was written in C++ and including OpenCV and the CGAL library as external libraries. 3D point clouds are imported to the program in the LAS (Log ASCII Standard) format. Program was examined with three types of point clouds with different level of density and accuracy.
|
| Surface normals (left), segmentation of surface normals in object space (right) |
Surface type |
Building roofs |
Flat terrain |
|---|---|---|
Standard deviation of validated points |
0.409 m |
0.473 m |
Mean Z shift of validated points |
0.0 |
0.0 |
Validation of hybrid point clouds
 |
| Segmentation in image space (left), surface boundaries using convex hull (right) |
 |
| Overlaying of segmentation results from the RGB and grayscale image, red color boundaries results from the grayscale image and blue color boundaries results from the RGB image |
 |
| Representation of building roofs boundaries with points (left), with polyline (right) using 2D alpha shapes |
 |
| Texturing of extracted building roofs (left), classification result (right) |
 |
| 3D visualization of segmented 3D point clouds in the VRML format |