3D object detection and representation in remote sensing: Probabilistic methods and applications

Nowadays, satellite images and three dimensional data are actively used in various areas. The most important of these is the detection of objects after a natural disaster using satellite images or three dimensional data. In fact, this information is also valuable for government agencies, city and regional planners even if no natural disaster occurs. Turkey has its own remote sensing satellites in the orbit. There are also plans to launch new and advanced remote sensing satellites in the near future. Although we have our own remote sensing satellites in the orbit, these can only provide raw images. Either an operator should extract information from them or the information may be extracted by software automatically. The first option is not applicable most of the times since the size of the raw images are huge. Also, the objects to be detected from them are tiny compared to the image size. In this thesis, novel methods for object detection and segmentation are proposed. Remote sensing objects are detected using combinations of local features and shapes in a novel probabilistic voting framework. The shape of the objects are extracted using satellite images and height data. First, we developed a novel back-projection method to obtain the shape of detected objects in satellite images. Then for height data, two novel segmentation and filtering methods are proposed. The first method depends on the probabilistic voting method with a novel morphological based region growing algorithm. The second method uses empirical mode decomposition (EMD) algorithm for filtering and segmenting DSM into ground and non-ground points. The proposed methods are tested on different satellite images (IKONOS, WorldView, QuickBird,) and three dimensional data (DSM, LIDAR). Compared with the methods in the literature, better results have been obtained.