Development of Algorithms for Automatic Object Segmentation and Recognition using 3-D Point Clouds
| Autor: | Nguyen Xuan Loc |
|---|---|
| Rok vydání: | 2012 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 100 In this dissertation we discuss a variety of 3-D image processing techniques that advance the state of the art in the fields of object segmentation and object recognition. The rapid development of technology has made powerful light detection and ranging devices become available in the market and the acquisition of 3-D range data has been more feasible and popular. High-resolution 3-D profile of objects, such as people, cars, trees, buildings and roads, can be captured, providing a great help for the tasks of object segmentation and recognition which are remaining as challenging research topics in computer vision. The first contribution of this research is the novel depth slicing technique for segmenting object with the real-time processing. The depth region containing objects can be initially segmented by employing depth slicing technique. For accurately marking object boundary, a region growing method is then applied through a recursive searching process. The second contribution is a new method for object segmentation employing region-growing algorithm and classification of surface characteristics. In order to solve the problem of digital background identification (DBI), the method proposes a novel criterion based on the distribution of normal surface vectors. According to this criterion, range data are classified into certain types of surface as an initial stage of evaluation for addressing all the points belonging to the background. By incorporating this criterion into the region-growing process, a robust range data segmentation algorithm capable of segmenting complex objects suffering huge amount of noises in outside condition is established. To detect accurately the object boundary, a recursive search process involving the region-growing algorithm for registering homogeneous surface regions is developed. The third contribution of this thesis aims at object recognition. The key breakthroughs for object recognition mainly lie in defining unique features that distinguish the similarity among various 3-D objects. In this research, the object recognition scheme is developed to identify targets underlining automated search in the range images using geometric constrains and curvature-based histogram. Since the accuracy of object recognition is generally limited by using a single viewpoint constraint of sensing device, the important feature of the proposed technique which can overcome this weakness is to employ a set of histograms from multiple views for representing the 3-D structure of objects. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|