الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 121 |  |  |
| | | from | 121 | | |
Abstract
The goal of this thesis work is to address the problem of Vision-based Simultaneous Localization and Mapping (SLAM) and to propose new algorithms to achieve accurate results with low computational complexity.
In this thesis, a novel framework for vision-based Simultaneous Localization and Mapping (SLAM) is presented which focuses on the class of indoor mobile robots using only a monocular camera to achieve multi-level topological map building and robot localization concurrently.
Local and global features are combined in the same topological framework. The Scale Invariant Feature Transform (SIFT) is a local feature extractor that is used to build up the global level based on the fact that the local features provide a high level estimation for the robot position (e.g. in which room/location the robot is).
Topological mapping is used and proposed to be decomposed into sub-maps, in which the horizontal, vertical and diagonal details of the 4th level of the 2D haar discrete wavelet transform is used to build up the local level based on the fact that global image signatures provide a low level estimation for the robot’s position (e.g. in which view the robot is). Some modi cations are applied to enhance the output of the global features, in which, the hough transform is applied on the wavelet details to extract the shape histogram, in addition to the color layout descriptor (CLD). the shape histogram consists of three bins (number of vertical lines, number of horizontal lines and number of circles).
The output topological map is validated with the ground truth of the environment.
The results were evaluated according to the confusion matrix based on the classi ed images in the right and left datasets, in which Accuracy AC, Recall TP and Precision P were calculated. The proposed system succeeded in achieving an overall matching of 92% and an overall retrieval accuracy of 97%. Moreover, the number of decomposition
levels of the wavelet transform is analyzed to conclude that the relation between level of wavelet decomposition and number of decomposed views is inversely proportional. As a result, Level 4 is chosen for decomposition because the trade-o between a compact representation and a reliability similarity computation. Another reason is the compromise between a reduced size of wavelet signature and a controlled number of the splitting nodes of the local map.