Most vision-based UAV (Unmanned Aerial Vehicle) navigation algorithms extract features, such as horizon and peak, from 2D input image, then match with features from DEM through registration. The difficulties of the horizon and peak extraction, which is the basic process of such local feature matching approaches, originate from the variations of input images such as noise, viewing direction, scale, existence of horizon. So, the success of the feature extraction will depend on its ability to cope with these variations. This thesis presents a new feature extraction method robust to these variations.
For the reliable horizon extraction, the thesis proposes to combine robust boundary extraction with high-level scene recognition by using the characteristics of the human visual system. For example, binding, which is the main process of the visual perception, is the process to bind two different properties by a series of attention. A computational model based on gestalt theory is developed to implement the binding process, in which the boundary extracted from global error minimization is combined with the high-level scene characteristic of separated regions. In sum, the proposed algorithm extracts the horizon robustly even in the presence of strong image noises.
For the peak extraction, we use the CSS (curvature scale space) corner detection algorithm which is stable to scale, viewing direction, noise variations. Extracting peak correspondence of the image sequence is implemented using curvature matching.
The proposed robust feature extraction algorithm has been verified through experiments using many IR (infrared) images of the natural terrain environments.
