Master’s Thesis Review

This review is based on a manuscript of 67 pages corresponding to the Master’s thesis of Jan Vakula. The work of this thesis was performed in the Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague, under the supervision of Dr. Tomáš Svoboda. The goal of this thesis was to test and improve a visual odometry algorithm developed by Jiri Divis in his Master’s thesis. The original and modified version of the odometry algorithm were exhaustive tested and compared in various types of environments and under different conditions. The main improvement of the algorithm was achieved by the enhancement of the image keypoints management. In order to do that, several issues were faced. First, a Fourier analysis was proposed to detect blurry images and prevent the algorithm use them. Blurry images are problematic since only few keypoints can be extracted from them. Second, for better feature matching the original algorithm was updated so that the closest keypoint is calculated as Manhattan distance instead of Euclidean distance. Third, unlike the original algorithm, the images are not divided into the squares where the nearest point is found, the nearest point is searched in the whole image. Finally, the guide-matching algorithm was changed to establish the correspondences between keypoints from the actual image and keypoints from the previous image, since the translation robot’s motion is as little as possible. Using these correspondences and the correspondences between keypoints from the key image and keypoints from the previous image, the correspondences between keypoints from the actual image and keypoints from the key image can be determined very precisely. This last improvement is perhaps the main contribution to the new enhanced version of the visual odometry algorithm proposed by Jiri Divis. The presented visual odometry algorithm was compared with other types of odometry algorithms (visual compass and ICP) and with ground-truth (obtained from VICON and Leica). The results of the performed experiments show that the implementation of the new enhanced version of the algorithm outperforms the original one. This occurs in all test scenarios, but especially in outdoor