Shape registration using optimization for mobile robot navigation

The theme of this thesis is shape registration (also called shape matching or shape alignment) using optimization-based algorithms. We primarily address this problem in the context of solving the mobile robot self-localization problem in unknown environments. Here the task is matching 2D laser range scans of the environment to derive the relative position and heading of the robot. The difficulties in this problem are that the scans are noisy, discontinuous, not necessarily linear, and two scans taken at different positions may not completely overlap because of occlusion. We propose two iterative scan matching algorithms which do not require feature extraction or segmentation. Experiments demonstrate that the algorithms are effective in solving the scan matching problem. Based on the result of aligning pairwise scans, we then study the optimal registration and integration of multiple range scans for mapping an unknown environment. Here the issue of maintaining consistency in the integrated model is specifically raised. We address this issue by maintaining individual local frames of data and a network of uncertain spatial relations among data frames. We then formulate an optimal procedure to combine all available spatial relations to resolve possible map inconsistency. Two types of sensor data, odometry and range measurements, are used jointly to form uncertain spatial relations. Besides the applications for mobile robots, we also study the shape registration problem in other domains. Particularly, we apply extensions of our methods for registration of 3D surfaces described by range images, and 2D shapes from intensity images.