This dissertation proposes solutions to three problems involved in a mobile robot equipped with an omnidirectional vision sensor. The first one is a matching problem of finding correspondences of features in omnidirectional images. Conventional methods based on feature tracking have limitations when the sensor motion becomes large. To produce reliable matching results even though there are large translation and rotation of a sensor, we propose a method that combines the advantages of Sum of Squared Difference (SSD) and Dynamic Time Warping (DTW). Dominant corresponding feature pairs are found using a proximity matrix and a similarity matrix based on SSD, and then the remaining feature matching is accomplished by DTW. Distortions due to the conic mirror can be well treated by DTW and ‘initial point constraint’ for DTW is imposed by SSD even with large sensor motion. Experimental results show that a zero failure rate of matching can be achieved in an indoor environment even though there are translational sensor motion larger than 10cm and any amount of sensor rotation. When a feature is identified at more than two sensor locations, 2D position of the feature can be estimated by triangulation. The experimental results of map building are given to demonstrate the validity of the proposed feature matching method.
The second problem we are dealing with is an absolute localization problem of a mobile robot. We devised a new linear method that can be used to find the position and orientation of a robot using only bearing measurements of landmarks. We also propose a method for finding correspondences between features in a 2D map and features in an image by integrating the proposed localization algorithm with ‘interpretation tree search’ algorithm. The primary advantage of the proposed method is that the localization is accomplished simultaneously in the matching phase. The localization algorithm and the feature matching method are presented and simulation results are added to show that considerable reduction of the search space can be achieved by proposed method
The last part of the dissertation is devoted to 2-D Structure From Motion (SFM) problem. In a mobile robot application, 2-D SFM is identical to the Simultaneous Localization and Map Building (SLAM) except that a scale cannot be determined in 2-D SFM. We propose a linear closed form solution for 2-D SFM problem using minimum five landmarks. The proposed SFM algorithm is faster and more stable than other methods that use nonlinear estimation. Simulations and experiments in a real environment verify the validity of the proposed algorithm.
