Christoffer Valgren
Topological Mapping and Localization Using Omnidirectional Vision
Licentiate Thesis, 2007
Abstract: A topological representation of the world (i.e., a map where only places and connections between places are stored) provides an efficient means for a robot to perform localization and path planning. A topological map can be constructed in various ways, depending on the available sensors and the task to be performed. Vision — the main sensor modality that humans use to navigate — is used in this work to perform topological map building and localization for mobile robots. The map is based on quantitative image matching. Images similar to each other are clustered together, forming nodes that closely correspond to the human concept of a “place”. The resulting topological map naturally captures the structure of the environment; in areas where the (visual) environment changes rapidly, many nodes will be created, while in large, open areas only a few nodes will be created. To cluster the images into topological nodes, the incremental spectral clustering algorithm is used. This is a general-purpose clustering algorithm that easily can be adapted to the problem of topological mapping. Incremental spectral clustering is one of the major accomplishments of the thesis. The resulting map is used to perform localization. The environment is challenging; seven large data sets, spanning indoor and outdoor areas, acquired during four distinct seasons, is used to evaluate the algorithms. Different types of local features are evaluated from the perspective of localization. It is shown that one of them, the U-SURF algorithm, outperforms the competition. Using this type of local feature extracted from high-resolution images allows the robot to localize with a high success rate, even under extreme seasonal variations. This cross-season localization is another, important contribution of the thesis.
Download: [PDF for printing(10.0MB)] [PDF for screen viewing (4.1MB)]
  author = {Christoffer Valgren},
  title = {Topological Mapping and Localization Using Omnidirectional Vision},
  school = {Örebro University},
  type = {Licentiate Thesis},
  year = {2007},