@MASTERSTHESIS\{IMM2006-04726,
    author       = "S. A. Gudmundsson",
    title        = "Robot Vision Applications using the {CSEM} SwissRanger Camera",
    year         = "2006",
    keywords     = "mathematical modelling, {3D} imaging, time-of-ight, range images, robot localization, pose estimation, range image segmentation, non-linear dimension reduction, Local Linear Embedding.",
    school       = "Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}",
    address      = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby",
    type         = "",
    note         = "Supervised Rasmus Larsen, and co-supervisor Jens Michael Carstensen, {IMM}.",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/4726-full.html",
    abstract     = "The SwissRanger is new type of depth vision camera using the Time-of-Flight (TOF) principle. It acquires in real time both normal intensity images and {3D} range images. It is an active range nder with a harmless light source emitting near infrared light at under {1W}. Most other active range finders are laser based and have much higher latency. The SwissRangers usefulness is proved here by solving two diverse robot vision applications: The mobile robot localization problem and the {3D} object pose estimation problem.

The robots localization is found by segmenting range images; into planar surfaces. The segmentation is done by calculating the local surface normals at each pixel, grouping the image into regions and robustly tting to planes using {RANSAC}. From these planes a map of the robots environment is constructed. For a robot to handle an object it has to recognize the objects pose or orientation in space. This is approached by using a dimensionality reduction method called Local Linear Embedding (LLE). A dataset, with range images of an object can be seen as points in a very high dimensional pixel space. It has been shown that for {3D} objects such points lie on nonlinear manifolds in the high dimensional space. The {LLE} technique reduces the dimensionality down to a true dimensionality of the manifold and reveals the separating characteristic in each point namely its pose. The pose of a new objects can then be detected by mapping it to this low dimensional space."
}