Pulli, Kari, Vision Methods for an Autonomous Machine Based on Range Imaging

Range imagery produces the kind of geometric information about the environment that an autonomous machine needs for operation. Not only does the range data provide explicit information about the obstacles that might hinder movements, but it also provides a means of recognizing objects that should be manipulated. TULKO, the \Machine of the Future" project, has chosen laser range imagery as it main source of vision data. In TULKO the location and orientation of the work space is determined with respect to a paper roll manipulator. The locations of individual paper rolls must also be determined in such a way that the manipulator can transfer them. In the actual application the manipulator can autonomously load the rolls from a platform to a ship. This thesis presents a Hough transform based method for locating standing paper rolls of a known radius directly from the depth data. The locations are transferred into the manipulator coordinate system by calibrating the sensor and manipulator coordinator systems. For object recognition from depth maps, a new method for range image segmentation was created. The method rst calculates local surface normals from the depth data using robust methods, and decomposes the normal vector into three orthogonal components. Based on two of the components it is possible to determine discontinuities in the surface orientation. A third component, a scaled depth value, is added to the two normal components, and the resulting triplets are considered as a single color vector with three \color" bands. This multi-band image is then segmented using a hierarchical connected component method. The resulting segmentation is re ned by robustly tting surface equations to the segments and thereby determining at the region borders to which segment the individual pixels belong.