Pose Tracking For Mobile Robot Localization From Large Scale Appearance Mosaics

Abstract

While visual odometry has unbounded error, navigation from pre-existing consistent scene models can generate extremely repeatable position estimates. This paper discusses a new approach to localization motivated by the fact that many man-made environments constain substantially flat, visually textured surfaces of persistent appearance.
For this important class of vision-based navigation problems the scene model can be reduced to a 2D surface painted with real textures - in other words, an image mosaic. Straightforward techniques from image-based localization and mosaicking are used to produce a field relevant AGV guidance system based on only vision and odometry. The visual tracking and localization aspects of the approach are described. We show that this approach to localization is able to exceed the speed barriers due to distortion and image overlap that are intrinsic to visual tracking and odometry. Speed can, however, become limited by a new mechanism - the inherent instability of visual tracking when operating in the regime beyond the overlap spped limit. Field trials currently demonstrate that the particular simplifications resulting from a downward looking camera configuration produce a guidance system repeatable to 1 mm throughout a 50,000 square foot facility with an MTBF (corresponding to loss of visual lock) of 106
images or five days of operation.