(Georgia Institute of Technology, 2013-11)
Ta, Duy-Nguyen; Ok, Kyel; Dellaert, Frank
Despite recent progress, autonomous navigation
on Micro Aerial Vehicles with a single frontal camera is still
a challenging problem, especially in feature-lacking environ-
ments. On a mobile robot with a frontal camera, monoSLAM
can fail when there are not enough visual features in the scene,
or when the robot, with rotationally dominant motions, yaws
away from a known map toward unknown regions. To overcome
such limitations and increase responsiveness, we present a
novel parallel tracking and mapping framework that is suitable
for robot navigation by fusing visual data with odometry
measurements in a principled manner. Our framework can
cope with a lack of visual features in the scene, and maintain
robustness during pure camera rotations. We demonstrate our
results on a dataset captured from the frontal camera of a quad-
rotor flying in a typical feature-lacking indoor environment.
(Georgia Institute of Technology, 2006-06)
Quennesson, Kevin; Dellaert, Frank
Users know what they see before where they are: it is more natural to talk about high level visibility information ("I see such object") than about one's location or orientation. In this paper we introduce a method to find in 3D worlds a density of viewpoints of camera locations from high level visibility constraints on objects in this world. Our method is based on Rao-Blackwellized importance sampling. For efficiency purposes, the proposal distribution used for sampling is extracted from a visibility preprocessing technique adapted from computer graphics. We apply the method for finding in a 3D city model of Atlanta the virtual locations of real-world cameras and viewpoints.