Graphics Reference
In-Depth Information
violated in practice, especially in visual systems, where the accuracy and uncertainty
of relative pose estimates depends heavily on viewpoint and scene structure.
We instead apply probabilistically sound graph reduction methods that limit the
complexity of the graph to a linear factor of the complexity of the explored space. Past
poses of the robot that are not used for view recognition can bemarginalized out of the
estimation, and their incident constraints are collapsed back into the graph. The mar-
ginalization procedure, equivalent to the update step of the Kalman filter or the vari-
able elimination step of the GraphSLAMalgorithm, is described by Konolige in [
11
].
Marginalization is used to systematically limit graph complexity in Kretzschmar
et al. [
13
], which employs an approximation rather than exact marginalization. The
approximate form is used to bound edge connectivity in the graph. Carlevaris-Bianco
and Eustice [
1
] propose an alternative via generic linear constraint node removal. In
contrast, we selectively prune edges incident to nodes of high degree, removing their
constraints from the GMRF in a conservative manner. The adaptive application of
marginalization and edge removal, discussed in Sect.
2.8
, is a significant feature of
this system.
2.3 System Overview
The input to the system is a sequence of images from the camera and a sequence
of differential motion estimates, derived from wheel odometry measurements or
other differential sensors. We refer to these differential measurements collectively
as odometry. The system outputs an incrementally updated estimate of the device's
current pose (localization) and estimates of a subset of its previous poses during
operation (mapping).
Two high-level components constitute the system: the visual recognition front end
and the constraint graph SLAMback end (Fig.
2.1
). The front-end processes the video
stream, yielding a global appearance database, a set of structured local appearance
Fig. 2.1
System structure overview
