Graphics Reference
In-Depth Information
Fig. 2.5
Example images from the Seq2 (
left
)andSeq3 (
right
). The reflector beacons are NAV200
fiducials used for ground truth estimation
2.9.1 Metrics
We measure both the accuracy of the incrementally estimated trajectory and of the
final view map. The view map is the set of poses of view nodes in the graph at the
end of the run, including incremental optimization but without any post-processing.
Comparing the trajectory to the reference reflects localization accuracy during
the run. Comparing the map to the appropriate subset of the reference indicates
how well the system can be expected to localize in the same environment given
subsequent operation. Though the latter metric is more common, and generally shows
smaller errors compared to the reference, it does not necessarily reflect how useful
the localization is during online operation.
The estimated and ground truth trajectories are compared by first finding the
rigid transformation between them that minimizes sum-squared position error, using
RANSAC and least squares. The view map corresponds to a subset of the total robot
trajectory, so the same method is used to compute the view map error over that
subset.
2.9.2 Results
Figure
2.7
shows a portion of the graph computed for
with and without reduc-
tion. The node and edge density is significantly lower in the latter.
Table
2.3
shows error metrics and graph complexity for full and reduced graphs.
In the “Full” columns, the graph is heavily optimized and no nodes or edges are
removed. The “Reduced” columns show the same metrics when the number of pose
nodes is bounded by the number of views plus ten, and the maximum permitted node
degree is eight. The graph complexity is greatly reduced with little or no loss of
localization accuracy. As expected, the error of the view map is smaller than that of
the causally estimated trajectory consisting of the best estimate at each timestep, as
the map has incorporated all the information up to the end of the run.
Seq2
