Graphics Reference
In-Depth Information
SLAM graph complexity during operation, using variable elimination and constraint
pruning with heuristic schedules. These methods keep optimization and storage costs
commensurate with explored area rather than with time of exploration while causing
minimal loss in mapping and localization accuracy.
An instantiation of the approach is demonstrated on real datasets with planar
ground-truth reference. The system operates successfully even at frame rates below
2Hz. Comparing the results with and without complexity reduction demonstrates
that the reduced graph yields similar localization accuracy at a small fraction of the
computational cost.
2.2 Related Work
2.2.1 View Recognition for SLAM
View recognition engines have proven attractive components for SLAM systems
because they permit robust and flexible loop closing. Instead of making correspon-
dences between individual features ormeasurements, visual or otherwise, viewrecog-
nition engines typically match constellations of features or entire images without
requiring feature tracking.
Williams et al. [
20
] rely on tracking for normal EKF SLAM operation, but use
view recognition to recover from failure. Several features are matched to the existing
map using appearance and structure constraints in order to reinitialize tracking.
The Parallel Tracking andMapping(PTAM) [
10
] systemalso employs view recog-
nition for recovery from tracking failure. Instead of using feature-based methods for
identifying candidate views, the system performs image-to-image correlation using
heavily blurred, low-resolution versions of the reference and query images. A crude
pose estimate is deduced from the result of the inverse-compositional matching,
following which tracking resumes.
Eade andDrummond [
4
] group subsets of features into local maps during tracking-
based SLAM. Correspondences are made between local maps to connect them or
to recover from tracking failures. The image-to-map matching first selects a subset
of local maps to consider using a bag-of-words ranking, and then performs local
matching to determine feature-to-feature correspondences. This two-step process is
common to many view recognition systems, often instantiated as a bag-of-words
prefilter followed by re-ranking using geometric constraints [
17
].
The above approaches rely on tracking and use view recognition as an out-of-
band method for failure recovery. Our approach instead performs recognition at
every time step as the primary source of observations. The system of Karlsson et al.
[
9
] is similar, constructing
landmarks
out of constellations of SIFT [
14
] features
and employing nearest neighbors and a simple Hough transform as the recognition
algorithm. The system is further refined by Eade et al. [
5
] by replacing the particle-
filter back end with a graph SLAM back end that is described in further detail in
