(Massachusetts Institute of Technology and Woods Hole Oceanographic Institution, 2014-02)
VanMiddlesworth, Mark
Motivated by inspection of complex underwater environments, we have developed a
system for multi-sensor SLAM utilizing both structured and unstructured environmental
features. We present a system for deriving planar constraints from sonar data,
and jointly optimizing the vehicle and plane positions as nodes in a factor graph. We
also present a system for outlier rejection and smoothing of 3D sonar data, and for
generating loop closure constraints based on the alignment of smoothed submaps.
Our factor graph SLAM backend combines loop closure constraints from sonar data
with detections of visual fiducial markers from camera imagery, and produces an online
estimate of the full vehicle trajectory and landmark positions. We evaluate our
technique on an inspection of a decomissioned aircraft carrier, as well as synthetic
data and controlled indoor experiments, demonstrating improved trajectory estimates
and reduced reprojection error in the final 3D map.
