We look at estimating occluded or otherwise missing regions of point-cloud terrain data using Gaussian Process Regression (GPR). The ability to estimate missing data allows motion planners to create more aggressive and cost efficient paths. Previous approaches in applying GPR to terrain estimation have large computational complexity or trade off accuracy of far-field data with computational complexity. We present Multiresolution Partitioned Gaussian Process Regression (MPGPR) to address these weaknesses. MPGPR improves upon past results by 1) using a quadtree structure to intuitively partition data in order to speed up parameter optimization times, as well as 2) using a hierarchical structure to capture long range data dependencies while preserving local detail. This hierarchical structure allows the method to perform well on practical datasets that may have nonuniform densities. Our method is evaluated on many artificially generated terrains as well as real data obtained from the Mars Curiosity Rover.
Multiresolution partitioned Gaussian process regression for terrain estimation
01.03.2018
1796809 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch