Space-carving Kernels for Accurate Rough Terrain Estimation

Abstract

Accurate terrain estimation is critical for autonomous offroad navigation. Reconstruction of a three-dimensional (3D) surface allows rough and hilly ground to be represented, yielding faster driving and better planning and control. However, data from a 3D sensor samples the terrain unevenly, quickly becoming sparse at longer ranges and containing large voids because of occlusions and inclines. The proposed approach uses online kernel-based learning to estimate a continuous surface over the area of interest while providing upper and lower bounds on that surface. Unlike other approaches, visibility information is exploited to constrain the terrain surface and increase precision, and an efficient gradient-based optimization allows for realtime implementation. To model sensor noise over varying ranges, a non-stationary covariance function is adopted. Experimental results are presented for several datasets, including groundtruthed terrain and a large 3D stereo dataset.