Approximate Continuous Belief Distributions for Precise Autonomous Inspection

Abstract

Precise inspection of cluttered environments by computationally-constrained systems requires an efficient and high-fidelity representation of the operating space. We propose a methodology to generate perceptual models of the environment that optimally handle the variation in clutter and provide a multi-resolution and multi-fidelity representation of the environment. Further, our approach is able to capture inherent structural dependencies thereby enabling efficient and precise inference. The approach employs a hierarchy of Gaussian Mixtures to approximate the underlying spatial distribution. We make use of techniques grounded in information theory to estimate the optimal size of the mixture model and to generate and update the hierarchy of mixtures. We show that the proposed approach is superior in terms of memory requirements and accuracy as compared to other state of the art 3D mapping techniques such as NDT occupancy maps and Gaussian Process occupancy maps.