Efficient, Multifidelity Perceptual Representations via Hierarchical Gaussian Mixture Models

This paper presents a probabilistic environment representation that allows efficient high-fidelity modeling and inference toward enabling informed planning (active perception) on a computationally constrained mobile autonomous system. The proposed approach exploits the fact that real-world environments inherently possess structure that introduces dependencies between spatially distinct locations. Gaussian mixture models are employed to capture these structural dependencies and learn a semiparametric, arbitrary resolution spatial representation. A hierarchy of spatial models is proposed to enable a multifidelity representation with the variation in fidelity quantified via information-theoretic measures. Crucially for active perception, the proposed modeling approach enables a distribution over occupancy with an associated measure of uncertainty via incorporation of free space information. Evaluation of the proposed technique via a real-time graphics processing unit based implementation is presented on real-world data sets in diverse environments. The proposed approach is shown to perform favorably as compared to state-of-the-art occupancy mapping techniques in terms of memory footprint, prediction accuracy, and generalizability to structurally diverse environments.