Towards Efficient Occupancy Mapping via Gaussian Process Latent Field Shaping

TL;DR

This paper introduces a novel Gaussian Process-based occupancy mapping method that directly manipulates the latent function to efficiently incorporate free space information, improving continuous environment representation.

Contribution

It proposes a new GP latent field shaping approach that directly integrates free space as a prior, distinguishing between free and unknown areas for improved occupancy mapping.

Findings

Competitive reconstruction accuracy in simulated environments

Efficient integration of free space information

Improved continuous occupancy representation

Abstract

Occupancy mapping has been a key enabler of mobile robotics. Originally based on a discrete grid representation, occupancy mapping has evolved towards continuous representations that can predict the occupancy status at any location and account for occupancy correlations between neighbouring areas. Gaussian Process (GP) approaches treat this task as a binary classification problem using both observations of occupied and free space. Conceptually, a GP latent field is passed through a logistic function to obtain the output class without actually manipulating the GP latent field. In this work, we propose to act directly on the latent function to efficiently integrate free space information as a prior based on the shape of the sensor's field-of-view. A major difference with existing methods is the change in the classification problem, as we distinguish between free and unknown space. The `occupied' area is the infinitesimally thin location where the class transitions from free to unknown. We demonstrate in simulated environments that our approach is sound and leads to competitive reconstruction accuracy.