Hierarchical Gaussian processes for robust and accurate map building

Abstract

This paper proposes a new method for building occupancy maps and surface meshes using hierarchical Gaussian processes. Previously, we found that a Gaussian process, one of the state-of-The-Art machine learning techniques for regression and classification, could serve a unified framework for occupancy mapping and surface reconstruction. Particularly, due to its high computational complexity, we partitioned both training and test data into manageable subsets and applied local Gaussian processes. However, when the local regions do not include any observations, this approach makes no predictions and thus generates blanks in the map. Therefore, in order to deal with missing data and make it more robust, we combine a global Gaussian process with local Gaussian processes. The global Gaussian process covers the overall trend of the whole observations, while the local Gaussian processes adapt to precise local observations. We demonstrate our method with a real dataset and compare accuracy and speed with OctoMaps and our previous method, GPmaps. Experimental results show that our method is relatively slow than the previous work but produces more robust and accurate robotic maps. Show more