A computationally efficient low-bandwidth method for very-large-scale mapping of road signs with multiple vehicles

Abstract

This paper provides a flexible solution to the problem of building and maintaining a very-large-scale map using multiple vehicles. In particular, we consider producing a map of landmarks on the scale of thousands of kilometres in an outdoor environment. The algorithm is distributed across multiple vehicles each given the task of producing and updating a local map. The vehicles are equipped with a range of sensors and selectively communicate maps to and from a central station in a bandwidth-constraint environment. The potentially overlapping local maps are asynchronously transmitted back to a central fusion centre where a global map repository is maintained. The work addresses two of the most common issues of mapping in large-scale environments, namely, computational complexity and limited communication bandwidth. The proposed communication architecture is scalable and is capable of dealing with time-varying overlapping map sizes. A general data fusion framework based on covariance intersection is proposed to tackle the problem of redundant information propagation that is caused by communicating sub-maps of arbitrary size in the network. We also provide an analysis on the applicability of covariance intersection, as compared to the optimal approach when no cross-correlation is known between estimates from different vehicles. We further analyse the solution using a number of illustrative examples.