Robust covariance estimation in sensor data fusion

Abstract

This paper addresses the robust estimation of a covariance matrix to express the uncertainty when fusing information from multiple sensors. This is a problem of interest in multiple robotics domains and applications, namely in Search and Rescue. The paper compares the Covariance Intersection (CI) and a class of Orthogonal Gnanadesikan-Kettenring (OGK) estimators. The performance of the two estimators is analyzed using the 2-norm of the covariance matrix. Simulation tests are presented showing that OGK tends to outperform CI when the correlation between sensors is significant and in the presence of outliers. The formal bounds found show that each of the two estimators outperforms the other depending on the region of the covariance matrix spectrum they are operating in. The conclusions point to the use covariance estimation systems with a hybrid of the two estimators.