Local reference filter for life-long vision aided inertial navigation

Abstract

Filter based system state estimation is widely used for hard-realtime applications. In long-term filter operation the estimation of unobservable system states can lead to numerical instability due to unbounded state uncertainties. We introduce a filter concept that estimates system states in respect to changing local references instead of one global reference. In this way unbounded state covariances can be reset in a consistent way. We show how local reference (LR) filtering can be integrated into filter prediction to be used in square root filter implementations. The concept of LR-filtering is applied to the problem of vision aided inertial navigation (LR-INS). The results of a simulated 24 h quadrotor flight using the LR-INS demonstrate longterm filter stability. Real quadrotor flight experiments show the usability of the LR-INS for a highly dynamic system with limited computational resources.