Multisensor Navigation Systems Multicore Architecture

Abstract

Multisensor navigation systems with an inertial navigation system (INS), employ GPS and other aiding sensors to correct INS errors and improve navigation accuracy. Kalman filter (KF) is used to fuse the aiding sensor measurements and estimate errors. Both the time update, in which the state vector and covariance matrix are propagated, and the measurement update, in which all the aiding sensor measurements are fused with the INS-derived variables, have high computational costs. Two forms of the KF, the original covariance KF and the newer information KF, are examined. The most efficient parts of each are combined into a hybrid covariance/information filter. The state and measurement spaces of the hybrid filter are partitioned by sensors leading to blocked matrix formulas for both time update and measurement update. The blocked equations lead directly to threaded architecture implementations. The threaded architecture permits either serial operation with a single processor or parallel operation with multiple processors. Formulas are derived and evaluated for the floating-point operation counts for a general unblocked-serial hybrid filter and its blocked-serial and blocked-parallel implementations for an INS aided by GPS and other sensors. The blocked-parallel hybrid filter shows a significant improvement compared to the blocked-serial and unblocked-serial implementations.