GPS Integrated Inertial Navigation System Using Interactive Multiple Model Extended Kalman Filtering

Abstract

This paper presents an implementation of a Global Positioning System (GPS) integrated inertial navigation system (INS) for vehicle state estimation. The INS uses Extended Kalman Filtering (EKF) of the linearized state space model for state estimation. The two INS EKF models have differently tuned noise parameters. The models operate in parallel using an interactive multiple model (IMM) approach. The IMM mixes the state and state covariance estimates from both models to yield a combined estimate of the system states. The mixing weights are based on the likelihood of each model correctly tracking the system states. The likelihoods are computed using the innovation and innovation covariance matrices of each model. The model with the higher likelihood has a larger influence on the overall state estimation. The KITTI Vision Benchmark dataset has been utilized for testing and validation. The GPS coordinates have been transformed into a local tangent frame position estimation. Orientation measurements are provided by the dataset for heading correction. The analysis shows that the INS system accurately tracks the position and orientation; the IMM filter generally outperforms the single EFK model estimator during turning maneuvers where the IMM filter produces a lower mean position error than a single EKF filter.