Study on three-dimensional terrain referenced navigation using machine learning

Abstract

Terrain referenced navigation (TRN) is a navigation technology that estimates position of aircraft by comparing the altitude measured by an altimeter with a digital elevation model (DEM). In this study, I proposed an interferometric radar altimeter (IRA) based TRN which is suitable for UAVs or cruise missiles. The IRA is a sensor that acquires the relative distance from the aircraft to the nearest point as three-dimensional data. I design an IRA-based TRN method to apply a Bank of Kalman filter (BKF) or particle filter (PF) and solve various problems for performance improvement through machine learning. Three main factors that determine the performance of TRN are a high-precision altimeter, a high-resolution DEM, and an optimized TRN algorithm. The IRA has a disadvantage in that the sensor output is significantly uncertain due to the signal processing process and environmental factors, which is in contrast to the advantage of providing the 3D information. In this study, the validity of the technique is verified by applying a valid measurement extraction method through a radial basis function and extreme learning machine with a BKF and PF-based TRN. A high-resolution DEM requires a large amount of memory space, and widely used interpolation methods that deal with this include unpredictable errors. In this study, I derived a regression model by learning the high-resolution DEM using a multi-layered radial basis function based extreme learning machine method. As a result, the memory space is remarkably reduced and the performance is improved. In addition, TRN can obtain stable performance even in a situation where a GNSS is no possible for a long time, but it has a disadvantage in that performance is degraded in flat and repetitive terrains. To address this disadvantage, I proposed a technique to adjust the noise covariances and the measurement model of the PF using the recurrent neural network (RNN) to design the filter so that it is stable in flat and repetitive terrains. I verify how close to the optimal design the algorithm is through an analysis of the Cramér-Rao lower bound (CRLB).