Aerospace and Electronic Systems Magazine August 2017 - 22

Feature Article:

DOI. No. 10.1109/MAES.2017.160176

Compressed Fusion of GNSS and Inertial Navigation
With Simultaneous Localization and Mapping
Jonghyuk Kim, Australian National University, Canberra, Australia
Jiantong Cheng, Key Laboratory of Complex Aviation System Simulation, Beijing, China
Jose Guivant, University of New South Wales, Sydney, Australia
Juan Nieto, Autonomous Systems Lab, ETH Zurich, Switzerland

INTRODUCTION
Unmanned aerial vehicles (UAVs) have attracted significant attention from both civilian and defense industries over the last few
decades. With the advances in low-cost inertial sensor technology
and the global navigation satellite system (GNSS), the six degreesof-freedom (6DOF) vehicle state can be estimated accurately by
fusing this information, which has been a crucial step toward realtime guidance and flight control [1], [2].
GNSS and inertial integrated navigation has been extensively
studied, with many commercial products in the market for mobile
and aerial vehicles [3]-[6]. Depending on the GNSS information
used in fusion, the integration architecture has been classified as
loosely coupled, tightly coupled, or deep integration. The stateof-the-art architecture fuses the dynamic information from inertial
sensors to aid code and/or carrier-phase tracking loops, thus enabling the fast acquisition of a signal in highly dynamic conditions
[5], [7].
Although these integrations have been quite successful, precise
navigation in cluttered environments is still challenging because
of the use of low-quality and low-cost inertial measurement units
(IMUs). If satellite signals are partially or fully blocked for an extended period, the system can provide reliable navigation solutions
only for a limited time, typically in the range of a few minutes if
the IMU is properly calibrated.
To address this problem in the robotics community, simultaneous localization and mapping (SLAM) has been developed;
it solves the mapping problem of unknown environments while

Authors' current addresses: J. Kim, College of Engineering and
Computer Science, Australian National University, Canberra,
ACT 0200, Australia, E-mail: (jonghyuk.kim@anu.edu.aul). J.
Cheng, Key Laboratory of Complex Aviation System Simulation, Beijing, 100076, China. J. Guivant, School of Mechanical
and Manufacturing Engineering, University of New South
Wales, Sydney, NSW 2052, Australia. J. Nieto, Autonomous
Systems Lab, ETH Zurich, Zürich 8092, Switzerland.
Manuscript received August 15, 2016, revised January 28, 2017,
March 21, 2017, and ready for publication May 3, 2017.
Review handled by A. Dempster.
0885/8985/17/$26.00 © 2017 IEEE
22

simultaneously localizing the vehicle using the constructed map
[8]-[12]. In contrast to the conventional navigation systems, such
as GNSS and terrain and image matching systems, no infrastructure or a priori information about the environment is required, thus
making it quite attractive for standalone or aided navigation in
global positioning system (GPS)-denied environments. Based on
the seminal work by [13] and [14], there have been two types of
approaches: filtering and graph optimization.
Because this article aims for a real-time solution, only the filtering-based approach is summarized here. Filtering-based SLAM
solutions are mostly based on the extended Kalman filter (EKF)
[15], unscented Kalman filter (UKF) [16], particle filter known
as FastSLAM [17], and sparse extended information filter (SEIF)
[18].
The EKF-based solutions have been extensively developed
in the robotics community because of their ease of implementation and relatively low computational cost [14]. The nonlinearity involved in the vehicle dynamics and measurements, however, can cause degraded performance or even divergence of the
filter because of the errors involved in the linearization process
and Jacobian calculation. This issue is also encountered in the
SEIF.
In contrast, the sampling-based solutions utilize a set of
samples to estimate the state and the uncertainty without explicit
linearization, thus making it a Jacobian-free method yet with an
added computational complexity. For example, unscented SLAM
deterministically draws samples and their weights for nonlinear
prediction and update. Numerous methods have been investigated
for unscented SLAM. For example, [19] presented a solution based
on the square-root unscented filter for a visual mono-SLAM problem in which the square root of the covariance matrix is propagated; it has a cubic complexity of O(N3), with N being the dimension of the state vector. To reduce the computational cost, [20]
proposed a hybrid filtering strategy in which only the vehicle pose
is heuristically sampled by the unscented filter, while the whole
state vector is updated using the conventional EKF. Linearization,
however, is still required to predict the cross-correlation between
the vehicle and the map. To circumvent this linearization problem,
[21] proposed a partial sampling strategy based on the linear regression model. This work showed that Jacobian matrices could be
inferred from the propagated sigma points, subsequently achieving

IEEE A&E SYSTEMS MAGAZINE

AUGUST 2017

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine August 2017