| Abstract |
Unmanned Aerial Vehicles (UAVs), particularly multi-rotor UAVs, have gained significant popularity in the autonomous
robotics research field. The small size and agility of such aircrafts makes them an ideal candidate for use in restricted
environments. With the ability to navigate freely in 3D space, UAVs offer the possibility to reach places that are otherwise
inaccessible to humans. As such, UAVs find numerous applications in commercial and research tracks, such as inspection,
surveillance, search and rescue (SaR) and aerial mapping. However, their remote operation is challenging for non-highly
trained personnel, due to the fast dynamics of UAVs and the inherent difficulty to estimate distances to far away objects.
In order for an autonomous UAV to safely and reliably navigate within a given environment, the control system must be
able to determine the state of the UAV at any given moment. The state consists of a number of extrinsic variables, such as
position, velocity and attitude of the UAV. The latter are commonly provided in outdoor operations via the Global Positioning
System (GPS). While GPS has been widely used for long range navigation in open environments, its performance
degrades significantly in constrained environments and is unusable indoors. Accordingly, state estimation for UAVs in
the named environments is a popular, contemporary research area. Many successful solutions have been developed using
laser-range finder sensors. These sensors provide very accurate measurements at the cost of increased power and weight
requirements. Visual sensors constitute an attractive alternative state estimation sensor; they offer high information content
per image coupled with light weight and low power consumption. As a result, many recent works have focused on state
estimation on UAVs, where a camera is the only exteroceptive sensor.
In this thesis we aim at advancing the navigation capabilities of UAV systems, through the development of a framework
that facilitates autonomous flights in previously unknown indoor and outdoor workspaces, using dual on-board cameras as
the primary sensors. Those sensors introduce interesting challenges and need to be considered in the design of each part of
the navigation framework. Therefore, the work presented in this thesis focuses on the problem of visual based navigation
for UAVs. For the sake of robustness in autonomous flights, visual information is fused with measurements provided by an
Inertial Measurement Unit (IMU), available on every UAV. The complementary nature of visual and inertial data renders
this setup a particularly powerful sensor combination, albeit posing challenges in their synchronization and calibration.
With this minimal sensory setup, in the current dissertation we have studied, implemented and evaluated methods for UAV
control, state estimation, and autonomous navigation.
The first contribution of this thesis regards state estimation and control. We have shown that both parts can strongly
benefit from a good model that provides valuable information about possible motions. Including the model in the state
estimator, combined with a pressure sensor, renders the velocity and the two inclination angles observable. We present a
versatile framework, using an Extended Kalman Filter (EKF) to fuse visual information with inertial measurements from
the on-board sensor suite. Given that UAVs should be able to estimate their state from visual observations and also share
acquired environment information with other members of the team, a common frame of reference for both positioning and
observations is also derived.
For UAV control, we study the necessary dynamics and differential flatness of multi-rotor systems. We discuss potential
abstractions in order to employ position- and trajectory controllers, on different types of multi-rotor platforms. Based on
our investigation, we propose a control approach based on dynamic inversion, which avoids the commonly used attitude
control loop and significantly reduces the mathematical operations necessary, rendering the approach appropriate for
constrained on-board hardware.
Furthermore, we present a novel framework for horizon line (HL) detection that can be effectively used for UAV navigation.
Our scheme is based on a Canny edge and a Hough detector along with an optimization step performed by a Particle
Swarm optimization (PSO) algorithm. The PSO’s objective function is based on a variation of the Bag of Words (BOW)
method to effectively consider multiple image descriptors and facilitate computational efficiency. More specifically, the
image descriptors employed are L * a * b color features, texture features and PHOW-SIFT features.
Finally, a formulation of a visual navigation technique that enables the Asctec Firefly UAV to localize and navigate in
outdoor environments is outlined. This is achieved through the use of a general, scalable, tracking and mapping system
(SLAM). The overall goal of this dissertation has been to combine and integrate the above partial contributions, and enable
UAV-navigate in rather large, unknown outdoor environments. The latter implies that the UAV is stable flying on its own,
while an operator at a ground station only gives high-level commands, such as simple way-points, desired velocities and
dynamic trajectories. These outdoor tests prove the validity of the presented real-time and on-board navigation system.
Also, they show the capability of the entire sensor-fusion framework to convert an aerial vehicle into a power-on-and-go
system for real-world, large-scale and long-term operations.
The results in this thesis, showed that vision only navigation is possible as long as we are only concerned about the local
consistency of our environment. The advantages of multi- fusion systems were demonstrated by switching between single
and dual visual sensor setups and tested the different configurations with state noises and erroneous measurements. SLAM
framework was able to initialize and to produce results in real-time on-board navigation, in self-similarity environments,
where futures like grass in rural areas and asphalt in urban areas, causing the tracker to fail tracking the current pose,
and the mapper to fail re-localizing. Furthermore the navigation system was shown to be able to operate, even with low
resolution and distort image sequences from visual sensors, in contrast to other visual SLAM approaches.
To summarize, in the current dissertation we studied and developed an integrated framework to facilitate autonomous UAV
navigation in large, unknown outdoor environments. Accordingly, robust and stable UAV flight has been accomplished,
while an operator at a ground station only gives high-level commands, such as simple way-points, desired velocities and
dynamic trajectories.
|
Search
