25-02-2013, 12:56 PM
Vision-Only Automatic Flight Control for Small UAVs
Vision-Only Automatic.pdf (Size: 1.89 MB / Downloads: 25)
Abstract
In this paper, a vision-based flight control system that
uses a skyline-detection algorithm is developed for application to
small unmanned aerial vehicles. The skyline-detection algorithm
can detect straight or uneven skylines. The system integrates
a remote controller, a remotely controlled airplane, a camera,
a wireless transmitter/receiver, a ground control computer, and
the proposed skyline-detection algorithm to achieve automatic
control of flight stability. Static and dynamic tests are conducted
to validate the system. In the static tests, the average accuracy
rate for skyline detection is 98.62% based on five test videos. In
the dynamic tests, straight and circular flights are used to verify
lateral and longitudinal stability for the proposed flight control
system. The experimental results demonstrate the performance
and robustness of the algorithm and the feasibility and potential
of a low-cost vision-only flight control system.
INTRODUCTION
OVER the last decade, the potential of unmanned aerial vehicles
(UAVs) has been demonstrated in D-cube missions
that are identified as dangerous, dirty, or dull. Recently, considerable
interest has arisen in small-scale lightweight low-cost
unpiloted aircraft (known as small UAVs) designed to fly at low
altitudes with limited payloads and power constraints. Aside
from military deployment, small UAVs are widely applicable
in civilian scenarios, because they are easy to build, carry,
and launch. When equipped with vision sensors, small UAVs
can execute numerous monitoring missions, including remote
sensing, traffic monitoring, forest fire surveillance, accident
reconnaissance, mapping, search, and rescue. Vision sensors
are indispensable for carrying out these missions. In addition, a
robust, reliable, and inexpensive autopilot system is necessary,
and attitude determination and control are the key to achieving
autonomy [1], [2].
UAV Subsystem
The UAV subsystem integrates a video acquisition module, a
remote control receiver module, a servo control module, and a
power supply module on a radio-controlled aircraft. The aircraft
is a SOARJET EP-19 (fuselage length: 87 cm), high-wing
(wingspan: 140 cm), four-channel (controlling rudder, ailerons,
elevator, and throttle), battery-powered, single-motor aircraft.
The subsystem is shown in Fig. 2. The aircraft weighs about
600 g and carries the four modules, which have a combined
weight of roughly 187 g. The video acquisition module consists
of a 1/3-in CMOS camera that is 20 × 20 × 20 mm in size and
15 g in weight, together with a 1.2-GHz wireless transmitter
and antenna.
SKYLINE-DETECTION ALGORITHM
The video images that are captured by a stationary chargecoupled
device (CCD) camera contain both the background and
moving object images. Because most of the background image
is motionless, the color components red ®, green (G), and
blue (B) of the background pixels should not change from one
frame to the next during the background extraction. Unfortunately,
several background pixels change in an image sequence
because of moving objects that pass through these background
pixels and the variation in illumination. Therefore, each pixel
in an image sequence will have as many different colors as the
candidates for a background pixel. For any given pixel of an
image sequence, the probability of the background color will
be higher than the probabilities of the colors counted when the
moving objects pass through in the image sequence. Previous
studies [1], [21], [22] used this concept to develop background
extraction algorithms. However, the memory consumption and
processing time required to obtain the initial background image
are major drawbacks, because these probability-based algorithms
use an image sequence in a fixed-time duration to count
the probabilities of each color for each pixel. The algorithm
described in this paper resolves the problems with regard to the
large-memory requirement and processing time.
Detection Algorithm for the Tracking Stage
In the tracking stage, a new ROI is confined to the estimated
region of the terminal points of the skyline to avoid image interference
and increase the processing speed. The linear Kalman
filter is employed to track the terminal points from consecutive
frames. The new ROI is the search area of size r × r pixels
surrounding the estimated point from the linear Kalman filter.
Here, r is set equal to 3c, and the initial value of c is set equal
to 2. Because the Kalman filter can estimate the locations of
the terminal points, the size of the search area (ROI) is reduced
to 9 × 9 (c = 2). When the terminal point is not detected from
the search area, the parameter c is increased by 1 to enlarge
the search area.
EXPERIMENTAL RESULTS
In this section, we use static and dynamic tests to demonstrate
that the proposed algorithm is effective to stabilize the
flight of a UAV. The system integration and static tests are
first described in detail. Then, the experimental results for
the skyline-detection algorithm are presented based on test
videos in various environments. The test videos indicate that
the algorithm can correctly identify the skyline more than 98%
of the time. Finally, in the dynamic tests, the algorithm is
employed to control the stability of a small airplane, and the
flight data show that the attitude of the airplane remains stable
during autonomous control.
System Integration and Static Tests
Because the roll angle and the pitch distance are detected
from the virtual horizon by the GCC, the flight control signals
of the airplane are analyzed by the GCC and are transmitted
through a remote controller. To manage the interface between
the GCC and the remote controller, the aileron and elevator
input–output (I/O) circuit of the remote controller is modified
to connect to the analog-to-digital/digital-to-analog (AD/DA)
board mounted on the GCC, as shown in Fig. 7. In addition, a
switch is installed on the remote controller to switch the control
mode between manual and automatic.
Results of the Test Flight
After the series of ground tests, a test flight is conducted to
evaluate the lateral and longitudinal stability of the proposed
vision-based flight control system in the real world. At the
beginning of the test, the UAV is hand launched and remotely
controlled in the manual mode by an experienced pilot until it
reaches a specific altitude of about 60mabove the ground. Once
the desired altitude and the correct initial skyline have been
attained, the mode is switched to automatic, and the control of
the ailerons and elevator is handed over to the GCC (although
the rudder and throttle are still operated by the pilot). Before
landing, the control mode is switched back to the manual mode,
and the UAV is again controlled by the pilot. A GlobalSat
DG-100, a Global Positioning System (GPS) device that weighs
115 g, is carried onboard to record the trajectory during the test
flight.
CONCLUSION
This paper has introduced a vision-based flight control system
for small UAVs using a skyline-detection algorithm. The
system integrates a remote controller, a remotely controlled
airplane, a camera, a wireless transmitter/receiver, a GCC, and
the proposed skyline-detection algorithm to achieve automatic
control of flight stability. As the experimental results indicate,
this low-cost flight control system offers the advantages of realtime
constraint, robustness to noise, and reliable detection. Furthermore,
the proposed skyline-detection algorithm can detect
both straight and uneven skylines, with an average accuracy rate
of 98.62%, based on the five test videos. The flight test results
demonstrate that the system controls and stabilizes a UAV
with vision-only flight control. The limitation of the proposed
algorithm is that the skyline must remain in the image frame
at all times. In future research, we will add rudder and throttle
controls and transmit the GPS signal to the GCC to fully control
the automatic flight of a UAV.