Obstacle Detection and Collision Avoidance For A UAV With Complementary Low-Cost Sensors
Obstacle Detection and Collision Avoidance For A UAV With Complementary Low-Cost Sensors
Obstacle Detection and Collision Avoidance For A UAV With Complementary Low-Cost Sensors
June 1, 2015.
Digital Object Identifier 10.1109/ACCESS.2015.2432455
ABSTRACT This paper demonstrates an innovative and simple solution for obstacle detection and collision
avoidance of unmanned aerial vehicles (UAVs) optimized for and evaluated with quadrotors. The sensors
exploited in this paper are low-cost ultrasonic and infrared range finders, which are much cheaper though
noisier than more expensive sensors such as laser scanners. This needs to be taken into consideration for
the design, implementation, and parametrization of the signal processing and control algorithm for such
a system, which is the topic of this paper. For improved data fusion, inertial and optical flow sensors are
used as a distance derivative for reference. As a result, a UAV is capable of distance controlled collision
avoidance, which is more complex and powerful than comparable simple solutions. At the same time, the
solution remains simple with a low computational burden. Thus, memory and time-consuming simultaneous
localization and mapping is not required for collision avoidance.
INDEX TERMS Collision avoidance, obstacle detection, ultrasonic, infrared, autonomous, UAV, quadrotor,
quadrocopter.
I. INTRODUCTION
2169-3536
2015 IEEE. Translations and content mining are permitted for academic research only.
Personal use is also permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
599
E to react
towards an obstacle w and control parameters C
properly on the threat of a collision. With this information the
distance control (DC) regulates the distance. Thus collisions
are avoided.
ex
ey
dx
dy
E
vy
i = vx
(3)
wx wy
Ex C
Ey
C
IV. CONCEPT
A. OVERVIEW
Fig. 3 depicts the concept of the system from a high level view
divided into obstacle detection (OD) and collision
avoidance (CA). The obstacle detection consists of the sensing and the situation awareness (SA). The sensing is the
reading, conditioning and filtering of the OD (IR + US),
OF (optical flow) and IMU (inertial measurement unit)
sensors, to receive distances d, position changes 1p and
orientations (roll) and (pitch). These data S (1) are then
the input to the situation awareness, whose task is to process
the multiple information from the sensing to determine the
E (2), which contains the minimum distance available
vector
to the front, right, rear and left as well as the longitudinal (x)
and lateral (y) velocity v.
d0US
d0IRL
d0IRS
d US
d1IRS
d1IRL
IRS
IRL
d US
d1
d1
IRS
IRL
d3US
d
d
2
2
US
IRS
IRL
d3
d4
d3
1pOF
1pOF
x
y
d US
IRS
IRL
d3
d3
..
..
.. ..
5
S=
.
.
. .
US
d4IRL
d4IRS
d6
OF 1pOF
1p
IRL
IRS
x
y
d7US
d5
d5
d5IRL
d5IRS
d8US
d US
IRL
IRS
d
d
6
9
6
IRL
IRS
d US
d7
d7
10
US
d7IRL
d11
d7IRS
(1)
dFront
dRight
E = dRear
dLeft
vx
vy
E
= df
Ev
(2)
WS (k) = a (N 1 + b qO N2 )
+ c (N 3 + b qO N4 ) + d N5
0
WS (k) = e WS (k) + (1 e) WS (k 1)
(5)
(6)
N9 = |D (k 1) + O(k) I (k)|
0
WI (k) = a (N 6 + b qO N7 )
+ c (N 8 + b qO N9 )
0
WI (k) = e WI (k) + (1 e) WI (k 1)
WU (k) I (k) + WI (k) U (k)
D(k) =
WI (k) + WU (k)
(7)
(8)
Ei = g T (sin() sin i
+ sin () cos i ) (10)
2
2
VOLUME 3, 2015
Algorithm 1 TAF
Input: distance main neighbor left: ml
Input: distance direct neighbor left: nl
Input: distance main: m
Input: distance direct neighbor right: nr
Input: distance main neighbor right: mr
Input: tolerance:
Savings: last weights: w0 , w1 , w2
Output: fused minimum distance: d
if (ml < nl ) then nl = end if
if (mr < nr ) then nr = end if
if (m < nl ) then
if (m < nr ) then w0 = 0
else w0 = 1
end if
else
if (nl < nr ) then w0 = 1
else w0 = 1
end if
endP
if
S= wi
i
w2 = w1
w1 = w0
d =m
if (S < 1) then d = nl end if
if (S > 1) then d = nr end if
return d
E9 E0
E3
E6
E11 E2
E5
E8
E
E6
E9
T = 0 E3
(11)
E7 E10
E1 E4
E3 E6
E9
E0
The task of the triple awareness fusion (TAF) is to
determine the four distances dEf (2) as one fundamental base
VOLUME 3, 2015
FIGURE 9. CA concept.
E. DISTANCE CONTROL
(13)
(d+ < dth ) & (d < dth ) & (d+ + d < 2 dmin )
d+ < d
(14)
(15)
FIGURE 15. Test Case A (corner, run 1). Top: Data fused distances to the
wall, input to the DC. Bottom: Measured angle from IMU, manipulated by
the DC.
pushed into a corner till the front and left distance was
below dmin . Consequently the collision avoidance automatically activated, controlled the flight of the quadrotor and kept
its position in the corner by controlling the distance to the
front and to the left. Table 1 shows the results. The mean
distances measured from the center of the quadrotor towards
the walls were 144cm and 135cm. Accordingly the mean
distance between the tip of the propellers and the wall was
about 1m as desired.
TABLE 1. Results: Test case A (stationary in corner).
VI. EVALUATION
The system has been evaluated for more than 100 experimental tests. In the first empiric experiments the parameters of
the system have been optimized. Then, for evaluation of the
system, its performance during three different test cases has
been analyzed and the results are presented.
The test cases are a stationary flight in a corner, a stationary
flight between two walls and a flight in a room with a person
moving towards the quadrotor. Every experiment has been
repeated 10 times, lasting about 3 minutes, but only 60s has
been display. These short time periods were chosen both to
have many repetitions with comparable data under the same
conditions and to avoid corruption of the data by the voltage
drop of the batteries.
During all experiments the quadrotor operated fully
autonomously without manual input or external devices.
An external computer operating as ground station was used
to receive telemetry data such as the evaluated values.
It is not claimed, that all parameters are optimized in
research prototype. That is why the performance of the
system is analyzed by its data fusion output and control
input, which can be seen as almost true, as otherwise
a failure would be clearly seen in the data as well
(compare Fig. 15 and Fig. 17).
A. STATIONARY: CORNER
The idea of the first test case was to investigate the obstacle
detection, collision avoidance and autonomous flight behavior in the simplest scenario. Since the system was designed for
distance controlled flight, stationary obstacles like one wall
in the front and on the left side are sufficient for autonomous
stationary flight. In this experiment the quadrotor has been
606
FIGURE 16. Test Case B (between two walls, run 1). Top: Data fused
distances to the wall, input to the DC. Bottom: Set point angle, output of
the DC.
Table 2 summarizes the results of this experiment. In comparison to Table 1 it can be seen, that the mean distance as
well as the standard deviation are lower. The explanation for
VOLUME 3, 2015
VOLUME 3, 2015
609