IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO.

1, JANUARY 2020 243

Robust Autonomous Navigation of Unmanned

Aerial Vehicles (UAVs) for Warehouses’
Inventory Application
Woong Kwon , Jun Ho Park , Minsu Lee , Jongbeom Her , Sang-Hyeon Kim , and Ja-Won Seo

Abstract—The inventory inspection using autonomous un-

manned aerial vehicles (UAVs) is beneficial in terms of cost, time,
and safety of human workers. However, in typical warehouses, it
is very challenging for the autonomous UAVs to do inventory task
motions safely because aisles are narrow and long, and the illumi-
nation is poor. Prior autonomous UAVs are not suitable for such
environments, since they suffer from either localization methods
prone to disturbance, drift and outliers, or expensive sensors. We
present a low-cost sensing system with an extended Kalman filter
(EKF)-based multi-sensor fusion framework to achieve practical
autonomous navigation of UAVs in warehouse environments. To
overcome the inherent drift, outliers, and disturbance problems of
naïve UAV localization methods, we suggest 1) exploiting compo-
nent test of Mahalanobis norm to reject outliers efficiently, 2) intro-
ducing pseudo-covariance to incorporate a visual SLAM algorithm,
and 3) recognizing floor lanes to get absolute information–as robust
data fusion methods. Exemplar results are provided to demonstrate
the effectiveness of the methods. The proposed system has been
successfully implemented for diverse cyclic inventory inspection
tasks in a materials warehouse.
Index Terms—Aerial systems, applications, aerial systems,
perception and autonomy, sensor fusion. Fig. 1. Our UAV system does inventory inspection tasks in fully autonomous
flights in a warehouse. It is able to fly within a narrow aisle (1.8 m width), move
long distance (55 m∼60 m length), and reach to the highest floor (12 m altitude)
I. INTRODUCTION without any collision, while scanning barcodes of all the stock.
HE unmanned aerial vehicles (UAVs) which have the
T autonomous navigation and environment perception capa-
bility are rapidly arising as promising robotic applications in the
of commercial (non-autonomous) UAVs are not appropriate for
industrial areas such as logistics, transportation, surveillance,
the use at warehouses because safe remote operation is very
etc. Most of current industrial UAVs are limited to outdoor use
difficult even by trained pilots [1]. Some of the drone makers [2],
where GPS data are available. Indoor autonomous navigation
[3] announced to sell drone solutions for inventory tasks, yet no
of UAVs has been actively studied in robotics community, but
articles are found regarding practical applications at warehouses.
there are only a few practical industrial cases.
Fully autonomous UAVs without piloted flights are thought
The inventory inspection using UAVs at warehouses has sev-
to be suitable solutions. Still, safe operation without collision
eral advantages over human workers’ manual inspection, in that
is very challenging because 1) an UAV should fly within very
it can save time and cost when compared with using high-reach
narrow and long aisles between racks as shown in Fig. 1(a) and
forklifts, and avoid injury risk of workers doing manual barcode
(b), 2) it should be close to stocks to read the barcode exactly
scanning on a fast moving high-reach forklift. However, most
as in Fib. 1(c), and 3) the illumination is commonly very poor
in the middle of the long corridors. Several autonomous UAVs
Manuscript received September 9, 2019; accepted November 13, 2019. Date
of publication November 21, 2019; date of current version December 6, 2019. targeting warehouses [1], [4], [5], [6] have been presented, which
This article was recommended for publication by Associate Editor A. Elfes and will be explained in detail in the following section. However,
Editor J. Roberts upon evaluation of the reviewers’ comments. (Corresponding none of them are demonstrated to work well in long and narrow
author: Woong Kwon.)
The authors are with the Factory Automation Technology Team, Global (e.g., 1.8 meters wide) aisles.
Technology Center, Samsung Electronics Co., Ltd., Suwon 16677, Republic of Towards safe and successful inventory inspection at indoor
Korea (e-mail: [email protected]; [email protected]; minsu46.lee@ warehouses, we suggest an Extended Kalman Filter (EKF)-
samsung.com; [email protected]; [email protected]; jawon.
[email protected]). based multi-sensor fusion framework using low-cost sensors:
Digital Object Identifier 10.1109/LRA.2019.2955003 three cameras, a 2D laser scanner, a 1D range sensor, and an
2377-3766 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.
 244 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 1, JANUARY 2020

IMU. We use a simple map which consists of pose information

of attached tags, and so there is no need to do piloted flights for
online map building. Furthermore, we propose three robust data
fusion methods – 1) outlier rejection using Mahalanobis norm’s
component test, 2) addition of visual SLAM to measurement
update by introducing ‘pseudo-covariance,’ and 3) recognition
of floor lanes – to overcome the problems of drift, outliers, and
disturbance of measurements, resulting to enhance the robust-
ness of our multi-sensor fusion framework.
We have been successfully performing regular inventory in-
spection with our autonomous UAV. Results of doing various
inventory task motions are provided to prove the performance
of our methods.
To the authors’ knowledge, such practical applications of
UAVs in a warehouse are the first case in robotics community. Fig. 2. Overall system architecture of our autonomous UAV.
The main contributions of this letter are as follows:
1) We suggest a low-cost multi-sensor system and EKF-
based multi-sensor fusion framework to get optimal pose and UAV, developed for automated inventory management of
estimations. warehouses. They use a 2D lidar, multiple cameras, ultrasonic
2) We propose robust data fusion methods to strengthen the sensors, and adaptive active IR markers. Their results are yet lim-
robustness and to remove the risk of collisions guaran- ited to accuracy test of the UGR-UAV collaborative localization
teeing successful inventory inspections inside materials methods rather than autonomous flights.
warehouse.
3) We build light and simple xml-based maps which are III. SYSTEM OVERVIEW
composed of only poses of pre-installed tags.
We describe the overall architecture of our system, the hard-
4) Our autonomous maneuvering UAV system can be ex-
ware setup, the tag map concept and the multi-sensor fusion
tended to diverse areas requiring autonomous navigation
based localization framework. We also mention why the robust
capability.
data fusion is necessary for complete autonomous flight.

II. RELATED WORK A. Overall Architecture

Recently, several groups have published their results for Fig. 2 shows our system architecture consisting of the sen-
the development of autonomous UAV system aimed at ware- sors, the software algorithm modules connected to the sensors,
house inventory applications. Beul et al. [1] presented high- EKF-based multi-sensor fusion frameworks, and the other extra
performance autonomous inventory UAV for operations inside components. The proposed robust data fusion methods, which
warehouses (funded by German Industry 4.0 project). Their will be clarified in detail in the following section, are shown
sensor setup includes RFID reader, two high-resolution cameras, within the blue dashed rectangle.
and an expensive 3D lidar. They performed experiments in an Acquired data from the sensors are fed into the measurement
operative warehouse with an external warehouse management update module of the EKF in the form of relative/absolute poses,
system and high-level inspection missions. Their method needs angles, or distance, via corresponding algorithms. For instance,
periodic piloted flights to build a 3D feature map, and has the Hector-SLAM [7] provides 6D relative pose(x, y, z, roll, pitch,
step to align the semantic map with the 3D feature map. and yaw) of the UAV and April Tag [8] recognition algorithm
Campos-Macias et al. [4] presented an autonomous nav- gives its 6D absolute pose. More details of the algorithms
igation framework for exploration in unknown 3D cluttered including visual odometry(SLAM) and lane recognition will be
environments. They used an RGBD camera for depth sensing elaborated in the Section IV.
and a tracking camera for visual-inertial odometry. Their demon- The EKF-based multi-sensor fusion framework in conjunc-
stration scenarios include warehouse applications for capturing tion with the robust data fusion methods sends optimally es-
inventory and locating out of place items, while focusing on timated UAV poses to the flight controller and the mission
exploration task to reach a goal. planner. The flight controller unit(FCU) outputs low-level motor
Welburn et al. [5] proposed a navigational system that commands for the quadrotor frame based on the feedback pose
enables UAVs to operate within a dark and GPS-denied data to make stable flights possible. According to the current
environment(warehouse-like) to improve safety for human per- UAV pose estimation, the mission planner updates the waypoint
sonnel. They use two 2D lidars and one camera for marker list and transfers it to the path planner to make path commands
recognition, which are similar to our setup. However, their UAV updated. Those are used as the flight controller’s reference
system has not realized real autonomous flights yet. trajectories. Since the mission planner, the path planner, and
Kalinov et al. [6] presented a high-precision UAV localiza- the flight control [16] is beyond the scope of this letter, we will
tion system interconnecting an unmanned ground robot(UGR) not mention them anymore.

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.
 KWON et al.: ROBUST AUTONOMOUS NAVIGATION OF UAVs FOR WAREHOUSES’ INVENTORY APPLICATION 245

Fig. 3. The sensing system of our autonomous UAV. Fig. 4. April Tags (in yellow circles) attached to the warehouse rack. Their po-
sition and orientation information is saved in an xml-based tag map beforehand
for the autonomous navigation of the UAV.
B. Hardware Setup
For the purpose of robust indoor autonomous navigation, our
Since this tag map is an xml file, its size is very small
UAV’s sensing system is composed of a 2D laser scanner, a 1D
(under 800 KB per warehouse) when compared with 3D feature
range sensor, three cameras(forward, upward and downward),
map. Also there is no need for piloted flights to build a map
and an IMU(Inertial Measurement Unit) as shown in Fig. 3. A
periodically in contrast to [1].
barcode reader is equipped for the inventory barcode scanning.
The 2D lidar(laser scanner) is Hokuyo UST-10LX with
D. Multi-Sensor Fusion Framework
270 degrees field of view and 10 meters range, which is much
cheaper than 3D lidars. The 1D range sensor(lidar), Terabee We suggest a multi-sensor fusion framework to optimally
TeraRanger One, measures the distance up to 14 meters between estimate the 6D pose of the UAV in the indoor(GPS-denied)
the UAV to the ground, to be used as an altitude measurement environments. The framework is based on the EKF where the
through UAV posture compensation. The forward and upward IMU pose is used as state variables of the time update phase -
cameras, mvBluFox-MLC, are 752 × 480 monochrome cameras prediction - and the other sensors’ data are processed in the
with global shuttering and equipped with flat and wide angle measurement update – correction - phase in a modular way [9]
lenses respectively. The downward camera, oCam-1CNG-U, is as depicted in Fig. 2.
1280 × 960 global shuttering color camera. We use the IMU We assume that the pose of the IMU sensor is equivalent to
sensor embedded within Pixhawk3 FCU. All the cameras and the that of the UAV. The state of the EKF consists of the position
sensors are also low-cost under 400 USD. The barcode reader, vector of the IMU piw in the world frame {W}, its velocity vector
i i
Cognex DM363, is an image-based barcode scanner with the vw , its orientation quaternion qw describing the rotation of IMU
maximum range of 500 mm. with respect to W, gyro bias bω and accelerometer bias ba . The
All the sensor drivers, software modules and algorithms run entire state yields a 16-dimensional state vector x:
on a single board computer, Intel NUC7i7BNH with 16 GB
x = [piw T vw q w bw T ba T ] T .
i T i T
(1)
RAM. We use the custom-made quadrotor frames as the flying
mechanism. The horizontal size of the UAV is 74 cm × 71 cm We derive the following differential equations:
and its weight including battery is 3.2 kg.
ṗiw = vw
i
(2)
i i
C. Concept of Tag Map v̇w = Rw (am − ba − na ) − g (3)
The April Tag [8] (abbreviated to “tag” later), which is shown i 1 i
q̇w = Ω(ωm − bω − nω )qw (4)
in yellow circles of Fig. 4, was devised to precisely calculate its 2
relative 3D positions and 3D orientations with respect to the cam- ḃω = nbω (5)
era frame by means of a computer vision algorithm. Conversely,
the forward camera and the tag recognition algorithm in Fig. 2 ḃa = nba (6)
output relative 6D poses (3D positions and 3D orientations) i
where Rw is a rotation matrix corresponding to quaternion
of the camera with respect to the tag frame. Through simple i
qw , am is a measured acceleration from IMU, Ω(ω) is the
coordinate transformation, we have the absolute 6D pose of the
quaternion-multiplication matrix of ω [10], ωm is a measured
UAV with respect to the world coordinate frame if the poses of
angular velocity from IMU, g is the gravity vector in the world
the tags in the world frame are provided beforehand.
frame, and na , nω , nbω , and nba are Gaussian noises. Equations
Prior to the operation, we build an xml-based map storing
(5) and (6) mean that the bias terms are modeled as random
each pose of tags in the world frame, which are attached to the
walk. Taking the expectations of differential equations (2)∼(6),
predetermined locations of the every rack as shown in Fig. 4.
and defining the error state vector Δx of (1), we have a linearized
Thus, the absolute 6D pose data of the UAV calculated from tag
error state equation
recognitions are used as one of the measurement update of the
EKF, once tags are observed. Δẋ = AΔx + Bn (7)

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.
 246 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 1, JANUARY 2020

Fig. 5. Images from the forward camera showing the recognition of an April
Tag while the UAV is flying horizontally. As the tag is seen far from the image
center, the possibility of wrong detection (i.e., outlier) increases.

where A and B are linearized matrices of (2)∼(6)(See [11] for

details). Consequently, we are able to obtain the discrete time
Fig. 6. The effect of rejecting an outlier from tag recognition in yaw angle
update process of the EKF from (7) [11]. estimation.
The measurement model relates the state vector and the sensor
measurement as
z = h(x) + n (8) As the tag is detected farther away from the image center, the
possibility of occurrence of tag detection outliers increases. We
where z is a measurement vector such as position, Euler angle, found that in our warehouse outliers of tag recognition also
and pose, h(x) is the corresponding measurement model and n is appeared due to bad attachments of the tags and bending of
Gaussian measurement noise. We employ Lynen’s [9] modular the tag plates. Besides, the 1D range sensor and the 2D laser
multi-sensor fusion framework where individual sensors’ mea- scanner also generate outliers.
surement updates are fused independently to get the optimal In order to reject such outliers of measurements [9], we cal-
estimation of the state. It is advantageous that we are able to add culate and monitor the Mahalanobis norm M of a measurement
our own more sensors to this framework. update
A minimal sensor suite for the valid measurement update of
our EKF is a camera which measures the pose of tags attached M 2 = (z − H x̂)T S −1 (z − H x̂) (9)
to the rack frames providing the 6D absolute pose, a 2D laser
where z is the measurement vector, H is the linearized mea-
scanner plus IMU which calculate the 6D relative pose from scan
surement matrix of h(x) in (8), S is the state covariance matrix,
matching techniques like Hector-SLAM [7], and a downward 1D
and x̂ is the state prediction. We propose that specific com-
range sensor which estimates the altitude of the UAV.
ponents of this norm - for example, yaw/pitch angles from
Theoretically, this concept should guarantee sufficiently con-
tag recognition or lateral positions from Hector-SLAM - be
sistent results because complete 6D poses are always available
individually checked during mission flights. If some of them
by combining the absolute and the relative pose measurements.
are larger than a predetermined threshold, our EKF framework
However, in practice, this is not the case, since there inevitably
will reject the corresponding component of the measurement.
exist drift, outliers, and disturbance of measurements due to
This component-based Mahalanobis norm test reduces the false
system uncertainty and incomplete sensing data. Moreover,
rejection ratio by separating a measurement vector into smaller
warehouse environments commonly shows very challenging
dimensional vectors.
situations for vision-based autonomous navigation, because the
Fig. 6 shows an exemplar result of tag’s yaw component
illumination is very poor in the middle of the rack, the corridors
outlier rejection. Unless the outlier is rejected, the multi sensor
are only 1.8 meters wide, and the rack is so long (55∼60 meters)
fusion output (yellow line) is deviated largely as in Fig. 6(a)
that a small errors or drifts may cause collisions. Thus, the above
because the absolute pose from “relative” Hector-SLAM data
sensing system does NOT make practical applications of au-
(not shown in this figure) is recalculated based on the recent
tonomously maneuvering UAVs possible – requiring extremely
tag measurements. However, if the outlier is rejected from the
high safety – without addressing such problems.
measurement update, the deviation of the EKF output due to the
outlier disappeared as in Fig. 6(b).
IV. ROBUST DATA FUSION
We suggest three approaches as the robust data fusion methods B. Pseudo-Covariance to Utilize Visual SLAM
to increase the robustness of our EKF, in such a way as to
To minimize the disturbance effects caused by the laser
overcome the drift, outliers, and disturbance issues. These help
scanner’s Hector-SLAM – usually occurring during vertical
the multi-sensor fusion framework to cope with the challenging
movement, we add a visual SLAM algorithm (the state-of-the-art
warehouse environments successfully.
is ORB-SLAM2 [12]) implemented by using the upward camera
capturing images toward the ceiling. The x, y, and z position;
A. Outlier Rejection by Mahalanobis Norm Test and yaw orientation data from this algorithm are fused into
The outliers usually occur from sensor faults, large recogni- EKF to weaken disturbance effects from Hector-SLAM or other
tion errors or bad measurement environments. Fig. 5 represents unknown sources.
a typical scenario of outliers in tag recognition, where the Since visual SLAM algorithms like ORB-SLAM2 are mostly
sequence of images during horizontal flying motion is listed. keypoint-based, they do not provide the covariance of the output,

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.
 KWON et al.: ROBUST AUTONOMOUS NAVIGATION OF UAVs FOR WAREHOUSES’ INVENTORY APPLICATION 247

Fig. 7. Images from the upward camera showing the ORB SLAM2 results.
The numbers of matching points in motion model are (a) 288 and (b) 47. As Fig. 9. Images from the downward camera showing the results of lane
the number decreases, the uncertainty of pose estimation tends to increase. recognition.
Therefore, we use this number to calculate the pseudo-covariance.

ratio of the physical gap of both lanes to their gap on the image
plane. The latter is used as yaw measurements of the UAV.
These two absolute measurements are very helpful to remove
the influences of drift of the estimation. This lane recognition is
especially effective in flights at the 2nd and 3rd floor altitude and
in the middle of corridors.

V. RESULTS IN A WAREHOUSE APPLICATION

Fig. 8. The effects of applying pseudo-covariance of ORB SLAM2. Using our EKF-based UAV localization including the three
methods to enhance robustness, we have been conducting vari-
which is required for the measurement update phase of EKF. We ous inventory inspection tasks, actually accomplishing safe au-
propose to use the number of matching points in motion model, tonomous navigation. Regarding the poses from tag recognition
as depicted in Fig. 7, to extract the uncertainty information as ground truth (approximately), we estimated the localization
of the ORB-SLAM2 - namely ‘pseudo-covariance’. Through accuracy of the proposed methods. The average and maximum
repetitive experiments and tuning process, we obtained the localization error is 3.12 cm and 25.68 cm, respectively, in a full
best proportional constant and offsets between the number of inspection task of a single rack (6 floors and 55 m long), which
matching points and the pseudo-covariance that works well. are sufficient for secure operations.
By exploiting pseudo-covariance of ORB-SLAM2, we can The proposed methods prevent the UAV from colliding into
apply its measurement update into EKF to reduce the effects near racks or walls by removing the drift of naïve localization
of disturbance. Fig. 8 illustrates the effects of applying ORB- methods. Outliers of tag recognition, occurring from bad at-
SLAM2. The large deviation of the multi-sensor fusion output tachments, bended tag plates, or recognition faults, were clearly
in Fig. 8(a) is effectively reduced in Fig. 8(b), as shown within rejected to keep the estimation from serious deviation. The
the red circles. bad effects of frequent disturbance of Hector-SLAM appearing
One more thing that should be kept in mind is that the mainly in ascending/descending phases were also minimized
ORB-SLAM2 often loses the tracking if there are not enough in the estimation results, consequently showing stable flights.
matching points. In that case we make the algorithm to try Moreover, the map building process was very simple and cost
to restart the tracking quickly. Since we use the monocular effective since our map is only composed of marker tags’ pose
camera, we should get the scale factor of the ORB-SLAM2 data. On the contrary, the approaches using 3D feature map, as
output data after the tracking starts as soon as possible. For the in [1], typically need big map files and periodic map building
calculation of the scale factor, we employ Maximum Likelihood steps through manual flights to reflect consecutively changing
Estimation(MLE)-based method offered by Engel et al. [13] piled boxes in a warehouse.
since its calculation considers the uncertainty information and Fig. 10 reveals a comparison result to verify the improvements
so gives consistent results. effects of the proposed robust data fusion methods. During an
ascending motion between floors, the previous UAV frequently
shows large deviating motions approaching to a rack as shown
C. Recognition of the Lane on the Floor in Fig. 10(a) due to drift or disturbance of the localization. This
To get rid of the drift effects along yaw and lateral direction situation is very dangerous because it may cause collisions into
in the rack aisles, we add a downward color camera to recognize the near rack or stocks. After applying the proposed methods, the
the lane information so as to extract lateral position and yaw UAV is able to maintain the center path safely as in Fig. 10(b).
orientation as in Fig. 9. Note that the yellow dotted lines are drawn along the vertical
By extracting edge features of the lane through Fast Fourier image center line for the easy visual understanding.
Transform (FFT) and likelihood computation, we obtain the Typical inventory inspection tasks are categorized into two
translational distance between the lane and image center lines, scenarios. One is the full inspection of stocks at the whole racks
and the twisting angle of the lane. The former can be interpreted of a warehouse. In our local warehouse, there are 32 racks in
as measurements of the lateral position by multiplying the scale total for stock storage. Each rack is 55 meters ∼60 meters long

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.
 248 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 1, JANUARY 2020

Fig. 10. An exemplar motion demonstrating the improvement of using the robust data fusion methods. While the previous UAV had shown a large disturbance
motion approaching to the rack dangerously (the 2nd red circle), the current UAV maintains the center path safely. The yellow dotted lines are drawn along the
vertical image center line for the visual understanding.

Fig. 11. A variety of tasks for inventory inspection conducted by our UAV system at the warehouse. For the easy visual understanding, we reconstruct the voxel
map [14], [15] and the path of the UAV from the data gathered in flights. The colors of the voxels reflect their altitude, and the black trajectory is the UAV’s
estimated pose. Note that, since a 2D laser scanner is used for environmental mapping, the 3D voxel map is created only within the laser scanner’s field of view.

and 6 floors high, and has 180 bins for pallets. Full inspection task motions done by our autonomous UAV targeting to bins at
means that all the bins of the racks are scanned at once. Common the 3rd floor. The UAV may start cycle counting tasks by taking
warehouses do full inspection periodically for complete inven- off at a place within the corridor in front of the target rack as in
tory management. Fig. 11(a) depicts a full inspection task of a Fig. 11(b), and at a home position outside of the aisles to reach
single rack performed by our autonomous UAV. In Fig. 11, for multiple racks in turn as shown in Fig. 11(c).
the intuitive understanding, we reconstruct the 3D voxel map With our autonomous UAV, we have been doing both full
[14], [15] and the path of the UAV from the data gathered in inspection and cycle counting tasks successfully in the materials
real autonomous flights. The colors of the voxels reflect indicate warehouse without any problems for one year, replacing the con-
their altitude, and the black trajectory is the estimated pose of ventional manual inventory inspection done by human workers.
the UAV. Since the environmental mapping is done by the 2D Fig. 12 shows several snapshots captured during a variety of the
laser scanner, the 3D voxel map is created only within the laser inventory inspection motions of our autonomous UAVs. Prior
scanner’s field of view. manual inspection tasks typically required high-reach forklifts,
The other category is cycle counting which is to count items its licensed driver, and one or two workers on board holding a
in specific bins within the warehouse without having to count hand-held barcode scanner. The application of the autonomous
the entire inventory. Fig. 11(b) and (c) describe cycle counting UAV enabled the warehouse to save such large equipment costs

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.
 KWON et al.: ROBUST AUTONOMOUS NAVIGATION OF UAVs FOR WAREHOUSES’ INVENTORY APPLICATION 249

Fig. 12. Photo captures of our UAV system while doing various motions for the inventory inspection tasks at the warehouse.

and human labors efficiently. The inspection time is also reduced [3] DroneScan, 2019. [Online]. Available: www.dronescan.co
since it takes only 9 minutes and 14 seconds to complete a [4] L. Campos-Macias et al., “Autonomous navigation of MAVs in unknown
cluttered environments.” [Online]. Available: https://arxiv.org/abs/1906.
full inspection of a single rack containing 180 bins. One of 08839
the most important consequences of our autonomous UAV is [5] E. Welburn et al., “A navigational system for quadcopter remote inspection
to get rid of the risk of human workers’ unpredicted fatal injury of offshore substations,” in Proc. 15th Int. Conf. Autonomic Auton. Syst.,
2019.
while doing barcode scanning on a vertically moving high-reach [6] I. Kalinov, E. Safronov, R. Agishev, M. Kurenkov, and D. Tset-
forklift. serukou, “High-precision UAV localization system for landing on a mo-
bile collaborative robot based on an IR marker pattern recognition,” in
Proc. IEEE 89th Veh. Technol. Conf., Kuala Lumpur, Malaysia, 2019,
VI. CONCLUSION pp. 1–6.
[7] S. Kohlbrecher et al., “Hector open source modules for autonomous
We proposed an autonomous UAV with a low-cost sensing mapping and navigation with rescue robots,” in Robot Soccer World Cup.
system and multi-sensor fusion framework to be used effectively Berlin, Germany: Springer, 2013, pp. 623–631.
for narrow and dark warehouse environments. To address the [8] J. Wang and E. Olson, “AprilTag 2: Efficient and robust fiducial detection,”
in Proc. IEEE/RSJ Int.Conf. Intell. Robots Syst., 2016, pp. 4193–4198.
problems of naïve UAV localization methods, we suggested [9] S. Lynen, M. W. Achtelik, S. Weiss, M. Chli, and R. Siegwart, “A robust
robust data fusion methods: outlier rejection using component- and modular multi-sensor fusion approach applied to MAV navigation,”
based Mahalanobis norm test, incorporation of visual SLAM by in Proc. IEEE/RSJ Int.Conf. Intell. Robots Syst., Tokyo, Japan, 2013,
pp. 3923–3929.
introducing pseudo-covariance, and recognition of floor lanes [10] N. Trawny and S. Roumeliotis, “Indirect Kalman filter for 3D attitude
for absolute lateral position and yaw measurements. These allow estimation,” Dept. Comput. Sci. Eng., Univ. Minnesota, Minneapolis, MN,
us to perform safe and fully autonomous flights for the cyclic USA, Tech. Rep. 2005-002, 2005.
[11] S. Weiss, “Vision based navigation for micro helicopters,” Ph.D. disser-
inventory inspections at our materials warehouse. As a fully tation, Autonomous Syst. Lab, Dept. Mech. Process Eng., ETH Zürich,
autonomous system, our method does not need piloted flight, Zürich, Switzerland, 2012.
and also reduces the cost, time and the risk of human’s high-floor [12] R. Mur-Artal and J. D. Tardós, “ORB-SLAM2: An open-source SLAM
system for monocular, stereo, and RGB-D cameras,” IEEE Trans. Robot.,
tasks. For further study, we plan to improve the system reducing vol. 33, no. 5, pp. 1255–1262, Oct. 2017.
the cost more, and extend our method to various indoor and [13] J. Engel, J. Sturm, and D. Cremers, “Scale-aware navigation of a low-
outdoor applications. cost quadrocopter with a monocular camera,” Robot. Auton. Syst. vol. 62,
pp. 1646–1656, 2014.
[14] H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, and J. Nieto, “Voxblox:
REFERENCES Incremental 3D euclidean signed distance fields for on-board MAV plan-
ning,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Vancouver, BC,
[1] M. Beul, D. Droeschel, M. Nieuwenhuisen, J. Quenzel, S. Houben, and Canada, 2017, pp. 1366–1373.
S. Behnke, “Fast autonomous flight in warehouses for inventory applica- [15] A Hornung et al., “OctoMap: An efficient probabilistic 3D mapping
tions,” IEEE Robot. Autom. Lett., vol 3, no. 4, pp. 3121–3128, Oct. 2018. framework based on octrees,” Auton. Robots, vol. 34, pp. 189–206, 2013.
[2] EyeSee, 2019. [Online]. Available: eyesee-drone.com [16] Pixhawk, 2019. [Online]. Available: pixhawk.org

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:32:00 UTC from IEEE Xplore. Restrictions apply.