Reducing Object Detection Uncertainty from RGB and Thermal Data

for UAV Outdoor Surveillance

Juan Sandino ‡ , Peter A. Caccetta ‡ , Conrad Sanderson †‡ , Frederic Maire  , Felipe Gonzalez 

 Queensland University of Technology, Australia

† GriffithUniversity, Australia
‡ Data61 / CSIRO, Australia

Abstract—Recent advances in Unmanned Aerial Vehicles (UAVs) gusts, unknown situational-awareness of surveyed environments,
have resulted in their quick adoption for wide a range of civilian and partial observability. Internal factors include sub-optimal
applications, including precision agriculture, biosecurity, disaster camera calibration settings, low image resolution, noisy camera
monitoring and surveillance. UAVs offer low-cost platforms with
flexible hardware configurations, as well as an increasing number frames during streaming, or imperfect detection outputs from
of autonomous capabilities, including take-off, landing, object computer vision detectors. As shown in Fig. 1, uncertainty
tracking and obstacle avoidance. However, little attention has sources that are poorly managed can compromise the behaviour
been paid to how UAVs deal with object detection uncertainties of UAVs and the flight mission itself [21]. Thus, it is essential
caused by false readings from vision-based detectors, data noise, to incorporate cognitive capabilities in UAVs to broaden their
vibrations, and occlusion. In most situations, the relevance
and understanding of these detections are delegated to human use in more real-world scenarios [12].
operators, as many UAVs have limited cognition power to interact The elevated number of stranded people and human loss is
autonomously with the environment. This paper presents a a problem that is far from solved [1]. In Australia alone, an
framework for autonomous navigation under uncertainty in average of 38,000 people per year are reported missing and
outdoor scenarios for small UAVs using a probabilistic-based around 2% of them (or 720 persons) are never located [4]. In
motion planner. The framework is evaluated with real flight tests
using a sub 2 kg quadrotor UAV and illustrated in victim finding the event of an emergency—where time management plays
Search and Rescue (SAR) case study in a forest/bushland. The a critical factor in the success of the rescue operation—the
navigation problem is modelled using a Partially Observable goal to identify and locate as many victims as quick as
Markov Decision Process (POMDP), and solved in real time possible. Thus, UAV technology for autonomous navigation
onboard the small UAV using Augmented Belief Trees (ABT) and victim detection in challenging environments could assist
and the TAPIR toolkit. Results from experiments using colour
and thermal imagery show that the proposed motion planner first-responders in locating as many victims as soon as possible.
provides accurate victim localisation coordinates, as the UAV has Research works on applied decision-making theory in UAVs
the flexibility to interact with the environment and obtain clearer is extensive and indicates that using Partially Observable
visualisations of any potential victims compared to the baseline Markov Decision Processes (POMDPs) onboard UAVs can
motion planner. Incorporating this system allows optimised UAV increase their cognitive capabilities for autonomous navigation
surveillance operations by diminishing false positive readings from
vision-based object detectors.
I. Introduction
Recent advances in autonomous navigation of Unmanned
Aerial Vehicles (UAVs)—also known as drones—have resulted
in their gradual adoption in a set of civilian and time-critical
applications such as surveillance, disaster monitoring, and
Search and Rescue (SAR) [11], [20], [24], [25]. UAVs offer
unique benefits such as compact sizes and low cost to scout out-
door and indoor environments, real-time telemetry and camera
streaming to monitor challenging and otherwise inaccessible
environments, extensive payload adaptability, and extensive
possibilities to augment navigation capabilities through soft-
ware [8], [16], [26], [31].
One critical challenge in deploying UAVs and robots in
general into real-world and time-critical applications is the
ever-presence of uncertainty. Factors that cause uncertainty
Fig. 1. Unmanned aerial vehicle (UAV) navigating in environments under
are diverse, and they can be classified as external or internal. uncertainty and partial observability. A small UAV with autonomous decision-
External factors come from sources beyond the scope of the making should be able to plan sequential sets of actions for optimal navigation
UAV, such as poor weather and illumination conditions, strong trajectories, despite limitations from imperfect sensor data.

∗
Published in: IEEE Aerospace Conference, 2022. DOI: 10.1109/AERO53065.2022.9843611
and object detection under uncertainty [5], [14], [27]. UAV II. Framework Design
frameworks for object detection and tracking have been The framework follows a modular system architecture for
tested in cluttered indoor environments and in the absence autonomous navigation onboard small UAVs as illustrated in
of Global Navigation Satellite System (GNSS) coverage [31], Fig. 2. This design extends an existing UAV framework for
[32], [34], [35]. POMDPs have also been applied to solve multi- autonomous navigation in cluttered environments under object
objective problems in UAVs, addressing tasks such as path detection uncertainty, tested in simulation and with real flight
planning, multiple object detection and tracking, and collision tests in a sub 2 kg quadcopter [32].
prevention [28], [29]. Fig. 2 illustrates the physical environment (or world) com-
In time-critical applications such as SAR, real-time camera posed by the UAV frame and any attached payloads (i.e.,
streaming is critical to comprehend the context of the envi- RGB or thermal cameras), the victim and obstacles. Acquired
ronment [22]. However, drone pilots have a strong reliance camera frames represent the visual interface (also called
on their communication systems to control most UAVs. If observations) of the surveyed environment by the UAV. The
communication systems fail, the usability of the UAV could UAV also contains the autopilot, which translates high-level
be seriously compromised [33]. Many approaches of POMDPs action commands into low-level signals that control the UAV
applied in UAVs for humanitarian relief operations have been motors. The last hardware component of the UAV frame is a
tested in simulation [3], [36] and very few systems have been companion computer, which is allocated to execute software
evaluated with real flight test using trivial targets [10]. algorithms in dedicated modules for computer vision, mapping,
Research efforts on onboard decision-making under object and real-time path planning. Action commands from the planner
detection uncertainty from Convolutional Neural Network are managed by the motion module, which interfaces with the
(CNN) models are scarce. Research conducted in [31], [32] flight controller of the autopilot.
described a framework and POMDP problem formulation for The following subsections discuss each of the proposed
a SAR application in GNSS-denied environments with a sub framework components. The UAV framework used in this work
2 kg UAV. However, the framework was only tested in cluttered is not limited to the hardware and software discussed below.
indoor scenarios. Other UAV frame designs, payloads, autopilots, vision-based
This paper describes a modular UAV framework for au- object detectors, planners, and software toolkits can also be
tonomous onboard navigation in outdoor environments under implemented.
uncertainty. The framework design aims to reduce levels of
object detection uncertainty using a POMDP-based motion A. UAV Airframe and Payloads
planner, which allows the UAV to interact with the environment The UAV airframe which offered the best combination
to obtain better visual representations of detected objects. CNN- between payload adaptability, size, and endurance for this
based computer vision inference and motion planning can be research is a Holybro X500 quadrotor kit (Holybro, China). As
executed in resource-constrained hardware onboard small UAVs. shown in Fig. 3, key components utilised from the kit include a
The framework is tested with real flight tests with a simulated Pixhawk 4 autopilot, Pixhawk 4 GNSS receiver, 2216 KV880
SAR mission, which consisted of finding an adult mannequin brushless motors, 22.86 cm plastic propellers, and a 433 MHz
in an open area and close to a tree. Three flight modes are Telemetry Radio. With dimensions of 41 cm × 41 cm ×
proposed to evaluate the feasibility of the framework for real- 30.0 cm, the UAV carries a four cell 5000 mAh LiPo battery,
world SAR operations. for an approximate flight autonomy of 12 min.
This paper extends the work in [30], [31], [32] with the The companion computer is an Intel UP2 , chosen for its
following contributions: (1) an extension of their evaluated price tag, number of peripherals and CPU architecture. Key
UAV framework—originally designed for navigation in GNSS- specifications include a 64-bit quad-core CPU at 1.1 GHz, 64
denied environments— for outdoor missions with GNSS signal GB eMMC SSD, 8 GB RAM, four FL110 USB 3.0 connectors,
coverage, and the design of a novel flight mode; (2) an two High-Speed UART controllers, and one mPCIe connector.
additional validation of preliminary results of their proposed The proposed framework was tested using two Red, Green,
UAV framework with comprehensive real flight tests; and (3) Blue (RGB) cameras, namely an Arducam B019701 and a
a scalability approach of the framework by adapting a thermal GoPro Hero 9. Thermal imagery is sourced from a FLIR Tau
camera and a custom object detector to locate victims using 2 connected to a ThermalCapture device for real-time frame
their heat signatures. streaming. The cameras, which can be interchangeably used in
The rest of the paper is structured as follows. Section II the proposed framework, are mounted onto an anti-vibration
details the UAV framework design for autonomous object bracket, pointing to the ground and in parallel to Earth’s nadir,
detection in uncertain outdoor environments. Section III sum- as seen in Fig. 4. Core properties for the cameras can be found
marises the implemented probabilistic-based motion planner in Tab. IV in the Appendix.
using a POMDP. The design of conducted experiments using
real flight tests is presented in Section IV. Obtained results and B. Vision Module
discussion of performance indicators are provided in Section V. This module consists of a deep learning object detector
Conclusions and future avenues for research are discussed in processing raw frames from the GoPro Hero 9 camera. Taking
Section VI. into account the performance limitations of running deep
 Companion Computer Environment

Agent
A CNN Object Detector
A

Vision Processing Obstacles

Unit (VPU)

Victim Detection UAV Frame Camera

Outputs Victim

Vision Module
Local Position
Estimator IMU /
GNSS
Compass

Observation Server Depth

Flight Controller B
Camera
Flight Management Unit Sensors
POMDP Motion
Planner
Autopilot
Motors
Action Commands
Operator
Planner Module
Flight Mode
Switch
3D Occupancy Map
B Motion Server
Mapping Module Ground Control Station
Motion Module

Fig. 2. Modular system architecture for autonomous navigation onboard UAVs in uncertain outdoor environments. The framework portrays the physical
environment (or world) composed of the UAV frame, attached payloads, world obstacles, and the victim. A companion computer is attached to the UAV to
execute software algorithms in dedicated modules for computer vision, mapping, real-time path planning, and a motion server that interfaces the companion
computer with the UAV autopilot.

2
3

4
5

Fig. 3. Framework implementation in a sub 2 kg quadrotor UAV. Primary

components include: (1) carbon fibre Holybro X500; (2) Pixhawk 4 GNSS
receiver; (3) Pixhawk 4 autopilot; (4) 433 MHz telemetry radio; (5) Intel UP2
companion computer; (6) payload.
Fig. 4. GoPro Hero 9 mounted onto an anti-vibration bracket, pointing to the
ground and in parallel to Earth’s nadir.
learning models in resource-constrained hardware, a Vision III. Planner Design
Processing Unit (VPU) is installed in the companion computer. This approach formulates the decision-making problem as a
Convolutional operations that normally run onboard a CPU or POMDP. The planner transmits UAV position commands to the
GPU are allocated to the VPU for inference of CNN models in motion planner derived from environment observations. The
resource-constrained hardware. In this implementation, the discussion presented in this section is adapted from [30] and
selected VPU is an Intel Movidius Myriad X, which is only essential parts are shown in this paper for completeness.
connected to the companion computer via the mPCIe slot. With a taken action a ∈ A, the UAV receives an observation
The detection module is programmed in Python and uses the o ∈ O encoded by the observation function Z(s0 , a, o) =
OpenVINO library to optimise code instructions to load CNN P (o | s0 , a). Every decision chain is then quantified with
models into the VPU. an estimated reward r, calculated using the reward function
The deep learning model architecture used to detect victims R(a, s). A POMDP uses a probability distribution over the
is an off-the-shelf Google MobileNet Single-Shot Detector system states to model uncertainty of its observed states. This
(SSD) [7]. This model is deployed in Caffe [15] and tuned with modelling is called the belief b(H) = P[s1 | H], · · · , P[sn |
pre-trained weights from the PASCAL VOC2012 dataset [9], H], where H is the history of actions, observations and
scoring a mean average precision of 72.7%. The dataset covers rewards the UAV has accumulated until a time step t, or
up to 21 class objects (including persons). However, only H = a0 , o1 , r1 , · · · , at−1 , ot , rt .
positive detections for the class person are evaluated. Acquired The motion policy π of the UAV is represented by mapping
camera frames are fitted into the input layer of the neural belief states to actions π : b → A. A POMDP is solved after
network (i.e., MobileNet SSD model) by downsizing the frames finding the optimal policy π ∗ , calculated as follows:
to 300 × 300 pixels. "∞ #!
X
∗ t
C. Mapping Module π := arg max E γ R (St , π (bt )) (1)
π
The Mapping module manages 3D occupancy maps, which t=0

are constituted by volumetric occupancy grids and displays where γ ∈ [0, 1] is the discount factor and defines the relative
the presence and localisation of objects in the surveyed importance of immediate rewards compared to future rewards.
environment. In this implementation, the 3D Occupancy Map A given POMDP solver starts planning from an initial belief
are requested by the Motion and Planner modules to evaluate b0 , which is usually generated using the initial conditions (and
the presence of obstacles at selected position coordinates in assumptions) of the flight mission.
the world coordinate frame. The maps are created through the
A. Assumptions
use of the Octomap library [13].
In this implementation, the formulated problem for explo-
D. Planner Module ration and object detection (i.e., victims) using multi-rotor
The planner module computes the motion policy of the UAV UAVs in outdoor environments assumes:
and contains three primary components: (1) an observation • An initial 3D occupancy map of the environment is pre-loaded to
server, which handles raw observations from the vision module the planner before the UAV takes off.
(i.e., detected victims, confidence and victim coordinates), • Observations come from processed camera frames (by the Vision
local position estimations of the UAV, and the state of the module), the 3D occupancy map (by the Mapping Module), and
3D occupancy map; (2) the POMDP motion planner, which the estimated local UAV position (by the Autopilot).
calls the observation server every time the planner requires • Only a single, and static victim can be detected at the same time.
new observations; and (3) action commands computed by the If more victims appear on processed camera frames, the planner
will only read data from the victim with the highest detection
motion policy of the planner and are read by the motion server. confidence values.
Complete details of the POMDP planner design can be found
• The motion planner starts once the UAV reaches known position
in Section III. setpoint (i.e., at one of the corners of the surveyed area).
E. Communication Interface • The planner stops computing a motion policy once: (1) the UAV
The UAV framework runs on open-source software tools. detects a victim whose detection confidence surpasses a set
threshold; (2) the UAV explores the whole search area extent
The companion computer runs under Linux Ubuntu 18.04 LTS without finding any victims; or (3) the UAV exceeds the maximum
O.S. and the Robot Operating System (ROS) melodic. ROS is a flight time on air (because of low levels of battery power).
middleware to communicate between the nodes of each module
(following the architecture design from Fig. 2). The Pixhawk 4 B. Actions
autopilot is powered by PX4, which communicates with the UAV actions are defined by seven position commands,
companion computer through MAVROS, a ROS implemen- namely forward, backward, left, right, up, down, and hover.
tation of the MAVLink protocol, which is industry standard UAV actions that are not included in the action space but are
for UAV communication and control [18]. The POMDP solver managed by the autopilot instead include arm, disarm, take-off,
implementation, which is described in Section III, also contains return to launch and land. Each action updates the position set
a ROS implementation to maximise the use of visualisation, point of the UAV in the world coordinate frame by calculating
telemetry and recording tools from ROS. and applying a change of position δ.
 TABLE I Algorithm 1 Reward function R for exploration and object
Applied reward values to the reward function R, defined in detection in outdoor environments.
Algorithm 1. 1: r ← 0
Variable Value Description 2: if fcrash then

rcrash −50 Cost of UAV crash 3: r ← rcrash . UAV crashing cost

rout −25 Cost of UAV breaching safety limits 4: else if froi then
rdtc +25 Reward for detecting potential victim 5: r ← rout . Beyond safety limits cost
rconf +50 Reward for confirmed victim detection 6: else if fdct then
raction −2.5 Cost per action taken 7: r ← rdtc h . Detected object reward
rfov −5 Footprint overlapping cost 
zu -zmin
i
8: r ← r + rdtc · 1 - zmax -zmin . UAV altitude reward
9: if cv ≥ ζ and a = Down then
The magnitude for δx and δy depends on the estimated 10: r ← r + rconf
overlap value between camera frame observations: 11: end if
12: else
δ = lFOV (1 − λ) (2) 13: r ← ractionh . Action cost
 i
where lFOV is the length of the projected camera’s Field of 14: r ← r − rdtc · 1 - zzmax
u - zmin
- zmin i . UAV altitude cost
View (FOV), and λ ∈ [0, 1) is the desired overlap value.
h 
15: r ← r − rdtc · 1 - 0.5 4·dv /dw . Horizontal
C. States distance cost
A system state s ∈ S is defined as: 16: r ← r + rfov · ε . Footprint overlap cost
17: end if
s = (pu , froi , fdct , pv , cv ) (3) 18: return r

where pu is the position of the UAV in the world coordinate

frame, fcrash is a flag raised when the UAV crashes with an
obstacle, froi is a flag indicating whether the UAV is flying victim detections. If the UAV detects a potential victim (Step 6),
beyond the flying limits, fdct is the flag raised if a potential R calculates a linear function (Step 8) which returns increased
victim is detected by the UAV. If fdct = True, the position of reward values as the UAV gets closer to the minimum allowed
the victim in the world coordinate frame is given in pv , with altitude. A higher reward value is returned if a potential victim
detection confidence cv ∈ [0, 1]. The system reaches a terminal is confirmed (Step 9 and 10).
state whenever cv ≥ ζ, where ζ is the confidence threshold.
The second component of the algorithm addresses explo-
D. Transition Function ration. In case there are no detections, R applies a set of cost
The motion dynamics of a multi-rotor UAV defines the functions to encourage a greedy horizontal exploration of the
transition from current to new states: environment. An exponential function in Step 14 calculates the
Manhattan distance between the UAV and the victim dv and
pu (k + 1) = pu (k) + ∆pu (k) (4) the maximum exploration distance dw which are defined as:
where pu (k) is the position of the UAV at time step k, and
Xn
∆pu (k) is the position change of the UAV between time steps. dv = |pi − qi |, pi =(xu , yu ), qi =(xv , yv ) (5)
This formulation does not contain any actions for heading Xi=1
n
changes. However, Eqn. 4 can be expanded if required by dw = |pi − qi |, pi =(xmax , ymax ), qi =(xmin , ymin ) (6)
i=1
adding the rotation matrix in multi-copters [6]. An illustration
of a problem formulation including the rotation matrix can be The overlap ε between the camera’s current footprint and
found in [31]. its correspondent location in the footprint map is defined as:
Pn
Fi (pu )
E. Reward Function ε = i=1 , ε ∈ [0, 1] (7)
n
The expected reward r after taking an action a ∈ A from
state s ∈ S is calculated using the reward function R(a, s) where Fi (pu ) are the pixel values of the projected FOV in the
defined in Algorithm 1 and Table I. This function critically footprint map, and n is the total number of projected pixels
influences the UAV behaviour during flight missions, and its in the footprint. A maximum overlap value of 1 indicates that
definition allows multi-objective task definition. A complete such action will place the UAV to a fully previously explored
discussion on the design considerations of the reward function area, and, as indicated in Step 16 of Algorithm 1, the whole
can be found in [30]. penalty value rfov will be added to the reward. A minimum
The order of the steps from Algorithm 1 classifies high-level value of 0 means that a given action will place the UAV in an
tasks into two components. The first one is object detection unexplored area and no penalty will be added to the reward.
and starts by evaluating any states which will negatively affect Intermediate values of ε represent partial overlapping, adding
the integrity of the UAV, followed by states indicating positive a partial penalty value rfov to the reward.
F. Observations adjusting the heading of the UAV mid-flight, and assuming
yaw estimation errors are negligible, Eqn. 12 is simplified as:
An observation o ∈ O is defined as:  0   0 
c (x) spu (x) + c(x)
o = (opu , odtc , opv , oζ , oobs ) (8) = (13)
c0 (y) s0pu (y) + c(y)
where opu is the estimated position of the UAV by the autopilot; The detection confidence oζ that comes as part of the output
odtc is the flag triggered by potential victim detections received data from the CNN object detector is modelled using:
by the CNN model; opv and oζ are the local position of the
victim and the detection confidence respectively, both of them (1 − ζmin ) (duv − zmin + ζmin )
oζ = (14)
defined only if there are any positive detections; and oobs is zmax − zmin
the flag triggered after processing the 3D occupancy map for
where ζmin is the minimally accepted confidence threshold, zmax
any obstacles located in front of the UAV.
and zmin are the maximum and minimum UAV flying altitudes
The detection confidence oζ measures the frequency of respectively, and d is the Manhattan distance between the
uv
positive detections between the last two observation calls: UAV and the victim.
Pn
odtci
oζ = i=1 (9) IV. Experiments
n
where n is the number of segmented frames between observa- This research validated the proposed UAV framework with
tion calls, and odtc is the flag indicating a positive detection real flight tests on the sub 2 kg quadcopter shown in Sec-
per processed frame i. tion II-A. The tests were designed under a ground SAR
application context, specifically, to locate a lost person last seen
G. Observation Model around a forest/bushland area. The subsections below present
This implementation uses Augmented Belief Trees (ABT) the location of conducted flights, environment setup, proposed
[19], an online POMDP solver that contains a model that flight modes for data collection and tuned hyperparameters of
generates T and Z using a modelled observation o given an the online POMDP solver.
action a and the next state s0 . The variables contained in the
A. Location and Environment Setup
generative model are the local position of the UAV s0pu , the
local position of the victim s0pv and the detection confidence oζ . Flight tests were conducted at the Samford Ecological
Potential victim detections and their subsequent position- Research Facility (SERF), 148 Camp Mountain Road, Samford
ing estimations are conditioned by the camera pose at the QLD 4520, Australia. As shown in Fig. 6, the 51 hectare
UAV frame and its projected footprint of the environment. property contains protected Dry Sclerophyll forest and grazing
Specifically, if the 2D local position coordinates of the victim zones, where the latter ones were utilised to fly the UAV.
s0pv (x, y) are within the projected footprint limits of the camera, The delimited flying area covers a mostly flat grazing
the victim is assumed to be detected. This estimation is done by zone featuring buffel grassland, a five-metre tree, and a car
calculating the sum of angles between a 2D point (i.e., s0pv ) and purposely placed as an additional obstacle. Flight tests were
each pair of points that constitute the footprint boundaries (the conducted between 19 Jul 2021 and 8 Sep 2021, in a rich range
footprint rectangular corners) [2]. The 2D projected footprint of illumination and weather conditions. Weather conditions
extent l of a vision-based sensor, illustrated in Fig. 5, can be included tests under clear and partly cloudy skies, with calm
calculated using:
  
h
ltop, bottom = s0pu (z) · tan α ± tan−1 (10)
2f
  
w
lleft, right = s0pu (z) · tan β ± tan−1 (11)
2f
where s0pu is the UAV altitude, α and β are the camera’s
pointing angles from the vertical z and horizontal x axis of
the World coordinate frame, w is the lens width, h is the lens
height, and f is the focal length.
The footprint corners c from the camera’s local coordinate
frame I are translated to the world’s coordinate frame W using: Ground

 0   0    
c (x) spu (x) cos(ϕu ) − sin(ϕu ) c(x)
= + (12)
c0 (y) s0pu (y) sin(ϕu ) cos(ϕu ) c(y) Fig. 5. Field of View (FOV) projection and footprint extent of a vision-based
sensor. The camera setup on the UAV frame defines α as the pointing angle
where s0puis the next UAV position state, and ϕu is the from the vertical (or pitch) and determines the coordinates of the footprint
Euler yaw angle of the UAV. However, as no actions involve corners c.
and gusty winds from 6 km/h up to 24 km/h respectively. The rectangular area, drafted in QGroundControl. Specific details
range of recorded temperatures ranged from 14°C to 25°C. of the survey pattern can be found in Fig. 15 and Tab. IV from
This implementation employed a static adult mannequin the Appendix. A diagram illustrating the functionality of tested
posing as the victim to be found for safety reasons. The flight modes is shown in Fig. 8.
mannequin was placed at two predefined locations, as depicted 1) Mission Mode: When mission mode is activated, the
in Fig. 7. The first location—referred from here as Location 1, UAV automatically follows a list of position and velocity
or L1—is a trivial setup with the mannequin free of any nearby waypoints which define the survey plan previously drafted in
obstacles and entire visibility from downward-looking cameras. QGroundControl and uploaded to the autopilot before starting
The second location—referred from here as Location 2, or the flight operation. This flight mode is traditionally supported
L2—introduces a complex setup as the mannequin is placed in many autopilots, and its out-of-the-box implementation
nearby a tree which causes partial occlusion during the flight serves as the planner baseline of this research. While mission
tests.

B. Flight Modes
The proposed UAV system is evaluated by collecting data
using three flight modes: mission, offboard, and hybrid. The
survey extent for these tests is delimited by a 6 m × 60 m

(a)

(a)

(b) (b)
Fig. 6. Location of conducted flight tests at the Samford Ecological Research Fig. 7. Adult mannequin placed in the surveyed area as the victim to be
Facility (SERF), QLD, Australia. (a) SERF and surveyed area boundary found. (a) Trivial victim location (L1) with the mannequin fully exposed in
extents (orange and red blobs respectively). (b) Aerial footage of surveyed the environment. (b) Complex victim location (L2) with the mannequin partly
area displaying buffel grassland and obstacles. occluded by a five-metre tree.
Fig. 8. Executed flight modes for exploration and object detection in outdoor environments. Mission mode is the baseline motion planner and lets the UAV
survey the SAR emulated area by following a lawnmower pattern. Offboard mode runs the POMDP motion planner by populating an initial victim position
belief across the entire flying area. Hybrid mode extends the functionality of mission mode by running the POMDP motion planner to inspect the area delimited
by the camera’s FOV.

mode is operated in the UAV, the object detector is running Once the motion server calls the motion planner after a
in parallel to record any positive detections while the UAV is first victim detection is received by the object detector, TAPIR
navigating in the environment and completing the survey. is booted by calculating an offline policy for four seconds.
2) Offboard mode: Offboard mode offers autonomous Afterwards, the observation server retrieves an observation,
navigation without a predefined survey plan of the environ- updates the motion policy and takes the action that returns
ment. This flight mode internally executes the POMDP-based the highest expected reward. An idle period of 3.4 seconds is
motion planner described in Section III by declaring as flight applied for the UAV to reach the desired position coordinate,
parameters the initial position waypoint where the UAV should and then, the process repeats itself by requesting a new
begin the survey, and the global coordinates of the survey observation from the observation server. The loop is broken
extents. The list of parameters can be found in Tab. V from once the detection confidence ζ exceeds a threshold. Specific
the Appendix. parameters from the TAPIR toolkit and ABT solver are shown
3) Hybrid Mode: This paper proposes the fusion of the in Table V from the Appendix.
provided capabilities between mission and offboard modes,
V. Results and Discussion
in a flight mode denominated hybrid. The aim of this flight
mode is to take advantage of the initial awareness and survey The proposed UAV framework is evaluated through the
coverage coming from mission mode in outdoor environments performance indicators listed as follows: (1) Successful de-
with GNSS signal coverage, and the autonomous navigation tections per flight mode; (2) Spatial distribution of recorded
capabilities of offboard mode. Instead of running the POMDP- GNSS coordinates via heatmaps; (3) Elapsed time taken by the
based motion planner covering the entire extent of the surveyed UAV to locate the victim per location; and (4) Scalability test
area, in hybrid mode the survey extent is only limited by the using thermal imagery. Real flight demonstrations of the UAV
extent of the camera’s FOV. Once a first detection is received framework can be found at https://youtu.be/U_9LbNXUwV0.
from the vision module, this flight mode triggers offboard Accuracy metrics of victim detections were recorded using
mode, boots the motion planner and passes action commands three variables: True Positives (TP), False Positives (FP), and
to the autopilot until the POMDP solver reaches a terminal False Negatives (FN). TP is defined here as the relative number
state (i.e.the UAV discards or confirms a victim). Afterwards, of flight runs where the victim was successfully detected at
the UAV resumes its survey by triggering back mission mode. the true location. FP is the relative number of flights which
The process repeats itself with new detection outputs until the recorded victim locations in other areas than the true position
UAV completes the survey in mission mode. of the victim. FN is the relative number of flights that did not
detect the victim at their real location. In this context, a given
C. POMDP solver flight test could report false positive detections and still detect
The navigation problem modelled as a POMDP is solved in the victim at the real location. A summary table of collected
real time through the use of the TAPIR toolkit [17]. TAPIR is metrics is depicted in Tab. II.
coded using the C++ programming language and encapsulates Accuracy metrics of the proposed framework provided
the Augmented Belief Trees (ABT) solver [19] to calculate and contrasting results at the tested victim locations. On flight
update the motion policy online. ABT reduces computational tests with the victim placed in a trivial location (i.e., L1), the
demands by reusing past computed policies and updating the UAV achieved 100% of positive victim detections in mission
optimal policy if changes to the POMDP model are detected. mode, and 20% of those recorded GNSS coordinates of false
Furthermore, formulated problems with ABT allow declaring victim locations. For tests in offboard and hybrid flight modes,
continuous values for actions, states, and observations. the true positive rates decreased in comparison with mission
 TABLE II
Accuracy metrics of the system to locate a victim at two locations
(L1 and L2). Here, TP are true positives, FP are false positives, and FN
are false negatives.

Flight Mode Runs TP (%) FP (%) FN (%)

Mission (L1) 5 100.0 20.0 0.0
Offboard (L1) 7 57.1 0.0 42.9
Hybrid (L1) 5 80.0 0.0 20.0
Mission (L2) 5 60.0 0.0 40.0
Offboard (L2) 5 80.0 0.0 20.0
Hybrid (L2) 7 71.4 28.6 28.6

mode. However, both setups achieved flight tests without any

false positive readings. An illustration of the spatial distribution
of recorded GNSS coordinates per flight mode for L1 is shown
in Fig. 10.
Flight tests with the mannequin located in a complex location
(i.e., L2) showed an overall improvement in TP rates for
offboard and hybrid modes compared to the baseline planner
(mission mode). For mission and offboard modes, there were
no flights which reported FP detections even though the flight
setup, payload, and object detector remained the same while
testing the framework with the mannequin located in L1.
Nevertheless, 28.6% of the flights in hybrid mode reported
false positive victim locations. The rate of false positives for all
flight tests were caused by limitations from the object detector
by recording other objects as humans (i.e., from the car placed
in the flight area). Conversely, most of the false negative records
during the flight tests occurred from excessive vibration in the
UAV frame caused by strong winds, as shown in Fig. 9. Fig. 9. Strong winds distorting RGB streaming in offboard mode. The top
A visual analysis of the spatial distribution of recorded image illustrates the traversed path of the UAV and the probability distribution
GNSS coordinates during the flight tests is performed using of the UAV and victim locations (orange and red points respectively). The
bottom image shows the latest (distorted) streamed frame from the RGB
heatmaps. The heatmaps, illustrated in Figs. 10 and 11, indicate camera.
a reduction in victim location uncertainty after operating the
proposed UAV framework using offboard and hybrid modes.
The presented POMDP-based motion planner from [32]— This work also studied the speed of the proposed framework
scaled in this work for UAV navigation in outdoor to find victims using mission, offboard, and hybrid flight modes.
environments—contributed on the reduction of object detection As presented in Tab. III, two primary insights are observed
uncertainty in flight tests using offboard and hybrid modes. An in elapsed times between victim locations and flight modes.
example of how the UAV inspects an area to confirm whether The first observed trend defines higher values of standard
a victim is truly located after receiving an initial detection deviation (especially in offboard mode) in flight tests with
is shown in Fig. 12. In the first time steps (Fig. 12a), a low the victim at location L1. These values were caused by slight
confidence value of 19.51% is retrieved because of the few inconsistencies in the traversed path from the motion planner,
number of pixels representing the mannequin while surveying as seen in Fig. 13. From the surveyed area, Location 1 is closer
at 16 m, and partial occlusion from the nearby tree. After to the survey limits, whereas Location 2 is placed closed the
taking actions commands from the POMDP policy computed upper centre. The motion planner lets the UAV move towards
in the planner module, the UAV is positioned closer to the the centre of the surveyed area before exploring its corners.
mannequin (i.e., 10 m) and with a better viewpoint of the scene, Despite test runs in offboard mode providing higher overall
retrieving a confidence value of 90.0% (Fig. 12b). The traversed accuracy values than tests in hybrid mode, surveyed patterns
path by the UAV also suggests the capability of the UAV to could become suboptimal if a victim is located close to the
adapt (or update) its motion policy while it interacts with the boundaries of the delimited flying area.
environment and receives new observations. Adjustments in The second observed trend is that hybrid mode recorded
the motion policy also occur from uncertainty sources, such longer times for detecting and confirming the victim, regardless
as unexpected strong wind currents, oscillating GNSS signal of its location in the surveyed area. This impact highly depends
errors, illumination changes, and false detections by the CNN on the recall properties of the vision-based object detector. The
model. number of false positive outputs while the UAV explores the
 (a) (a)

(b) (b)

(c) (c)
Fig. 10. Heatmaps of recorded GNSS coordinates in a trivial victim location Fig. 11. Heatmaps of recorded GNSS coordinates in a complex victim location
(L1) using (a) mission, (b) offboard, and (c) hybrid flight modes. (L2) using (a) mission, (b) offboard, and (c) hybrid flight modes.
 TABLE III
Elapsed time by the UAV to locate a victim at two locations (L1 and L2)
per flight mode. Here, SD stands for Standard Deviation and SE stands
for Standard Error.

Flight Mode Mean (s) SD (s) SE (s)

Mission 169.67 – –
Offboard (L1) 146.24 128.70 64.35
Hybrid (L1) 392.39 130.20 65.10
Offboard (L2) 148.30 25.58 12.79
Hybrid (L2) 289.26 76.84 34.36

(a)

Fig. 13. Example traversed path in offboard mode while no victims are found.
The UAV moves towards the centre of the surveyed area before exploring its
corners.

environment defines the number of inspections, which will

increase the flight time until the survey is complete. Using
other object detectors tuned from airborne UAV datasets is
expected to reduce the survey duration in hybrid mode, as these
models should provide higher recall values than the MobileNet
SSD detector implemented in this paper.
Limitations in the implemented object detector to output
positive detections have conditioned the maximum altitude for
the surveys, and consequently, impacted the overall speed of the
system to complete the mission. Other restrictions are defined
by the gimbal configuration, which scopes the footprint of the
camera’s FOV, and the image resolution of streamed frames.
Therefore, reduced times to accomplish the victim finding
(b)
mission can be accomplished by improvements in the object
Fig. 12. Reduction of object detection uncertainty from RGB camera after detector, higher image resolution and oblique camera angles
executing the motion policy. (a) Initial detection of a potential victim with that provide a wider viewpoint of the scene. Nonetheless, such
a confidence value of 19.51%. (b) Increased confidence value (90.0%) at
subsequent detections after the UAV executes actions from the computed
improvements might involve an increment in computational
motion policy and gets closer to the victim. power and further research should evaluate any impacts from
better detectors and processing of high-resolution frames in
embedded systems such as the UP2 .
 The scalability test of the presented framework was achieved
with preliminary flight tests using thermal imagery. The
modular components replaced for this trial consisted of the
UAV payload, the vision-based detector, and the flight plan
generated by QGroundControl to match the desired overlap
values previously tested with the RGB camera.
The object detector is a custom implementation of a
TinyYoloV2 model architecture from Microsoft Azure Custom
Vision services. A total of 5175 labelled images were used in
the training process. The thermal dataset contains images of
adults with a rich set of posing configurations: adults lying on
the ground and waving their arms; adults standing up waving
their arms; altitude and camera viewpoint variations from the
UAV, ranging from 5 to 40 m, and gimbal angles of 45° and
0° from Nadir.
Thermal preview tests were conducted between 5:30 a.m.
and 7:30 a.m. from 25 Aug 2021 to 23 Sep 2021, with ambient
temperatures between 7°C and 14°C. An overview of the system
navigating in hybrid mode to increase values of object detection
confidence from streamed thermal frames can be found in
Fig. 14. In this instance, the UAV starts inspecting the area
covered by the FOV of the thermal camera (Fig. 14a). In
subsequent time steps, the aircraft gets closer to have a better
visual of the scene and confirm the presence of the victim (a)
by retrieving higher detection confidence values (Fig. 14b).
A demonstrative preview video of flight operations with thermal
imagery can be found at https://youtu.be/yIPNBwNYtAo.
The proposed UAV framework constitutes a novel approach to
autonomous navigation for exploration and target finding under
uncertainty. Real-world environments are full of uncertainties
such as illumination conditions, strong wind currents, collision
from static and dynamic obstacles, occlusion, and limitations
in object detectors, which can negatively affect the success
of the flight mission. While many approaches handle object
detection uncertainty by fine-tuning CNN models with labelled
airborne datasets [23], this research suggests augmenting the
cognition power onboard UAVs from imperfect sensor and
detection output observations. Moreover, augmenting autonomy
capabilities in small UAVs might open more approaches to
automated surveying in outdoor environments by using a swarm
of UAVs which will not require permanent supervision by pilots
to photo-interpret streamed camera frames to identify and locate
potential victims.
Several benefits can be offered through the use of the
framework in time-critical applications to SAR squads. A first
assessment of the accessibility conditions of the surveyed
environment, identification, localisation, quantification and
conditions of victims and external hazards might be obtained (b)
rapidly thanks to real-time telemetry of processed camera
Fig. 14. Reduction of object detection uncertainty from thermal camera after
frames. After completing a flight, the list of GNSS coordinates executing the motion policy. (a) Initial detection of a potential victim with
(depicted in Figs. 10 and 11) can be shared to SAR squads a confidence value of 31.75%. (b) Increased confidence value (83.33%) at
to coordinate better response strategies, as well as launching subsequent detections after the UAV gets closer to the victim.
additional UAVs for critical deployment of medicines, food or
water.
 VI. Conclusions TABLE IV
Flight plan parameters based on RGB camera properties.
This paper discussed a modular UAV framework for au-
tonomous onboard navigation in outdoor environments under Property Value
uncertainty. The system showed how levels of object detection UAV altitude 16 m
uncertainty were substantially reduced by calculating a motion UAV velocity 2 m/s
policy using an online POMDP solver and interacting with the Lens width 2.06 mm
Lens height 1.52 mm
environment to obtain better visual representations of potential Camera focal length 4.7 mm
detected targets. CNN-based computer vision inference and Image resolution 640 by 480 px
motion planning can be executed in resource-constrained Overlap 30%
hardware onboard small UAVs. The framework design was Bottom right waypoint -27.3892746°, 152.8730164°
validated with real flight tests with a simulated SAR mission, Top right waypoint -27.3887138°, 152.8730164°
Top left waypoint -27.3887138°, 152.8727722°
which consisted in finding an adult mannequin in an open area Bottom left waypoint -27.3892765°, 152.8727722°
and close to a tree. Collected performance indicators from
three flight modes suggest that the system reduces levels of
object detection uncertainty in outdoor environments whether
information about the surveyed environment is available. This
framework was also extended by adapting the payload with a
thermal camera and converting a customised people detector
from thermal imagery in the vision module.
The presented framework extends the contributions in [31],
[30] by: (1) extending their tested UAV framework in GNSS-
denied environments for outdoor missions with GNSS signal
coverage with a novel flight mode (i.e., hybrid mode), (2)
further validating preliminary results of the presented frame-
work using real flight tests, and (3) demonstrating scalability
opportunities of the modular framework design by adapting a
thermal camera and custom object detector to locate victims
using their heat signatures.
Future work should evaluate the performance of the UAV sys-
tem using tailored object detectors from network architectures
different than MobileNet SSD and high-resolution streamed
frames. Further assessments with multiple victims and more
complex environment configurations (e.g. slope terrain, type
and number of obstacles) are encouraged. A comparison study
of various online POMDP solvers, or other motion planners,
could improve understanding the limits of using the ABT solver
and POMDP solvers for motion planning under uncertainty.
Appendix
The survey design and flight plan parameters for conducted
tests in mission and hybrid flight modes are shown in Fig. 15
and Tab. IV, respectively.
The set of hyper-parameters used in TAPIR and initial
conditions to operate the UAV using offboard and hybrid modes Fig. 15. Flight plan following a lawnmower pattern using the RGB camera
are shown in Tab. V. properties from Tab. IV.

Acknowledgements
The authors wish to express their gratitude to Sharlene Lee-Jendili This research was funded by the Commonwealth Scientific and
for her implementation of the thermal-based people detector used in Industrial Research Organisation (CSIRO) through the CSIRO Data61
this project. The authors acknowledge continued support from the PhD and Top Up Scholarships (Agreement 50061686); the Australian
Queensland University of Technology (QUT) through the Centre for Research Council (ARC) through the ARC Discovery Project 2018
Robotics. Special thanks to the Samford Ecological Research Facility “Navigating under the forest canopy and in the urban jungle” (grant
(SERF) team (Marcus Yates and Lorrelle Allen) for their continuous number ARC DP180102250); and the Queensland University of
assistance and equipment provided during the flight tests. The authors Technology (QUT) through the Higher Degree Research (HDR)
would also like to gratefully thank the QUT Research Engineering Tuition Fee Sponsorship. Special thanks to Hexagon through the
Facility (REF) team (Dr Dmitry Bratanov, Gavin Broadbent, Dean Hexagon SmartNet RTK corrections service that enabled high accuracy
Gilligan) for their technical support that made possible conducting surveying and positioning data using the EMLID Reach RTK receiver
the flight tests. during the experimentation phase.
 TABLE V [16] J. Jiménez López and M. Mulero-Pázmány. Drones for conservation in
Set of hyper-parameters used in TAPIR and initial conditions to protected areas: Present and future. Drones, 3(1):10, 2019.
operate the UAV in offboard and hybrid flight modes. [17] D. Klimenko, J. Song, and H. Kurniawati. TAPIR: a software toolkit for
approximating and adapting POMDP solutions online. In Australasian
Variable Description Value Conference on Robotics and Automation, pages 1–9, 2014.
zmax Maximum UAV altitude 16 m [18] A. Koubaa, A. Allouch, M. Alajlan, Y. Javed, A. Belghith, and M. Khalgui.
zmin Minimum UAV altitude 5.25 m Micro air vehicle link (mavlink) in a nutshell: A survey. IEEE Access,
pu0 Initial UAV position (-27.3897972°, 152.8732300°, 20m) 7:87658–87680, 2019.
ϕu UAV Heading 0° [19] H. Kurniawati and V. Yadav. An online POMDP solver for uncertainty
δz UAV climb step 2m planning in dynamic environment. In M. Inaba and P. Corke, editors,
λ Frame overlap 40% Robotics Research. Springer Tracts in Advanced Robotics, volume 114,
pages 611–629. 2016.
α Camera pitch angle 0°
ζmin Minimum detection confidence 10% [20] S. Lee, D. Har, and D. Kum. Drone-assisted disaster management:
Finding victims via infrared camera and lidar sensor fusion. In 3rd
ζ Confidence threshold 85% Asia-Pacific World Congress on Computer Science and Engineering,
γ Discount factor 0.95 pages 84–89, Nadi, Fiji, 2016.
∆t Time step interval 4s [21] S. Macdonald and A. Stevens. How to explore planets with drones.
tmax Maximum flying time 10 min Astronomy & Geophysics, 59(3):3.18–3.22, 2018.
[22] S. Mayer, L. Lischke, and P. W. Woźniak. Drones for search and rescue.
In 1st International Workshop on Human-Drone Interaction, pages 1–7,
2019.
References [23] P. Mittal, R. Singh, and A. Sharma. Deep learning-based object detection
in low-altitude UAV datasets: A survey. Image and Vision Computing,
[1] Australian National Search and Rescue Council. Volume 2 - search and 104:104046, 2020.
rescue operations. In National Search and Rescue Manual, chapter 3, [24] N. H. Motlagh, M. Bagaa, and T. Taleb. UAV-based iot platform: A crowd
pages 73–353. Canberra, Australia, 2020. surveillance use case. IEEE Communications Magazine, 55(2):128–134,
[2] P. Bourke. Polygons and meshes. 2017.
http://paulbourke.net/geometry/polygonmesh, 1997. [25] G. Pajares. Overview and current status of remote sensing applications
[3] R. Z. B. Bravo, A. Leiras, and F. L. Cyrino Oliveira. The use of UAVs in based on unmanned aerial vehicles (UAVs). Photogrammetric Engineering
humanitarian relief: An application of POMDP-based methodology for & Remote Sensing, 81(4):281–330, 2015.
finding victims. Production and Operations Management, 28(2):421–440, [26] M. G. Pensieri, M. Garau, and P. M. Barone. Drones as an integral part
2019. of remote sensing technologies to help missing people. Drones, 4(2):15,
[4] S. Bricknell. Missing persons: Who is at risk? Technical report, Australian 2020.
Institute of Criminology, Canberra, Australia, 2017. [27] C. Ponzoni Carvalho Chanel, A. Albore, J. T’Hooft, C. Lesire, and
[5] M. Chen, E. Frazzoli, D. Hsu, and W. S. Lee. POMDP-lite for robust F. Teichteil-Königsbuch. Ample: an anytime planning and execution
robot planning under uncertainty. In International Conference on Robotics framework for dynamic and uncertain problems in robotics. Autonomous
and Automation, pages 5427–5433, 2016. Robots, 43(1):37–62, 2019.
[6] A. Chovancová, T. Fico, L. Chovanec, and P. Hubinsk. Mathematical [28] S. Ragi and E. K. P. Chong. UAV path planning in a dynamic environment
modelling and parameter identification of quadrotor (a survey). Procedia via partially observable markov decision process. IEEE Transactions on
Engineering, 96:172–181, 2014. Aerospace and Electronic Systems, 49(4):2397–2412, 2013.
[7] Y. Chuanqi. Caffe implementation of Google MobileNet SSD de- [29] S. Ragi and E. K. P. Chong. UAV guidance algorithms via partially
tection network, with pretrained weights on voc0712 and map=0.727. observable markov decision processes. In K. Valavanis and G. Vacht-
https://github.com/chuanqi305/MobileNet-SSD, 2020. Accessed 2020- sevanos, editors, Handbook of Unmanned Aerial Vehicles, chapter 73,
08-30. pages 1775–1810. Dordrecht, 2015.
[8] M. Erdelj and E. Natalizio. UAV-assisted disaster management: Appli- [30] J. Sandino, F. Maire, P. Caccetta, C. Sanderson, and F. Gonzalez. Drone-
cations and open issues. In International Conference on Computing, based autonomous motion planning system for outdoor environments
Networking and Communications, pages 1–5, 2016. under object detection uncertainty. Remote Sensing, 13(21):4481, 2021.
[9] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, [31] J. Sandino, F. Vanegas, F. Gonzalez, and F. Maire. Autonomous UAV
J. Winn, and A. Zisserman. The PASCAL visual object classes challenge: navigation for active perception of targets in uncertain and cluttered
A retrospective. International Journal of Computer Vision, 111(1):98– environments. In IEEE Aerospace Conference, pages 1–12, 2020.
136, 2015. [32] J. Sandino, F. Vanegas, F. Maire, P. Caccetta, C. Sanderson, and
[10] A. Gupta, D. Bessonov, and P. Li. A decision-theoretic approach to F. Gonzalez. UAV framework for autonomous onboard navigation and
detection-based target search with a UAV. In International Conference people/object detection in cluttered indoor environments. Remote Sensing,
on Intelligent Robots and Systems, pages 5304–5309, 2017. 12(20):3386, 2020.
[11] L. Hanson and K. Namuduri. Real-world applications. In K. Namuduri, [33] K. P. Valavanis and G. J. Vachtsevanos. Future of unmanned aviation. In
S. Chaumette, J. H. Kim, and J. P. G. Sterbenz, editors, UAV Networks K. P. Valavanis and G. J. Vachtsevanos, editors, Handbook of Unmanned
and Communications, pages 194–213. Cambridge, UK, 2017. Aerial Vehicles, chapter 126, pages 2993–3009. Dordrecht, 2015.
[12] M. Hassanalian, D. Rice, and A. Abdelkefi. Evolution of space drones [34] F. Vanegas, D. Campbell, M. Eich, and F. Gonzalez. UAV based target
for planetary exploration: A review. Progress in Aerospace Sciences, finding and tracking in GPS-denied and cluttered environments. In
97:61–105, 2018. International Conference on Intelligent Robots and Systems, pages 2307–
[13] A. Hornung, K. M. Wurm, M. Bennewitz, C. Stachniss, and W. Burgard. 2313, 2016.
Octomap: an efficient probabilistic 3D mapping framework based on [35] F. Vanegas and F. Gonzalez. Uncertainty based online planning for
octrees. Autonomous Robots, 34(3):189–206, 2013. UAV target finding in cluttered and GPS-denied environments. In IEEE
[14] U. Ilhan, L. Gardashova, and K. Kilic. UAV using dec-POMDP model Aerospace Conference, pages 1–9, 2016.
for increasing the level of security in the company. Procedia Computer [36] S. Waharte and N. Trigoni. Supporting search and rescue operations with
Science, 102:458–464, 2016. UAVs. In International Conference on Emerging Security Technologies,
[15] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, pages 142–147, 2010.
S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast
feature embedding. In International Conference on Multimedia, pages
675–678, 2014.