Constructing Full-Coverage 3D UAV Ad-Hoc

Networks Through Collaborative Exploration in

Unknown Urban Environments
Narangerelt Batsoyol, YeonJin Jin, and HyungJune Lee
Department of Computer Science and Engineering
Ewha Womans University, Seoul, South Korea
Email: hyungjune.lee@ewha.ac.kr

Abstract—We consider a 3D network construction problem Step 1: Collaborative Exploration Step 2: Network Construction

in the post-disaster scenario, where large urban areas are UAV Sensing Range
communication-wise isolated from the outside environment due
to the severely damaged network infrastructure. Our main goal
is to reconnect the isolated regions with the outside environment
using unmanned aerial vehicles (UAVs) by building 3D aerial ad-
hoc networks. Prior to network construction, we aim to capture
the global map information over region of interests (RoI) by
Task
exploring all obstacles in the unknown region. We propose an Division
Network
Connection
efficient technique for collaborative 3D terrestrial exploration Base Station
using multiple UAVs based on our distributed path planning
Fig. 1. System overview with collaborative exploration and network con-
algorithm, which finds collision-free exploration paths. Then, struction using UAVs
we present an optimal full-coverage 3D aerial ad-hoc network
construction by deploying the minimum number of UAVs to using UAVs: randomly sampling search algorithms such as
indispensable spots while obtaining maximum network coverage. rapidly-exploring random tree and probabilistic roadmap, and
Simulation results demonstrate that our proposed exploration
scheme outperforms several counterpart algorithms in terms optimal search algorithms [8], [9]. These algorithms, however,
of traversal time and redundant visit rate. Also, our network have not explicitly explored the area with obstacles.
construction algorithm guarantees almost full coverage toward The coverage and connectivity problem in 3D ad-hoc net-
terrestrial space with only minimal UAV usage. works have attracted a great attention among researchers in
recent years. Especially, constructing a reliable network in
I. I NTRODUCTION
atmospheric or underwater environments is a challenging task.
In a post catastrophic disaster situation, network infrastruc- Underwater acoustic sensor networks have been studied in [1],
ture tends to be severely collapsed. Geographical regions in while climate monitoring and weather forecasting problems
a large urban area occupied with various low-rise, mid-rise, have been solved by deploying 3D aerial networks [2], [3],
and high-rise buildings can be communication-wise isolated, [6]. Most of previous works related with 3D coverage problem
or totally secluded from the outside environment. It would be have used some space filling approach by dividing a space into
a promising approach to reconnect these isolated regions with equal polyhedrons based on transmission range [2], [10]. How-
the outside environment (toward base stations) by constructing ever, the environment is assumed to be free from obstacles,
ad-hoc networks using unmanned aerial vehicles (UAVs). and thus, they have limitations in directly applying to more
Since any global map information over the cluttered region practical settings.
is unknown, it is essential to explore the given region of In this paper, we aim to solve two key problems of 1)
interests (RoI) to capture all the obstacles’ distribution status how to explore an unknown complex 3D space occupied with
including their height information without missing any unvis- obstacles in a fast manner using multiple UAVs, and 2) how
ited areas in a fast manner. UAVs can also be used as flexible to construct a full-coverage 3D ad-hoc network in an isolated
communication relays to form an aerial ad-hoc network due cluttered environment.
to their less movement constraint and high maneuverability. We propose an efficient method for collaborative exploration
Regarding efficient area exploration, many researchers in the over unknown urban environment using multiple UAVs based
robotics field have conducted space exploration with obstacle on a novel path planning algorithm. The proposed algorithm
avoidance mostly in two dimensions, whereas the space explo- finds optimal and collision-free exploration paths in 3D space.
ration in three dimensions (3D) has not been actively studied. The main goal is to completely explore the whole area over
Several algorithms were implemented in 3D environments RoI and build a complete height map as soon as possible,
while avoiding redundant visit at each specific location.
This work was supported by Basic Science Research Program through the
National Research Foundation of Korea (NRF) funded by the Ministry of Then, we present a heuristic yet efficient full-coverage 3D
Education (NRF-2015R1D1A1A01057902). aerial network construction algorithm. The challenge is to

978-1-5386-3180-5/18/$31.00 ©2018 IEEE

find the best 3D deployment location for each UAV such that
UAVs can extend wireless coverage toward all the terrestrial
buildings and spaces, and can also communicate with other
UAVs and base stations. We aim to minimize the number of
deployed UAVs, while guaranteeing maximum coverage.
Our paper is organized as follows: After presenting our sys-
tem model in Sec. II, we describe our collaborative exploration
with path planning in Sec. III and network construction in (a) Naive Algorithm (b) Our Algorithm
Sec. IV. We validate our algorithms in Sec. V and finally
conclude our work in Sec. VI. Fig. 2. Basic movement trajectory in a space with no obstacle

II. S YSTEM M ODEL

We consider the problem of constructing full-coverage ad-
hoc networks using UAVs in unknown urban environments.
Our goal is to perform an efficient 3D exploration for capturing
the height dynamics over the RoI and dispatch UAVs for
extending wireless coverage toward terrestrial areas and also
constructing aerial ad-hoc networks. (a) Naive Algorithm (b) Our Algorithm
For this work, we employ a theoretical unit disk model for
the communication model with communication range 𝑅. A Fig. 3. Basic movement trajectory in a space with obstacles
UAV can communicate with other UAVs and base stations
inevitable to have redundant cell visits, leading to inefficient
located at the corner of RoI via a wireless radio such as 802.11
coverage, and delayed exploration. However, since two or
or 802.15.4. We let UAVs traverse over virtual 3D grid cells
more UAV agents can communicate each other within the
to facilitate an otherwise complex problem.
radio range and exchange information about their visited cells,
The 𝑥𝑦 axis represents a cross-sectional 2D representation
the redundant visit can be reduced. For this, some negotiation
from the above, while the 𝑧 axis is the height level in 3D
protocol for sharing their visited cells and the height of them is
space. Each individual grid cell’s information is provided by
necessary. Further, collaborative exploration based on almost
0 or 1 depending on whether an obstacle is located at the cell.
equal task division of visiting the remaining cells among
It is assumed that UAVs can sense any obstacles in its directly
the UAVs would increase exploration efficiency. All these
adjacent cells within the sensing range using radar sensors or
factors should work together to help other UAVs to minimize
laser scanners [4].
redundant visits in the future, and the overall exploration
The problem of constructing full-coverage aerial ad-hoc
procedure can finish as early as possible.
networks can then be described with two sub-problems: 1) 3D
exploration for height mapping by UAVs and 2) aerial ad-hoc A. Path Planning
network construction, as illustrated in Fig. 1.
We aim to design an efficient path planning algorithm for
III. C OLLABORATIVE E XPLORATION a single UAV to construct a complete 2D height map by
To construct aerial ad-hoc networks with UAVs in an exploring the RoI on the 3D space in a cluttered city with
unknown territory occupied with various low-rise, mid-rise, buildings. We suppose that each UAV can detect an obstacle
and high-rise buildings, it is essential to explore the given RoI in its directly adjacent cells, which means it can sense its
without missing any unvisited areas in a rapid manner. This surrounding 3×3×3 cells in 3D space. Our algorithm performs
can be achieved by finding the height of cluttered buildings in exploration with two modes: non-obstacle exploration and
a geographically regular basis and constructing a height map obstacle exploration depending on whether an obstacle is
over the RoI. Further, if we are allowed to use multiple UAVs detected.
for the exploration process, they would rather collaborate with 1) Non-obstacle Exploration: For exploration, each UAV
each other to find optimal and collision-free exploration paths sets its basic movement direction randomly from clockwise or
in 3D space for even faster execution. The main goal is to counter-clockwise direction. The UAV starts exploring from
complete the height mapping process over the given RoI as the border region of the RoI, and travels toward the center
soon as possible, while avoiding redundant visits at a specific in a spiral manner, following the basic direction under the
location. non-obstacle circumstance. Since the UAV can sense directly
Each individual UAV agent follows its own distributed path adjacent cells, it may well follow not along the border line, but
planning decision. Whenever the UAV agent visits a certain along non-border cells that are one cell away from the border
cell as close as possible on the surface of 3D obstacles from line. A naive scanning exploration can visit each cell at a
the air, it updates its visited cell list and records the height of time in a regular basis for complete coverage as in Fig. 2(a),
the obstacle located below at the same 𝑥𝑦 coordinate. whereas our proposed exploration can reduce the exploration
If each UAV agent follows its own distributed path planning completion time by quickly covering the RoI area with a spiral
without considering other agents’ exploration progress, it is horizontal movement as in Fig. 2(b).
 For the vertical exploration, the UAV starts exploring from
the second level since it can sense cells in the lower and
the upper levels. In case of detecting an obstacle during this
procedure, the UAV changes to the obstacle exploration mode
and gradually flies upward up to the top of the currently
detected obstacle.
A UAV can move in North, South, West, and East direction
while mapping, and saves its sensed cells in the sensedList. To
determine its movement direction among these four directions,
it counts the number of unsensed cells in its North, South, East,
and West direction, and goes to the direction with the largest
number of unsensed cells. If there are multiple candidate Fig. 4. Collaborative exploration using 3 UAVs for faster height mapping with
20 × 20 grid size and 4 sub-areas of 10 × 10 grid size where communication
directions, the UAV selects a more highly prioritized direction range 𝑅 is 10
based on its initially selected basic movement direction at the
beginning of the non-obstacle exploration. For each direction, are repeated until all the cells from innerList are explored.
we hold some priority order for directions among North, South, Illustrative examples are shown in Fig. 3.
East, and West. We prioritize the clockwise direction with the Once a group of obstacles for a building are fully explored,
following order of North → East → South → West, while the the UAV descends back to the second level and continues to
counter-clockwise direction has the following priority order explore all other cells over RoI, while transitioning back and
of South → East → North → West. The UAV continues to forth to the non-obstacle exploration mode.
move based on the selected direction until it meets a border A more detailed path planning algorithm is described in
cell or an already-sensed cell. Upon encountering one of these Algorithm 1.
cells, it finds a new direction based on the aforementioned
procedure and follows the trajectory until it encounters an B. Task Division
obstacle or completes the mapping process (for obtaining the We extend our path planning algorithm for a single UAV
height information for all the cells). to the one with multiple UAVs. By letting multiple UAVs
2) Obstacle Exploration: Once a UAV encounters an obsta- explore RoI in a collaborative way, the overall area exploration
cle located at adjacent cells within its sensing range during the for obtaining all the height information of cells over RoI can
non-obstacle exploration, it changes the mode to the obstacle be reduced. We propose a hierarchical task division method
exploration mode. It first tries to fully encircle the base of that divides a territory area into smaller sub-areas and assigns
a building of which the obstacle is a part. After finishing sub-areas to visit almost equally to communicable UAVs.
exploring the base of the building (before moving upward), the Considering the communication range 𝑅, we divide the RoI
UAV obtains the knowledge of how large the obstacle group into multiple sub-areas with the size of 𝑅 × 𝑅 so that a UAV
is and how its belonging cells are located. Except for directly located at the center of a sub-area can communicate with
visited cells during the encircling exploration, the UAV can another at that of adjacent sub-areas.
list up a set of inner cells, innerList that need to be explored At the start, each UAV thinks that it is responsible for
afterwards. exploring every sub-area in RoI and follows its own ex-
Upon the completion of encircling at the lowest level, it ploration path planning according to Sec. III-A. Each UAV
ascends up to the fourth level since the information of cells maintains assignedList that saves the sub-area IDs that it needs
located at the third level has already been obtained during the to explore, and visitedSubareaList that updates already-visited
lowest encircling. Then, the next cell to visit is determined by sub-area IDs. Each UAV is subject to completely explore one
choosing a cell with the most adjacent unexplored cells from sub-area first, and then is allowed to move to another. To
the innerList at the current level. If there are several candidate determine which sub-area to visit next time, the UAV finds
cells, we choose the closest one among them. the nearest sub-area from assignedList. It continues to visit
Then, the UAV generates a path from the currently visiting the remaining sub-areas from the list one by one until it meets
cell to the selected destination cell and starts moving horizon- another UAV, or every sub-area from the list is explored.
tally and then vertically, or vertically and then horizontally. When several UAVs get encountered each other in the air
We choose one of two movement behaviors that has the within their communication range, they exchange the infor-
larger number of unvisited cells from innerList. In case of mation of visitedSubareaList and which sub-area is currently
encountering another even higher obstacle during the journey being explored by each own UAV. Based on the shared
on the way, it performs the hill climbing and descending until information, the UAVs perform the task division procedure that
reaching the current destination cell. The UAV climbs the distributes the remaining sub-areas to visit into them with the
hill until there is no obstacle cell found within its sensing similar workload by updating assignedList. It should be noted
range and follows its path. In case that there is no obstacle that if there exists any obstacle between two UAVs within the
cell right below, the UAV descends until finding an obstacle communication range, we consider them non-communicable
cell right below or reaching the second level. These steps each other.
 We represent the allocation of sub-areas to available UAVs
Algorithm 1 Distributed Path Planning Algorithm as making bids at each iteration step. At each iteration, each
1: Input: Starting position of UAV, 3D cartesian coordinate UAV 𝑘 is allocated to one sub-area and registers the sub-area’s
2: Output: 2D height map 𝐻
ID at its 𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑𝐿𝑖𝑠𝑡𝑈 𝐴𝑉𝑘 . We calculate a weight measure
//Set Initial State
//𝐵𝑎𝑠𝑖𝑐𝐷𝑖𝑟 shows priority of direction for each sub-area 𝑖 to prioritize sub-areas to allocate to a better
//𝐶𝑙𝑜𝑐𝑘𝑤𝑖𝑠𝑒 = [North-East-South-West] UAV as follows:
//𝐶𝑜𝑢𝑛𝑡𝑒𝑟-𝐶𝑙𝑜𝑐𝑘𝑤𝑖𝑠𝑒 = [South-East-North-West]
//If the cell at 𝑥𝑦𝑧 is an obstacle, then 𝐶𝑥𝑦𝑧 = 1, else 𝐶𝑥𝑦𝑧 = 0 𝑤𝑒𝑖𝑔ℎ𝑡𝑆𝑖 =𝐷(𝐿𝑜𝑐𝑈 𝐴𝑉𝑘 , 𝐶𝑆𝑖 )+
3: 𝑧𝑖𝑛𝑖𝑡𝑖𝑎𝑙 of UAV = 2; ∑
4: 𝐵𝑎𝑠𝑖𝑐𝐷𝑖𝑟 = random(𝐶𝑙𝑜𝑐𝑘𝑤𝑖𝑠𝑒, 𝐶𝑜𝑢𝑛𝑡𝑒𝑟-𝐶𝑙𝑜𝑐𝑘𝑤𝑖𝑠𝑒); (𝐷(𝐶𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑𝐿𝑖𝑠𝑡𝑈 𝐴𝑉𝑘 (𝑗) , 𝐶𝑆𝑖 ) (1)
5: while (𝐻 is not completed) do 𝑗
6: // I. Select direction 𝐷𝑖𝑟 to move
7: 𝐷𝑖𝑟 = Direction (N/S/E/W) of 𝑚𝑎𝑥(# of unexplored cells); where 𝑆𝑖 denotes sub-area 𝑖, 𝐿𝑜𝑐𝑈 𝐴𝑉𝑘 is the current
8: if # of 𝐷𝑖𝑟 ≥ 2 then location of UAV 𝑘, 𝐶𝑆𝑖 is the center of sub-area 𝑖,
9: Choose 𝐷𝑖𝑟 which comes first in 𝐵𝑎𝑠𝑖𝑐𝐷𝑖𝑟 list;
10: end if 𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑𝐿𝑖𝑠𝑡𝑈 𝐴𝑉𝑘 (𝑗) is the 𝑗th sub-area from assignedList of
11: // II. Follow trajectory UAV 𝑘, and 𝐷(𝐴, 𝐵) denotes geographical distance between
12: while (𝐻 is not completed && no obstacle cell detected && 𝐴 and 𝐵. This weight measure means the cumulative distance
not reached border cell && not reached already-sensed cell)) do
13: Follow the path in 𝐷𝑖𝑟;
measure for all possible combinations from the current UAV
14: Update 𝑠𝑒𝑛𝑠𝑒𝑑𝐿𝑖𝑠𝑡 location, or already-assigned, but not-visited-yet sub-areas to
15: end while a candidate sub-area.
16: // III-i. Obstacle detection In case that several UAVs select the same sub-area with the
17: if (At least one 𝐶𝑥𝑦𝑧 = 1 in sensor range) then
18: // i) Obstacle base encircling nearest distance, a UAV with the smallest number of visited
19: 𝐺= current position (starting location of obstacle exploring); sub-areas is chosen. If there still exist ambiguities, a UAV
20: while 𝑛𝑒𝑥𝑡 𝑐𝑒𝑙𝑙 != 𝐺 && 𝑛𝑒𝑥𝑡 𝑐𝑒𝑙𝑙 != border cell do is randomly chosen and allocated to the sub-area. From the
21: Find Cardinal Direction (of UAV’s sensing area) which has
𝐶𝑥𝑦𝑧 = 1; second iteration, the weight measure is used to quantify the
22: if (𝐶𝑥𝑦𝑧 = 1 is at East or NorthEast side) then distance measure from a group of its already-allocated sub-
23: Go to North; areas to a sub-area candidate and to allocate the remaining
24: else if (𝐶𝑥𝑦𝑧 = 1 is at South or Southeast side) then
25: Go to East; sub-areas to visit to UAVs.
26: else if (𝐶𝑥𝑦𝑧 = 1 is at West or Southwest side) then It should be noted that if multiple UAVs become aware of
27: Go to South; exploring the same sub-area within the communication range,
28: else if (𝐶𝑥𝑦𝑧 = 1 is at North or Northwest side) then
29: Go to West; we let a UAV with the larger number of already-visited cells
30: end if in the sub-area be allocated to the current sub-area, whereas
31: Update 𝑠𝑒𝑛𝑠𝑒𝑑𝐿𝑖𝑠𝑡 other UAVs stop exploring at the current sub-area and leave for
32: Save the detected obstacle cell in 𝑜𝑏𝑠𝑡𝑎𝑐𝑙𝑒𝐿𝑖𝑠𝑡;
33: end while another sub-area from their own assignedList by task division.
34: // ii) Obstacle height exploration An example of collaborative exploration is illustrated in Fig. 4.
35: Increase 𝑧 to 4 & update 𝑠𝑒𝑛𝑠𝑒𝑑𝐿𝑖𝑠𝑡; After finishing exploring all the sub-areas from subareaList,
36: 𝑖𝑛𝑛𝑒𝑟𝐿𝑖𝑠𝑡 = cells encircled by 𝑜𝑏𝑠𝑡𝑎𝑐𝑙𝑒𝐿𝑖𝑠𝑡;
37: 𝑆 = 𝑥𝑦 location of cell in 𝑜𝑏𝑠𝑡𝑎𝑐𝑙𝑒𝐿𝑖𝑠𝑡 (which has max(# of
each UAV goes back to the nearest corner cell where a base
unexplored adjacent obstacle cells)); station is located and finishes its height mapping process via
38: while (𝑆 exist) do exploration. All the UAVs share their local map information
39: 𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙 = [𝑥 of 𝑆, 𝑦 of 𝑆, current 𝑧 level];
40: Count # of cells in 𝑜𝑏𝑠𝑡𝑎𝑐𝑙𝑒𝐿𝑖𝑠𝑡 on the way if we go to
through their connected base station and merge into a global
𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙’s 𝑥 first with current 𝑦 then go to 𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙’s 𝑦 ; or height map over RoI.
go to 𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙’s 𝑦 first with current 𝑥 then go to 𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙’s 𝑥.
A 𝑝𝑎𝑡ℎ with max # of cells is chosen, else selected in random. IV. N ETWORK C ONSTRUCTION
41: while (Go to 𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙 following the 𝑝𝑎𝑡ℎ) do Once the height mapping process for the entire cells over
42: if There exist higher level obstacles on the way to 𝑑𝑒𝑠𝑡𝐶𝑒𝑙𝑙
then
RoI is completed, we focus on constructing a full coverage 3D
43: Do hill climbing and descending until there is an obstacle ad-hoc network using UAVs. The challenge is to find the best
right below; location for each UAV to be deployed in 3D space such that
44: 𝐻𝑥𝑦 = 𝑧 − 1, with 𝑥𝑦 of current position;
45: end if
UAVs can extend wireless coverage to all terrestrial buildings,
46: Update 𝑠𝑒𝑛𝑠𝑒𝑑𝐿𝑖𝑠𝑡 and can communicate each other where at least one UAV is
47: end while connected to a base station. We aim to minimize the number
48: Update 𝑆 (from unvisited cells ∈ {𝑜𝑏𝑠𝑡𝑎𝑐𝑙𝑒𝐿𝑖𝑠𝑡 ∪
𝑖𝑛𝑛𝑒𝑟𝐿𝑖𝑠𝑡});
of deployed UAVs, while guaranteeing maximum wireless
49: end while coverage. We suppose that a UAV can cover all nearby cells
50: // iii) Descend to the lowest level within the communication range 𝑅 where there is no obstacle
51: If there is unvisited border/near border cell nearby, UAV goes to between them.
that cell. Else should go to the closest unvisited cell at current 𝑧.
52: Decrease UAV’s 𝑧 value until 𝑧 = 2 & Update 𝑠𝑒𝑛𝑠𝑒𝑑𝐿𝑖𝑠𝑡; Motivated by some previous works on wireless coverage in
53: end if 3D space [2], [10], we rely on the same hierarchical sub-area
54: end while structure used for task division in Sec. III-B. The network
construction problem using UAVs are divided into three steps:
 2) Mid-rise and High-rise Obstacle Coverage: After fin-
ishing the ground space coverage, we start covering obstacle
buildings higher than 2𝑅 + 1 level. We define 𝑚 = 𝑅 + 1,
which is the lowest level of a UAV’s deployment for the
ground space coverage, and 𝑛 = 2𝑅 + 1, which is the
maximum coverage level from the level 𝑚 by the UAV. In
general, if we deploy a UAV at 𝑚 + 𝑖 × 𝑛 level (where 𝑖 is
a natural number), its resulting wireless coverage ranges from
(a) Subarea-based deployment (b) Obstacle coverage 𝑖 × 𝑛 to (𝑖 + 1) × 𝑛 level.
Depending on a building’s height level, we deploy the UAV
either to the top of the building or at the side. If the building’s
height is between 𝑖 × 𝑛 and 𝑚 + 𝑖 × 𝑛 level, the UAV can be
placed on the top of the building. Otherwise, if the building
height is between 𝑚 + 𝑖 × 𝑛 and (𝑖 + 1) × 𝑛 level, the UAV
cannot be physically placed inside the building, and thus, is
relocated at the side. In this case, we need (𝑖 − 1) additional
UAVs to deploy at the levels of 𝑚, 𝑚+𝑛, . . ., 𝑚+(𝑖−1)×𝑛.
To determine its 𝑥𝑦 coordinate for the UAV deployment, we
(c) Network hole filling (d) Refinement
check which position leads to covering the largest number of
Fig. 5. Evolutionary network construction optimization after necessary steps obstacle cells. In case that multiple UAVs are assigned to the
where the circles represent deployed UAVs, and solid lines show valid same building at the side, they are deployed at the different
communication links where base stations are located at four corners of RoI
levels with the same 𝑥𝑦 coordinate to get connected each other.
1) sub-area coverage for fully covering the ground space with If a UAV cannot fully cover a building with wide cross-section
or without obstacles, 2) network hole coverage by deploying located at the same level, other UAVs are deployed to locations
UAVs into the detected network holes, and 3) refinement with different 𝑥𝑦 coordinates at the same level.
process for discarding unnecessary UAVs. At this point, we achieve 100% sub-area coverage towards
A. Sub-area Coverage ground space and even building areas by dispatching additional
Our network construction procedure is initiated with the UAVs to the sectored sub-area regions.
first step of sub-area coverage. By deploying only one UAV
to each sub-area, we check whether one UAV is enough to B. Network Hole Coverage
cover its responsible sub-area. In case that there exist non- Although all the deployed UAVs are responsible for extend-
reachable cells shadowed by nearby obstacles from the current ing wireless coverage within their own subarea, the problem
UAV’s location at the sub-area, we deploy more UAVs for of constructing communication links among those UAVs and
guaranteeing 100% wireless coverage with a more aggressive connecting to at least one base station have not been addressed
manner. Further, to extend wireless coverage even to mid-rise yet. In this section, we present network construction among
and high-rise obstacles themselves, the deployment positions UAVs by finding out network holes and deploying additional
for additional UAVs should be determined. UAVs as relays.
1) Ground Space Coverage: The ground space coverage At the base station side, we list up all possible connections
consists of iterative steps to fully cover each sub-area by among UAVs and base stations. Based on the information,
finding out the best UAV positions where they can cover the we run the Dijkstra’s algorithm [7] to construct shortest path-
maximum number of cells in each sub-area. For the ground based communication networks.
space coverage, a UAV is deployed at 𝑅 +1 level, and extends Based on the extracted network information, we first find
their wireless coverage down to the bottom level and up to out isolated UAVs or base stations within 2𝑅 range to connect
2𝑅 + 1 level. via additional UAV deployment. In case that an isolated UAV
At the first iteration, we assign each UAV to every sub-area, is located together with other UAVs and base stations within
and each UAV counts all possible coverage cells considering 2𝑅 range, respectively, it chooses the closest base station to
obstacle shadowing for each deployment cell position inside make a direct connection to the base station. If there is no
its sub-area. In this stage, we need the same number of UAVs base station to connect, but exist multiple non-isolated UAVs
as the number of sub-areas as shown in Fig. 5(a). within the range, we choose one UAV among them that has
We further perform a few more iterations to achieve 100% the smallest number of hops toward its connected base station.
full coverage by deploying more UAVs to unfulfilled sub- Otherwise, if there exist only isolated UAVs to connect, we
areas. We find out the best positions to cover the obstructed select the closest UAV.
cells in the ground space and deploy additional UAVs at After choosing candidates to connect, we explore where to
the locations where the number of obstructed cells can be deploy additional UAVs as relays. We create a virtual 3D cube
minimized. We continue to run this iterative procedure until where an isolated UAV, and its selected candidate are located
there is no obstructed cell on the ground space. at diagonal vertexes. Then, we search over all possible cells
 1000
within the cube to find out one cell location that does not Hierarchical
100 Hierarchical
Centralized Centralized

Redundancy Percentage (%)

have an obstacle cell and not interrupted by any obstacle cells 800 Distributed
80
Distributed

Travel Time (sec)

for direct communication link. We place an additional UAV 600
60
at the chosen cell as relay. If this additionally deployed UAV
400
becomes connected not only with the original isolated UAV, 40

but with other isolated UAVs, we do not place more UAVs for 200 20

the other UAVs since they have already get connected with 0 0
1 2 3 4 1 2 3 4
luck. We continue this procedure until all the isolated UAVs # of UAVs # of UAVs

get connected toward at least one base station. (a) Travel time performance (b) Redundant cell-visit rate

C. Refinement Fig. 6. Travel time and redundant visit rate performance according to task
division by varying the number of UAVs on 30 × 30 grid size test sets
During the previous steps presented in Secs. IV-A and
IV-B, we have constructed complete communication networks
that connect all the UAVs to at least one base station while We analyze exploration performance based on total 100
extending wireless coverage to all the 3D spaces in the RoI randomly generated test sets where 20 test sets for each
as illustrated in Fig. 5. different grid size from 10 × 10 to 50 × 50 are used. As in
Fig. 7(a), our 3D path planning algorithm outperforms a naive
We perform the last optimization step to minimize the
scanning algorithm with a factor of up to 2. Both algorithms
number of UAVs deployed for network construction, while
lead to full exploration over all the cells, and the constructed
preserving the almost full wireless coverage property. We want
height topology through each exploration exactly matches the
to differentiate some redundant UAVs from core ones that play
ground-truth in our test datasets.
key roles in connecting many other connection points.
We investigate how task division improves exploration
We first check all closely located UAVs within 𝑅/2 range
efficiency in terms of travel time and redundant visit rate.
and pairs of UAVs that have at least 60% wireless coverage
We compare three different task division mechanisms: 1) our
overlap. If a UAV has the smaller number of communication
hierarchical task division, 2) a centralized task division, and
links, and does not create any network hole upon discarding
3) a distributed task division. The centralized task division
it, we remove the UAV from our network construction. We
algorithm, considered as a theoretically optimal bound, divides
continue to do this refinement process, and find the optimal
the exploration space into pre-determined sub-areas of which
number of UAVs and complete the whole network construc-
each one is assigned to a responsible UAV. The distributed task
tion.
division allocates unvisited cells using 𝐾-means clustering to
V. E VALUATION the encountered UAVs within communication range after they
exchange the visited cell information, without using any sub-
We validate our proposed algorithms in a simulated envi-
area division.
ronment. Our 3D environment is located over 30 × 30 grid
We use 20 randomly generated test sets with 30 × 30 grid
cell topology where communication range 𝑅 is up to 10 cells,
size and 9 sub-areas. As in Fig. 6, the centralized task division
and the sub-area size is 10 × 10 grid size. We use the UAV’s
performs best in terms of travel time and redundant visit
relative flying speed of 1 𝑐𝑒𝑙𝑙/𝑠𝑒𝑐. In our experiments, we
rate because the regions to explore with each UAV are pre-
have prepared 100 different basic building patterns and have
determined in an optimal way using a centralized manner. On
randomly generated test datasets with different grid size and
the other hand, our hierarchical task division also performs
also randomly chosen height of the building.
well comparable to the centralized task division, whereas the
Our evaluation is divided into two parts: 1) collaborative distributed task division performs worst. This demonstrates
exploration performance and 2) network construction perfor- that exploiting not a cell, but a sub-area as the minimal
mance. For collaborative exploration, we quantify total travel allocation unit and exploration based on it have indeed avoided
time for full 3D exploration as the farthest travel time among possible redundant path overlap among UAVs.
UAVs, and the average redundant visit rate among UAVs by
varying the number of UAVs. For network construction, we B. Network Construction
measure total coverage rate based on how many cells are We evaluate network construction performance in terms of
covered by UAVs. network coverage rate and the number of deployed UAVs. We
categorize network coverage in two aspects: 1) sensing cov-
A. Collaborative Exploration erage on how deployed UAVs extend their wireless coverage
To validate our path planning algorithm, we compare it toward terrestrial areas, and 2) communication coverage on
against a naive scanning algorithm that follows the zigzag whether deployed UAVs can communicate each other toward
trajectory [5]. The naive scanning algorithm is considered as at least one base station.
the basic motion planning method used mostly in agriculture, As in Fig. 7(b), our network construction algorithm keeps
search and rescue, and mapping task using UAVs as in optimizing over iterations in terms of both sensing and com-
Fig. 2(a). We adapt this algorithm to fit into the 3D space munication coverage, reaching 100% coverage at the network
with obstacles as shown in Fig. 3(a). hole filling step. Through the last refinement process, the
 5000

Network Coverage (%)

100
4000
Travel time (sec)

80
3000
60
2000
40

1000 20
Ours Sensing Coverage
Naive Communication Coverage
0 0
10 20 30 40 50 Iter1 Iter2 Iter3 ObsFill HoleFill Refine
Grid Size (N x N) Coverage Steps

(a) Path planning performance (b) Network coverage

50
# of UAVs per each step 100 Coverage per size of Lattice

Network Coverage (%)

# of UAVs per size of Lattice 800
40
90 (a) Constructed network topology using our algorithm
# of UAVs

600

# of UAVs
30

80 400
20

10 70 200

0 60 0
Iter1 Iter2 Iter3 ObsFill HoleFill Refine 10x10x10 6x6x6 5x5x5 3x3x3
Coverage Steps Size of Lattice (1 UAV per ρ x ρ x ρ)

(c) # of deployed UAVs (d) Lattice-based space filling

Fig. 7. Path planning efficiency and network coverage performance

(b) Lattice topology with 𝜌 = 10 (c) Lattice topology with 𝜌 = 6

coverage performance slightly decreases to 98% due to the Fig. 8. Network topology comparison between lattice-based space filling and
removal of 4 less effective UAVs as in Fig. 7(c). our algorithm
Lastly, we compare our network construction against a
space filling approach in 3D space with obstacles after we full coverage toward terrestrial space with only minimal UAV
adapt from [2]. We let one UAV deployed to one lattice cube usage.
consisting of 𝜌 × 𝜌 × 𝜌 cells. As UAVs are more densely For future work, to prepare some possible network failure
deployed by decreasing the lattice size 𝜌 from 10 to 3 in and extend the current problem, we may design a more conser-
Fig. 7(d), both sensing and communication coverage have vative wireless coverage by covering a specific location with at
improved. To achieve full coverage around 100%, however, least 𝑘 UAVs. Also, it would be interesting to reflect realistic
a tremendously large number of UAVs (i.e., 952) are required. problems such as battery outage and sensing failure during the
Even for the coverage rate of 91%, 116 UAVs are required for mission to design a more adaptive network algorithm.
the lattice-based space filling, whereas ours requires only 30
UAVs for the coverage of 98%. Fig. 8 shows the visualization R EFERENCES
of each obtained network topology. [1] I. F. Akyildiz, D. Pompili, and T. Melodia. Underwater acoustic sensor
networks: research challenges. Ad hoc networks, 3(3):257–279, 2005.
VI. C ONCLUSION [2] S. Alam and Z. J. Haas. Coverage and connectivity in three-dimensional
networks. In ACM MobiCom, pages 346–357. ACM, 2006.
We have presented an optimal full-coverage 3D aerial ad- [3] H. M. Ammari and S. Das. A study of k-coverage and measures of
hoc network construction method using multiple UAVs in connectivity in 3d wireless sensor networks. IEEE Transactions on
Computers, 59(2):243–257, 2010.
an urban environment. We have proposed ways to deploy [4] N. Gageik, P. Benz, and S. Montenegro. Obstacle detection and collision
the minimum number of UAVs to connect UAVs and base avoidance for a uav with complementary low-cost sensors. IEEE Access,
stations all together, while guaranteeing maximum network 3:599–609, 2015.
[5] D. Jeong, S.-Y. Park, and H. Lee. DroneNet: Network reconstruction
coverage. Prior to the network construction procedure, we have through sparse connectivity probing using distributed UAVs. In IEEE
proposed an efficient technique for collaborative 3D terrestrial PIMRC, pages 1797–1802, 2015.
exploration using multiple UAVs based on distributed path [6] D. Pompili, T. Melodia, and I. F. Akyildiz. Three-dimensional and
two-dimensional deployment analysis for underwater acoustic sensor
planning. Our collaborative exploration algorithm finds effi- networks. Ad Hoc Networks, 7(4):778–790, 2009.
cient collision-free exploration paths in 3D space in a rapid [7] S. Skiena. Dijkstras algorithm. Implementing Discrete Mathematics:
manner by reducing redundant visit as much as possible via Combinatorics and Graph Theory with Mathematica, Reading, MA:
Addison-Wesley, pages 225–227, 1990.
aerial task division. [8] K. Yang and S. Sukkarieh. 3D smooth path planning for a UAV in
Our experiment results show that our path planning algo- cluttered natural environments. In IEEE/RSJ Intelligent Robots and
rithm for height mapping significantly outperforms a baseline Systems (IROS), pages 794–800. IEEE, 2008.
[9] L. Yang, J. Qi, J. Xiao, and X. Yong. A literature review of uav 3d path
naive scanning algorithm. Also, our collaborative exploration planning. In Intelligent Control and Automation (WCICA), 2014 11th
procedure using multiple UAVs finds an effective way to World Congress on, pages 2376–2381. IEEE, 2014.
perform task division, while reducing the overall traversal time [10] C. Zhang, X. Bai, J. Teng, D. Xuan, and W. Jia. Constructing low-
connectivity and full-coverage three dimensional sensor networks. IEEE
and the redundancy percentage. We have also demonstrated Journal on Selected Areas in Communications, 28(7), 2010.
that our network construction technique guarantees almost