Surface-driven Next-Best-View planning for

exploration of large-scale 3D environments

Guillaume Hardouin1,2 , Fabio Morbidi2 , Julien Moras1 ,
Julien Marzat1 , El Mustapha Mouaddib2
1
ONERA-DTIS, Université Paris-Saclay, Palaiseau, France
Email: [email protected]
2
MIS laboratory, Université de Picardie Jules Verne, Amiens, France
Email: {fabio.morbidi, mouaddib}@u-picardie.fr

Abstract: In this paper, we propose a novel cluster-based informative path planning algorithm
to simultaneously explore and inspect a large-scale unknown environment with an Unmanned
Aerial Vehicle (UAV). Most of the existing methods address the surface inspection problem
as a volume exploration problem, and consider that the surface has been scanned when the
corresponding volume has been covered. Unfortunately, this approach may lead to inaccurate
3D models of the environment, and the UAV may not achieve global coverage. To overcome
these critical limitations, we introduce a 3D reconstruction method based on TSDF (Truncated
Signed Distance Function) mapping, which leverages the surfaces present in the environment to
generate an informative exploration path for the UAV. A Probabilistic Roadmap planner, used
to solve a TSP (Travelling Salesman Problem) over clusters of viewpoint configurations, ensures
that the resulting 3D model is accurate and complete. Two challenging structures (a power plant
and the Statue of Liberty) have been chosen to conduct realistic numerical experiments with a
quadrotor UAV. Our results provide evidence that the proposed method is effective and robust.

Keywords: Next-Best-View planning, 3D reconstruction, Truncated Signed Distance Function,

Probabilistic Roadmap, Unmanned Aerial Vehicle (UAV).

1. INTRODUCTION Distance Function (TSDF) is considered. The surface of

the obstacles is estimated by incrementally fusing the
range measurements into a volumetric distance field that
Autonomous robots are being increasingly employed to- represents the signed distance to the closest surface. This
day for time-consuming and dangerous tasks usually per- field can then be transformed into a surface representation.
formed by human operators. For instance, aerial robots In Newcombe et al. (2011); Klingensmith et al. (2015),
with different on-board sensors (RGB-D and stereo cam- the authors have shown that an optimized implementation
eras, laser range finders, etc.) hold great potential for of TSDF mapping, running on an embedded computer,
modeling large-scale 3D structures. Recent applications can provide an accurate mesh reconstruction in real-time.
include digital cultural heritage, exploration of confined Furthermore, TSDF data can be useful to identify miss-
and cluttered environments, and structural inspection for ing parts of the model during the online reconstruction
preventive maintenance (Tabib et al. (2016); Song and Jo phase, as in Monica and Aleotti (2018). Thanks to the
(2018); Bircher et al. (2018)). In Next-Best-View (NBV) volumetric representation, it is possible to generate a list
planning (Connolly (1985)), the robot iteratively computes of candidate sensor configurations from the missing parts
the best viewpoint configurations to fully reconstruct the in the identified free space. We propose an NBV planning
3D environment, which can be (partially) known in ad- method which allows to visit all these configurations and
vance or completely unknown. The robot then progres- which solves a Travelling Salesman Problem (TSP) based
sively discovers the surrounding environment during its on the known map: an optimal path visiting the viewpoint
mission. The focus can be either on the exploration of configurations is computed by taking the surface coverage
the 3D volume (exploration problem) or on the consistency they provide and their location in space, into account. As
and completeness of the reconstructed surface (inspection the robot follows the path, it performs successive scans and
problem), for which different path-planning strategies have completes the map. The initial path is updated by lever-
been developed in the literature. aging the gathered information, until the whole unknown
In this paper, we study the problem of incremental ex- surface is covered.
ploration and surface reconstruction of an unknown 3D In summary, the original contributions of this work are
environment for inspection purposes, with an Unmanned threefold:
Aerial Vehicle (UAV). To address this problem, a map of
the environment is required for collision avoidance and • We only use a TSDF representation of surfaces to
to verify if the reconstruction is complete. In this work, measure the quality of reconstruction and to identify
a map representation based on the Truncated Signed the unknown or incomplete areas from which new
 viewpoint configurations for the UAV, can be gen- of the TSDF and on the RRT* expansion Karaman and
erated. Frazzoli (2011). They also compared the reconstruction
• We propose a novel objective function for NBV plan- quality/completeness of their method with that of some
ning, which incorporates the TDSF-based informa- recent volumetric approaches.
tion gain at each robot configuration. Volumetric exploration aims at exploring and at building
• We design a new algorithm, based on the computation a map of a large unknown environment, by using a 3D
of successive (approximate) solutions to the TSP in grid model (such as Octomap Hornung et al. (2013)) to
conjunction with a Probabilistic Roadmap planner identify known, unknown or occupied areas. The majority
(Lazy PRM*, Hauser (2015)), to visit the clusters of of recent exploration methods are sampling-based Bircher
viewpoint configurations for surface reconstruction. et al. (2016); Papachristos et al. (2019). They expand
a random tree (RRT/RRT*) of sampled sensor configu-
The remainder of this paper is organized as follows. In Sec- rations in the free space, and keep the NBV trajectory
tion 2, we review the state of the art in autonomous inspec- that guarantees the best volume coverage. Pose coverage
tion and exploration of unknown environments. Section 3 is evaluated via ray tracing, and the volume is explored as
is devoted to the problem formulation. The proposed algo- the UAV moves along the assigned path. These methods,
rithm is described in Section 4. A detailed evaluation of the which rely on a coarse volumetric map of the environment
algorithm in realistic simulation environments is reported for navigation purposes, are fast and efficient, but they do
in Section 5. Finally, in Section 6, the main contributions not explicitly account for the accuracy of the reconstructed
of the paper are summarized and some possible avenues surface. Some recent works have tried to combine volumet-
for future research are outlined. ric exploration (Octomap) and surface inspection for large-
scale 3D reconstruction. In Bircher et al. (2018), a prelim-
inary exploration of the unknown volume is followed by
2. RELATED WORK an inspection phase. The path that ensures the maximum
coverage of the surface is selected, and the inspection stops
The ability to plan informative paths for online exploration when the resolution of the facets of the reconstructed mesh
and modeling of 3D environments, is an essential prerequi- is optimal. An algorithm based on the RRT* expansion
site for truly autonomous robots. Numerous strategies for is proposed in Song and Jo (2017) for the generation of
3D modeling of an unknown environment with a sensor- an exploration path, where the nodes are sampled poses.
equipped robotic platform exist in the literature. In this A minimal number of viewpoint configurations is chosen,
paper, we are specially concerned with NBV methods, which guarantees the coverage of surface cells and the local
which determine the best viewpoints to visit depending shortest path for the aerial robot. However, the quality of
on the environment and the nature of the robot mission. the obtained 3D model is not evaluated. In Song and Jo
Surface inspection methods analyze the reconstructed sur- (2018), the same authors improved their previous method
face for viewpoint definition. The goal is to ensure that the by considering the incomplete areas of the 3D model,
reconstruction is accurate and complete. Frontier-based in their sampling-based pose generation algorithm. The
methods generate viewpoint configurations on the frontier volume coverage is evaluated via an Octomap, and the
of the observed surface, which satisfy some orientation, surface model is built from the TSDF volume which is
positioning, and sensing constraints. These configurations, used, in turn, to generate a point cloud. Such a point cloud
visited by the sensor-equipped robot, provide new infor- is exploited on the fly to determine the incomplete areas
mation about the surface, with some overlapping to guar- that the robot should visit next.
antee continuity (Connolly (1985); Vasquez-Gomez et al. Our literature review shows that in the existing methods
(2014); Border et al. (2018)). However, the information for simultaneous exploration and inspection of an unknown
gain associated to each viewpoint cannot be predicted to environment, volume exploration has received by far the
evaluate the NBV. Some works have addressed this issue most attention, while surface inspection has been only
by estimating the unknown surface for viewpoint selection addressed indirectly. In this paper, we push the envelope
(Pito (1996); Kriegel et al. (2015)), but generally, only and explicitly solve the surface inspection problem by con-
small objects are reconstructed. In addition, these works sidering a frontier-based approach coupled with an NBV
rely on strong assumptions on the location of the objects planning strategy for guidance in large-scale environments.
in space, and the only use of a surface representation Viewpoint configurations are clustered according to their
makes it difficult to extend these approaches to robotic location in space, to evaluate how the visit of specific areas
applications, since the free space cannot be determined. is informative and find a suitable path. This search is
The authors in Vasquez-Gomez et al. (2014) were the formalized as a TSP which is successively solved with a
first to consider a volumetric representation for frontier- heuristic algorithm until the reconstruction is completed.
based object reconstruction. This approach has been used Moreover, instead of expanding a random tree, we use the
in Yoder and Scherer (2016) for a large-scale environment, Lazy PRM* planner (Hauser (2015)), to find the shortest
but the quality of the 3D reconstruction obtained with path between two configurations by taking the map of
a UAV has not been assessed. Recently, in Monica and the environment into account. This planner is faster than
Aleotti (2018), new poses are generated by directly lever- RRT, and it is multi-query, making it ideal for path check-
aging the properties of the TSDF. Finally, in Schmid et al. ing between clusters, and thus for the resolution of the
(2020), the authors exploited the TSDF representation for TSP. Realistic numerical experiments featuring two large-
surface reconstruction and the Euclidean SDF (Oleynikova scale structures (an industrial plant and a statue) and a
et al. (2017)) for navigation, and they proposed a path- quadrotor UAV, show the effectiveness of our approach.
generation strategy based on the volumetric properties
 3. PROBLEM FORMULATION

In this paper, we consider a UAV with 4 degrees of freedom

(the position [x, y, z]T ∈ R3 and the yaw angle ψ ∈ S 1 ),
equipped with a depth sensor, e.g. a stereo camera or a
RGB-D sensor. The aerial robot should scan an unknown
but spatially-bounded environment V ⊂ R3 , and it should
accurately reconstruct the surface that it contains. Its 3D
exploration path should also be as short as possible. Figure 1. General flowchart of our algorithm: the internal archi-
tecture of the planner is shown inside the shaded box.
We denote by Q the set of all collision-free configurations
q = [x, y, z, ψ]T of the UAV, which are assumed to Definition 1 (Incomplete surface element). A voxel
be exactly known. The vector-valued function pji (s) : v ∈ M which satisfies (1), represents an Incomplete
[0, 1] → R3 × S 1 defines the path from configuration i Surface Element (ISE). We denote by C ⊂ M the set of
to configuration j, where pji (0) = qi and pji (1) = qj . all ISEs.
We assume that pji (s) is collision-free and feasible for the Definition 2 (Remaining incomplete surface). Let
aerial robot (i.e. the kinematic/dynamic constraints of the Qc ⊆ Q be the set of all configurations q from which voxel
aerial vehicle are satisfied along the path). A localisation v ∈ C can be completed. The S remaining
 incomplete surface
system provides the (exact) pose of the UAV with respect is then defined as Crem = v ∈ C v | Qc = ∅ .
to a world reference frame. Definition 3 (Scanned element). A voxel v ∈ M which
The forward-facing depth sensor is rigidly attached to the satisfies w(r) ≥ Wth , ∀ r ∈ Nv6 , is called a scanned
aerial robot, and its pose calibrated with respect to the element.
body frame of the UAV. The sensor provides a depth Problem 1 (UAV exploration & inspection). Find a
map and it has a limited field of view (FOV) and sensing collision-free path pn0 between the initial configuration q0
range. The depth maps are incrementally integrated in and the final configuration qn , which allows the UAV to
a TSDF volumetric map M , which consists of a voxel scan the set Cins = C\Crem of all ISEs.
grid where each voxel contains a truncated signed distance
value φ ∈ R and a weight w ≥ 0. Depending on the values 4. PROPOSED APPROACH
of φ and w, one can determine if a voxel is unknown,
occupied or empty. In fact, voxel v ∈ M is, During the exploration, the map is incrementally built
and the UAV detects sets of incomplete surfaces or con-
 unknown if w(v) = 0,

tours (cf. Section 3). The configurations which allow to
occupied if w(v) > 0 ∧ φ(v) ≤ 0, complete them are generated and clustered, depending
if w(v) > 0 ∧ φ(v) > 0,

empty on their location in the 3D space. A directed graph is
created and continuously expanded to represent the travel
where w(v) and φ(v) are the weight and signed distance
utility between clusters in the free space. A path ensuring
of voxel v, respectively. Note that φ(v) = 0 implicitly
collision-free navigation is extracted from this graph to
defines a surface: hence, the TSDF volume is a volumetric
maximize the utility. As the map evolves, new ISEs are
representation of a surface. Every time a voxel is scanned,
discovered, and the path to complete them is updated.
its associated weight is incremented by 1/zc2 (v) to mini-
The 3D model is considered complete when there are no
mize the sensing error, where zc (v) is the distance between
more ISEs.
voxel v and the current sensor position. To filter out noise
As depicted in Fig. 1, the planner includes two modules.
and reduce the number of false-positive voxels, we consider
The first module is the ISE extractor which identifies
voxel v to be known if its weight is greater than or equal
incomplete areas to visit from the TSDF map, generates
to a threshold Wth , i.e if w(v) ≥ Wth > 0. The value of
candidate viewpoint configurations, and clusters them.
Wth depends on the sensing range of the depth sensor,
The second module is the graph-based planner, which
and it can be determined experimentally. In Monica and
finds the optimal path for 3D reconstruction. It relies on a
Aleotti (2018), the authors defined a contour as the set of
single objective function, which trades off travel cost and
empty voxels that are neighbors to both an occupied and
utility of visiting a new cluster. The path is updated if
an unknown voxel. More precisely, voxel v ∈ M is said to
obstacles are detected along the way or if the area which
belong to a contour if the following three conditions are
still needs to be scanned has been already completed.
fulfilled:
The two modules are described in more detail in the next
a) w(v) ≥ Wth ∧ φ(v) > 0, two sections.
b) ∃ r ∈ Nv6 s.t. w(r) < Wth , (1)
18 4.1 ISE extractor
c) ∃ o ∈ Nv s.t. w(o) ≥ Wth ∧ φ(o) ≤ 0,
where Nv6 and Nv18 denote the 6- and 18-connected voxel Viewpoint generation from an incomplete element. Fol-
neighborhoods of v, respectively. In fact, if one observes lowing (Monica and Aleotti, 2018, Sect. 3.1), we adopt a
the space around the occupied voxels next to a frontier, fast local approach to determine a direction nv for observ-
the perception of the corresponding surface of the object ing the ISE v. This approach is based on the gradient of the
improves. weight function ∇w(x, y, z), which can be computed as,
X c−v
The following three definitions are introduced to support nv = w0 (c) , (2)
our problem statement. 26
kc − vk
c ∈ Nv
where Nv26 is the 26-connected neighborhood of v (in order
to minimize noise and sampling effects). We propose the
following variant of the “modified” weight function w0 ,
−Wth if voxel c is occupied,

w0 (c) =
Wth otherwise.
Note that nv in (2) points away from the known surface
(see Fig. 2), instead of being orthogonal to it, as in a
classical frontier-based approach. Such a direction guaran-
tees a better surface coverage when the incomplete areas
are located on sharp edges. A new sensor configuration is
generated along the direction nv at a specified distance
δpose from voxel v. The sensor points towards the incom- Figure 3. Example of weighted directed graph G with three clusters.
plete element: in this way, new informative cues about the
surface can be extracted, while still guaranteeing that the in Bresenham (1965), a set of rays is traced inside the
previously-explored regions are visible. Parameter δpose voxel map located at the viewpoint sensor frustum. When
can be selected according to the technical specifications a ray crosses a voxel, the attributes of this voxel are
of the depth sensor (e.g., range, FOV, etc., see Chen et al. stored, e.g. known or unknown, free or occupied. This
(2008)). Unfortunately, the poses qi , qj ∈ Q generated approach is often used in volume-exploration tasks, to
from the ISEs vi , vj ∈ C, respectively, may occasionally obtain the information gain of each viewpoint. To evaluate
be very close, i.e. dist(qi , qj ) <  for a small  > 0, or the a viewpoint configuration, we use the ray-tracing method
viewing directions are almost parallel, i.e. |nvi · nvj | ' 1. from a frontier-based perspective, i.e. we simply count the
These configurations are then aggregated into a single number of ISEs that can be seen from that location. More
viewpoint by averaging their positions and orientations. formally, let Cq S
be the set of all ISEs seen from viewpoint
q and let Cu = q ∈ u Cq . We then define the gain g(u) of
Cluster of configurations. Large-scale environments may cluster u as,
result in a number of possible viewpoint configurations, |Cu |
which is impractical for planning purposes. To overcome g(u) = , (3)
|C|
this problem and easily identify those areas of the surface where |Cu | denotes the cardinality of the set Cu .
which are more promising, the viewpoints are grouped into
clusters uj , j ∈ {1, 2, . . . , Nc }, depending on their location
in space. We denote by U = {u1 , u2 , . . . , uNc } the set 4.2 Graph-based path planner
of all clusters. We say that configuration qi belongs to a
generic cluster u if ∃ qj ∈ u s.t. d(τij ) < dν , where d(τij ) The planner schedules the visit of clusters of a given
TSDF map. The associated TSP is solved and the path
denotes the length of the path τij between qi and qj on
which maximizes an utility function is calculated. The
the directed graph we will define in Section 4.2, and dν is
utility of visiting a cluster depends on its gain and on
an upper bound on the distance. If no neighbors are found,
the cost of reaching all its viewpoint configurations. The
dν is increased up to the maximum value of dmax ν .
cost corresponds to the length of the feasible paths which
Cluster evaluation. For a given path to follow, and connect configurations computed with a PRM planner.
depending on the current TSDF map and surface shape, To formalize this idea, let us introduce the weighted
some viewpoint configurations will be more “useful” than directed graph G = (U, E, {auv }(u, v) ∈E ), where U is the
others. A possible way to determine the priority of a set of nodes (in our case, the clusters), E is the set of edges,
viewpoint, is to measure its level of informativeness. For and {auv }(u, v)∈E is a collection of weights for the edges E
instance, in the classical ray-tracing method proposed (see Fig. 3). Each edge euv ∈ E is directed and links
cluster u to v, with u, v ∈ U . Let us assume that the initial
configuration of the UAV belongs to one of the clusters of
G, i.e. q0 ∈ U . Let qk be a configuration of cluster u, and
ql , qm two configurations of cluster v. Then, the weight
auv between cluster u and v is the 6-tuple defined as,
auv = τkl , τlm , g(v), d(τkl ), d(τlm ), fuv ,

(4)
where
• τkl denotes the path from qk ∈ u to ql ∈ v, i.e. the
path between cluster u and v. In particular, we select
the configuration ql in v, which guarantees that τkl is
Figure 2. [left] Two-dimensional example of voxel contour v (filled the shortest possible path,
green square). Its 2D neighborhood is represented by a dashed
green square. Unknown voxels are black, occupied voxels are
• τlm denotes the shortest Hamiltonian path (God-
gray, and empty voxels are white. The reconstructed surface sil and Royle (2001)) including configurations in v,
is depicted as a blue segment, and the sensor configuration which starts at ql and ends at qm ,
and its frustum as dark blue triangles; [right] Direction from • g(v) is the gain of the cluster v, as defined in (3),
the contour nv , and corresponding viewpoint configuration at • d(τkl ) is the cost associated to the inter-cluster path τkl
distance δpose (light blue triangle). (the length of τkl ),
 • d(τlm ) is the cost associated to the intra-cluster
path τlm (the length of τlm ),
• fuv is the utility function defined as,
fuv = g(v) exp −λtc d(τkl ) − λic d(τlm ) , (5)

where λtc and λic are positive penalty terms for

the inter-cluster and intra-cluster costs, respectively,
which can be used to promote the visit of clusters far
apart or large clusters.
Note that the utility function (5), was originally proposed
in González-Banos and Latombe (2002), and it has been
adapted here to our new formulation. The information Figure 4. Gazebo simulation environments: [left] Scenario 1,
Powerplant; [right] Scenario 2, the Statue of Liberty.
gain provided by a cluster is penalized by the distance
to reach it and the travel cost associated to the visit of all The path of the UAV is computed with the Lazy PRM*
configurations belonging to it. planner from the Open Motion Planning Library (Şucan
The weights (4) of the directed graph G, quantify the et al. (2012)), which finds the shortest path between two
potential benefit of taking a path to pursue the 3D re- configurations by taking the structure of the TSDF map
construction (see Fig. 3). The goal is to maximize the into account (the collision radius is 1 m). Lazy PRM*
total utility function. Note that since G is directed, we allows multi-query path planning to all destination points,
should solve an Asymmetric Travelling Salesman Problem which is useful for reachable-path checking and distance
(Punnen (2007)). After converting it to a symmetric one evaluation, because of the reduced computational com-
(i.e. to a standard TSP), the Hamiltonian path is found plexity with respect to RRT (the average runtime is below
using the Lin-Kernighan heuristic (Helsgaun (2000)). 1 s). Model Predictive Control (Kamel et al. (2017)) has
Successive scans along the path allow to generate new been used to track the generated paths, the reference
viewpoint configurations, and the information gain asso- translational velocity being fixed at 0.5 m/s.
ciated to the visit of clusters changes as the map grows. The virtual environments considered in Scenario 1 and 2
The path is locally updated if new areas are discovered as are shown in Fig. 4. In Scenario 1, the Powerplant model 2
the TSDF map is completed, or if a cluster has already has been scaled to fit in a box of size 65 × 42 × 15 m3 (as a
covered. The path is fully recomputed if the Nc clusters result, the five flues have the same height). Because of its
have been visited (cf. Fig. 1). The reconstruction process narrow passages, high walls and roof, large flues and thin
stops when no more ISEs remain. gantries, Powerplant is challenging for both navigation and
reconstruction (occlusion problem). In Scenario 2, we con-
5. NUMERICAL EXPERIMENTS sidered a model of the Statue of Liberty 3 (20×20×60 m3 ),
which contains multiple sharp edges and fine details (dia-
The proposed surface-driven method has been validated dem and gown). The results of the numerical experiments
via realistic numerical experiments. We chose a benchmark are reported in Table 2. To obtain statistically-significant
industrial plant, which is widely used in the volumetric values, 10 trials per scenario have been carried out. As in
exploration literature (Scenario 1). In order to study the the exploration methods (cf. Section 2), we evaluated the
impact of the penalty terms in (5), on the reconstruction total length of the path of the UAV, and the completion
accuracy/completeness, we also considered a monumental time, i.e. the time necessary to cover the 3D environments.
statue (Scenario 2). The simulation parameters used in We ran our algorithm on a Dell Precision 7520 with 2.90
the two scenarios are reported in Table 1. To simulate GHz Intel Core i7 processor, 16 GB RAM and Quadro
the aerial vehicle, we used the ROS-Gazebo 1 environ- M2200 graphics card. The reconstructed 3D surface has
ment. For the quadrotor UAV with an on-board stereo been evaluated with CloudCompare 4 using the M3C2
camera, we leveraged the model provided by the RotorS (Multiscale Model to Model Cloud Comparison) algorithm
simulator (Furrer et al. (2016)). The TSDF volume was (Lague et al. (2013)). To quantify how well the surface
generated with the algorithm proposed in Zeng et al. has been recovered, we sampled dense point clouds on
(2017), where the reconstruction is performed with March- the reconstructed and ground truth (GT) meshes, and
ingCubes (Lorensen and Cline (1987)) and the weight we measured the deviation by performing a cloud-to-cloud
increment has been modified as detailed in Section 3. comparison (see Fig. 5). For a fair evaluation, we pruned
all the invisible surfaces of the GT mesh beforehand (e.g.
Table 1: Parameters used in the numerical experiments.
the interior floor and walls), and we restricted our analysis
Parameter Scenario 1 Scenario 2 to the exterior surface mesh only. We consider that a point
Voxel resolution ρv [m] 0.3 0.15 belonging to the GT point cloud has been covered by a cor-
Threshold Wth 0.3 0.3 responding one in the reconstructed cloud, if their absolute
Camera range [m] [1.6, 8] [1, 5] distance is less than the length√ of the half diagonal of a
Camera FOV [deg] (H, V) 90 × 60 90 × 60 voxel, i.e. less than emax = ρv 3/2, where ρv is the voxel
δpose [m] 4.7 3.6 resolution. Note that 2 emax is the maximum reconstruc-
dν [m] 2.0 2.5 tion error provided by the MarchingCubes algorithm by
dmax
ν [m] 5 5 default (see, Lorensen and Cline (1987)). The quality of
Penalty term λtc 0.3 0.17
2 http://models.gazebosim.org/
Penalty term λic 0.03 0.15 3 https://free3D.com/
1 https://ros.org/ , http://gazebosim.org/ 4 https://www.danielgm.net/cc/
Figure 5. [top] Scenario 1 and [bottom] Scenario 2: [left] Reconstructed mesh and 3D exploration path pn0 (blue) of the quadrotor
UAV; [right] Signed distance error. The colour coding shows the error in meters with respect to the ground truth, computed with
CloudCompare’s M3C2 algorithm.

the recovered surface is evaluated in Table 2, by reporting all the regions that are accessible to the UAV are covered.
the average and standard deviation of the signed distance In addition, in keeping with the recent analysis in Schmid
error with respect to the GT point cloud, and the root- et al. (2020), it turns out to be competitive with the state-
mean-square error (RMSE). of-the-art approaches in terms of overall 3D reconstruction
The choice of the penalty terms λtc and λic appearing quality (Schmid et al. (2020) report an RMSE of 6.4 ±
in the utility function (5), depends on the nature of 0.8 cm). However, further work is needed to perform a
the 3D environment where the UAV evolves. Scenarios 1 comparative study under identical simulation conditions.
and 2 are, in this respect, quite representative. In wide It is finally worth pointing out here, that the quality of
box-like environments as Scenario 1, the ISEs tend to 3D reconstruction is resolution dependent: in fact, it is
appear in the proximity of occluded regions and sharp inversely proportional to the size of TSDF voxels. A small
edges, and large extents of known surface may separate resolution amounts to a large number of voxels to be
these sites. To minimize the total distance traveled, inter- integrated in the TSDF map, which is a resource-intensive
cluster utility should then take priority over intra-cluster process. Therefore, if the quadrotor UAV explores a large-
utility, i.e. λtc  λic . On the other hand, the pedestal scale environment using only on-board sensing and pro-
of the statue excluded, Scenario 2 predominantly consists cessing, a trade-off between reconstruction quality and
of round surfaces and the average distance between two computational efficiency should be found.
clusters is much smaller than in Scenario 1. As a conse-
quence, similar penalty terms should be selected this time Table 2: Numerical results (averages over 10 trials).
(i.e. λtc ' λic , see Table 1). The upper bound dmax
ν on the Criterion Scenario 1 Scenario 2
distance between cluster configurations dν , changes during
the reconstruction, and its default value has been deter- Path length [m] 780 547
mined empirically by considering the spatial distribution Completion time [min.] 32 36
of viewpoint configurations. The algorithm by Song and emax [cm] 25.98 12.99
Jo (2018) exhibits similar completion times (around 35 Surface coverage [%] 91.5 92.3
min.) to ours, for Scenario 1. The gap is more important M3C2 Avg. error [cm] 0.14 0.29
in Scenario 2: in fact, our algorithm took 36 min., while M3C2 Std. dev. error [cm] 5.85 3.41
that of Song and Jo (2017), around 53 min. However, in RMSE [cm] 5.86 3.43
our case, the trajectory of the quadrotor UAV is longer
(780 m vs. 324 m, in Scenario 1), and more jagged. This is 6. CONCLUSIONS AND FUTURE WORK
not surprising, since the viewpoint configurations have
been generated for accurate 3D reconstruction and not for In this paper, we have presented a new surface-driven
navigation purposes as in Song and Jo (2018). Moreover, Next-Best-View planning algorithm for the exploration
no trajectory refinement (e.g. smoothing) is performed (to and inspection of large-scale environments with a UAV.
that effect, we plan to use a receding-horizon formulation In particular, a novel cluster-based 3D reconstruction gain
in future works). Nevertheless, our method guarantees that and cost-utility formulation has been proposed. Realistic
numerical experiments with ROS and Gazebo have suc- Karaman, S. and Frazzoli, E. (2011). Sampling-based algorithms for
cessfully validated the proposed method on two challeng- optimal motion planning. Int. J. Robot. Res., 30(7), 846–894.
ing outdoor environments. Klingensmith, M., Dryanovski, I., Srinivasa, S., and Xiao, J. (2015).
Chisel: Real Time Large Scale 3D Reconstruction Onboard a
There are several promising directions for further research Mobile Device using Spatially Hashed Signed Distance Fields. In
we would like to explore in the future. Before implementing Proc. Robotics: Science and Systems, volume 4.
our method in real-time on a hardware platform (a quadro- Kriegel, S., Rink, C., Bodenmüller, T., and Suppa, M. (2015). Ef-
tor UAV), extensive numerical experiments will be carried ficient next-best-scan planning for autonomous 3D surface recon-
struction of unknown objects. J. Real-Time Image Pr., 10(4),
out in the presence of noisy measurements and localization
611–631.
uncertainty. The full-attitude control of the aerial robot Lague, D., Brodu, N., and Leroux, J. (2013). Accurate 3D compar-
along a computed path, is another subject of ongoing ison of complex topography with terrestrial laser scanner: Appli-
research. Finally, we are planning to adapt our approach to cation to the Rangitikei canyon (N-Z). ISPRS J. photogramm.,
a multi-robot cooperative setting (cf. Corah and Michael 82, 10–26.
(2019)), in order to speed up simultaneous exploration and Lorensen, W.E. and Cline, H.E. (1987). Marching cubes: A high
3D reconstruction. resolution 3D surface construction algorithm. In Proc. ACM
SIGGRAPH Computer Graphics, volume 21, 163–169.
Monica, R. and Aleotti, J. (2018). Contour-based next-best view
ACKNOWLEDGMENT planning from point cloud segmentation of unknown objects.
Auton. Robot., 42(2), 443–458.
This work has been supported by ONERA DTIS and by Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D.,
the Hauts-de-France region, through the research project Davison, A.J., Kohli, P., Shotton, J., Hodges, S., and Fitzgibbon,
ScanBot, “Scanners Robotisés pour la Numérisation Au- A.W. (2011). KinectFusion: Real-time dense surface mapping and
tomatique du Patrimoine” (2018-2021). tracking. In Proc. 10th IEEE Int. Symp. Mixed Augmen. Reality,
volume 11, 127–136.
REFERENCES Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., and Nieto, J.
(2017). Voxblox: Incremental 3D Euclidean Signed Distance Fields
Bircher, A., Kamel, M., Alexis, K., Oleynikova, H., and Siegwart, for on-board MAV planning. In Proc. IEEE/RSJ Int. Conf. Intel.
R. (2016). Receding Horizon “Next-Best-View” Planner for 3D Robots Syst., 1366–1373.
Exploration. In Proc. IEEE Int. Conf. Robot. Automat., 1462– Papachristos, C., Kamel, M., Popović, M., Khattak, S., Bircher,
1468. A., Oleynikova, H., Dang, T., Mascarich, F., Alexis, K., and
Bircher, A., Kamel, M., Alexis, K., Oleynikova, H., and Siegwart, R. Siegwart, R. (2019). Autonomous Exploration and Inspection
(2018). Receding horizon path planning for 3D exploration and Path Planning for Aerial Robots Using the Robot Operating
surface inspection. Auton. Robot., 42(2), 291–306. System. In A. Koubaa (ed.), Robot Operating System (ROS),
Border, R., Gammell, J.D., and Newman, P. (2018). Surface Edge volume 3, 67–111. Springer.
Explorer (SEE): Planning Next Best Views Directly from 3D Pito, R. (1996). A Sensor-Based Solution to the “Next Best View”
Observations. In Proc. IEEE Int. Conf. Robot. Automat., 6116– Problem. In Proc. 13th Int. Conf. Pattern Recogn., volume 1,
6123. 941–945.
Bresenham, J.E. (1965). Algorithm for computer control of a digital Punnen, A.P. (2007). The Traveling Salesman Problem: Applica-
plotter. IBM Syst. J., 4(1), 25–30. tions, Formulations and Variations. In G. Gutin and A.P. Punnen
Chen, S., Li, Y.F., Wang, W., and Zhang, J. (2008). Active Sensor (eds.), The Traveling Salesman Problem and Its Variations, vol-
Planning for Multiview Vision Tasks, volume 1. Springer. ume 12, 1–28. Springer.
Connolly, C. (1985). The Determination of Next Best Views. In Schmid, L., Pantic, M., Khanna, R., Ott, L., Siegwart, R., and
Proc. IEEE Int. Conf. Robot. Automat., volume 2, 432–435. Nieto, J. (2020). An Efficient Sampling-based Method for Online
Corah, M. and Michael, N. (2019). Distributed matroid-constrained Informative Path Planning in Unknown Environments. IEEE
submodular maximization for multi-robot exploration: Theory Robot. Autonom. Lett., 5(2), 1500–1507.
and practice. Auton. Robot., 43(2), 485–501. Song, S. and Jo, S. (2017). Online Inspection Path Planning for
Şucan, I., Moll, M., and Kavraki, L. (2012). The Open Motion Autonomous 3D Modeling using a Micro-Aerial Vehicle. In Proc.
Planning Library. IEEE Rob. Autom. Mag., 19(4), 72–82. IEEE Int. Conf. Robot. Automat., 6217–6224.
Furrer, F., Burri, M., Achtelik, M., and Siegwart, R. (2016). RotorS Song, S. and Jo, S. (2018). Surface-based Exploration for Au-
– A Modular Gazebo MAV Simulator Framework. In A. Koubaa tonomous 3D Modeling. In Proc. IEEE Int. Conf. Robot. Au-
(ed.), Robot Operating System (ROS): The Complete Reference, tomat., 4319–4326.
volume 1, 595–625. Springer. Tabib, W., Corah, M., Michael, N., and Whittaker, R. (2016).
Godsil, C. and Royle, G. (2001). Algebraic Graph Theory. Springer. Computationally efficient information-theoretic exploration of pits
González-Banos, H.H. and Latombe, J.C. (2002). Navigation strate- and caves. In Proc. IEEE/RSJ Int. Conf. Intel. Robots Syst.,
gies for exploring indoor environments. Int. J. Robot. Res., 21(10- 3722–3727.
11), 829–848. Vasquez-Gomez, J.I., Sucar, L.E., Murrieta-Cid, R., and Lopez-
Hauser, K. (2015). Lazy Collision Checking in Asymptotically- Damian, E. (2014). Volumetric Next-best-view Planning for 3D
Optimal Motion Planning. In Proc. IEEE Int. Conf. Robot. Object Reconstruction with Positioning Error. Int. J. Adv. Robot.
Automat., 2951–2957. Syst., 11(10), 159.
Helsgaun, K. (2000). An effective implementation of the Lin– Yoder, L. and Scherer, S. (2016). Autonomous Exploration for
Kernighan traveling salesman heuristic. Eur. J. Oper. Rer., Infrastructure Modeling with a Micro Aerial Vehicle. In D.S.
126(1), 106–130. Wettergreen and T.D. Barfoot (eds.), Field and Service Robotics:
Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss, C., and Bur- Results of the 10th Int. Conf., 427–440. Springer.
gard, W. (2013). OctoMap: An efficient probabilistic 3D mapping Zeng, A., Song, S., Nießner, M., Fisher, M., Xiao, J., and Funkhouser,
framework based on octrees. Auton. Robot., 34(3), 189–206. T. (2017). 3DMatch: Learning Local Geometric Descriptors
Kamel, M., Stastny, T., Alexis, K., and Siegwart, R. (2017). Model from RGB-D Reconstructions. In Proc. IEEE Conf. Comp. Vis.
Predictive Control for Trajectory Tracking of Unmanned Aerial Pattern Recogn., 1802–1811.
Vehicles Using Robot Operating System. In A. Koubaa (ed.),
Robot Operating System (ROS) The Complete Reference, vol-
ume 2, 3–29. Springer.