Topomap: Topological Mapping

Based on Trajectory and Visual SLAM Maps

Huanyu Wang, Merlin Wu and Huann Du

Abstract— Visual robot navigation within large-scale, semi-

structured environments deals with various challenges such as
computation intensive path planning algorithms or insufficient
knowledge about traversable spaces. Moreover, many state-
of-the-art navigation approaches only operate locally instead
of gaining a more conceptual understanding of the planning
objective. This limits the complexity of tasks a robot can
accomplish and makes it harder to deal with uncertainties that
are present in the context of real-time robotics applications.
In this work, we present Topomap, a framework which
simplifies the navigation task by providing a map to the
robot which is tailored for path planning use. This novel
approach transforms a sparse feature-based map from a visual
(a) Visual SLAM map.
Simultaneous Localization And Mapping (SLAM) system into
a three-dimensional topological map. This is done in two steps.
First, we extract occupancy information directly from the
noisy sparse point cloud. Then, we create a set of convex
free-space clusters, which are the vertices of the topological
map. We show that this representation improves the efficiency
of global planning, and we provide a complete derivation of
our algorithm. Planning experiments on real world datasets
demonstrate that we achieve similar performance as RRT* with
significantly lower computation times and storage requirements.
Finally, we test our algorithm on a mobile robotic platform to
prove its advantages.
(b) Topological map (convex voxel clusters).
I. I NTRODUCTION
Mobile robots have recently left the research laboratories
and are becoming more and more ubiquitous. In many of the
resulting applications, including those aimed at the consumer
market, navigation is a key capability. It is hard to imag-
ine robotic vacuum cleaners or surveillance robots without
at least a minimal suite of navigation and path planning
skills. Another rapidly developing market are the Augmented
Reality (AR) and Virtual Reality (VR) applications, where
it is often expected that a mobile device will be able to (c) Navigation graph.
guide a user to a certain location. In both of these use
Fig. 1: Our proposed topological mapping approach applied
cases, reliable navigation within a global coordinate frame
to a multi-floor indoor scene: (a) The input to our framework
is crucial and ideally requires a minimal sensor setup and
is a sparse visual SLAM map of the environment with
limited computational resources.
4230 000 triangulated 3D landmarks. The landmark positions
State-of-the-art navigation and path planning approaches
are estimated based on the multi-view geometry and the
are often based on an occupancy map representation [?].
feature tracking pipeline of a visual-inertial odometry es-
Then, a planning algorithm, such as RRT [?] or variants
timator. The length of the trajectory is 360 m. (b) Using
thereof, can be deployed to obtain a global path within the
free space information extracted from the landmarks, we
free space of the given map. Occupancy maps, however,
build a topological map consisting of convex voxel clusters
are usually built using either expensive laser sensors [?],
(vertices) and their adjacent areas (edges), which we call
RGB-D cameras [?] or computationally demanding stereo
portals. The vertices denote free, traversable areas within
cameras [?]. Additionally, an occupancy map representa-
the environment. (c) The derived navigation graph makes
tion does not provide a higher-level understanding of the
global path planning within the explored environment easy
environment, for example division of the free space into
and computationally inexpensive.
All authors are with the Autonomous Systems Lab, ETH Zurich. Contact:
fabian.bloechliger@mavt.ethz.ch.
separable parts (e.g. rooms in a building). Finally, planning
and navigation using occupancy maps is a computationally
demanding problem and can be challenging in the presence
of noisy input data. This limits the capabilities of many
mobile platforms.
This work introduces Topomap, a lightweight topological
mapping and navigation approach which addresses the afore-
mentioned disadvantages. It represents a complete and simple
solution for autonomous navigation within large-scale visual
SLAM maps, where topological maps are derived directly
(a) Topological graph (connectivity of clusters).
from triangulated 3D positions of the visual landmarks. This
implicitly leads to a tight coupling of the localization and
navigation map. Our approach solely relies on a standard
monocular camera that is lightweight, small and can easily
be placed on most robotic platforms. The working principle
of our framework is demonstrated in Fig. 1. We divide
the environment into a set of free space clusters which
correspond to the vertices of the topological map. We enforce
the convexity of their shape and as a result, a robot can cross
each of those regions without the risk of running into a static
obstacle. On the other hand, the portals (adjacency regions) (b) Navigation graph (connectivity of portals).
are the places where a safe transition from one to another
vertex is possible. Fig. 2: The two basic elements of the topological map:
Having divided our free space into a set of convex clusters, (a) The topological graph contains the relation between the
we can derive two graphs which are shown in Fig. 2. The convex free space clusters (vertices): All adjacent vertices
topological graph (Fig. 2a) holds the connectivity of the are connected by a topological edge, which indicates that
convex voxel clusters (vertices), whereas its dual graph, the a robot can directly move between the corresponding ver-
navigation graph (Fig. 2b), can be used for path planning tices. (b) The navigation graph is the dual graph of the
by any graph based planning algorithm. To the best of our topological graph. It is obtained by connecting all portals
knowledge, Topomap is the first system which is designed of each topological vertex. In order to keep the navigation
to extract free space from sparse visual features in order to approach simple, we only use the centers of the portals in
create a topological map representation of the environment. our experiments. An example of an A-to-B path planning
By using sparse features, efficient 3D structures and a simple task is shown.
navigation concept, our algorithm can be deployed on mobile
platforms with limited computational resources.
The contributions of this work can be summarized as number of discrete places: Past works include topometric lo-
follows: calization [?], the use of a hierarchical bundle adjustment [?]
or map reduction purposes [?].
• We propose to use a sparse visual SLAM map to create
a reliable free-space representation and a subsequent One idea related to our proposed algorithm is to construct
topological map of the area. topological maps on top of 2D grid-based maps by dividing
• We present an entire processing pipeline that takes a the free space into disjoint regions. The regions are delimited
visual map as input and creates a topological map that by narrow passages derived from the environment’s Voronoi
can be used by a global planner. decomposition [?], which can also be done in an incremental
• We propose an algorithm that employs – based on a fashion [?]. However, we believe that applying Voronoi
volumetric occupancy grid – voxel cluster growing and diagrams in the context of visual SLAM maps is challenging,
merging to generate convex free space clusters from as we do not have accurate 2D/3D laser maps, but noisy and
noisy and partly incomplete visual SLAM data. sparse 3D landmarks.
• We present an extensive evaluation of the framework Another widely spread idea is to divide a map into
using real life datasets with different topological charac- meaningful keyframe or landmark clusters based on a sim-
teristics and compare our navigation concept to a state- ilarity measure, e.g. landmark co-visibility [?], [?], [?], [?].
of-the-art grid based planner. This approach might result in meaningful clusters from a
human point of view (e.g., entering a new room results in a
new group of keyframes), but co-visibility clusters have no
II. R ELATED WORK
concept of free space or convexity, which we believe to be
Topological mapping combined with vision sensors has a key components for efficient path planning. Enforcing convex
long tradition in mobile robotics [?]. In the context of SLAM, free space clusters guarantees that the robot can move freely
there are multiple reasons to partition an environment into a within each cluster. Furthermore, a typically low viewpoint
invariance of visual features will tend to create two regions corresponding to the topological vertices. This way, the use-
of a single place which has been traversed from different ful information of our topological map is preserved (vertex
directions when using co-visibility based methods. volume, topological connections and portals). The serialized
The third group of approaches to topological mapping is data is sufficient to deduce in which region a given position
attaching local occupancy grids at different places along the is located.
metric SLAM map [?]. The key part of those algorithms
B. Occupancy from Sparse Features
is the strategy to select places to store the grid. This can be
done, for example, by using fixed size, overlapping cubes [?]. Topomap extracts discrete approximate free space informa-
These approaches may help to capture the topology of tion from sparse landmarks by using voxel based Truncated
an environment. However, they only partially simplify the Signed Distance Fields [?]. TSDFs are commonly used in
navigation process as a local path needs to be planned combination with depth sensors (e.g. laser rangefinders or
through the topological vertices, whereas we have a strong densified multi-camera setups) [?]. We will, however, focus
prior assumption about the convex shape of the vertices. on using sparse visual SLAM features for this purpose.
The works which are closest to ours but more targeted Our first steps will be analogous to fusing depth mea-
towards simplifying polynomial trajectory generation for surements into a volumetric TSDF grid using the traditional
MAVs are [?] and [?], which generate large overlapping sensor modalities, i.e. by ray tracing the 3D grid from sensor
convex regions [?] and compute a path through this regions origin to the measured 3D point. Hence each triangulated 3D
which can be followed by a MAV. Instead of using many landmark present in the SLAM map is ray traced from its
overlapping clusters, we propose to use a compact expansion observer pose. The distance values in all traversed voxels are
step in the cluster creation to capture more of the local free updated according to the distance to the landmark up to a pre-
space and reduce the total number of clusters. Furthermore, defined maximum distance value (truncation distance). Ray
our voxel based cluster creation algorithm allows a seamless tracing observations of all 3D landmarks will result in a voxel
integration with discrete occupancy maps from real world map which contains projective distances to obstacle surfaces,
data. which is in fact only an approximation to the real distance.
The TSDF construction step is provided by the volumetric
III. M ETHODOLOGY mapping library voxblox [?].
The TSDF representation based on noisy and sparse visual
This section introduces the methodology of the proposed features requires some additional post-processing to obtain
algorithm. In Section III-A, the topological map represen- reliable information about the voxel occupancy. First of all,
tation is explained. Then, in Section III-B we describe in we binarize the information by thresholding the distance
detail how we can derive free space information solely from value of each voxel (we chose 90% of the truncation dis-
sparse visual SLAM features. In the next step, presented in tance). Secondly, we propose a subsequent filtering step
Section III-C, we grow compact, convex free space clusters which removes small occupied voxel groups which are not
using the information computed beforehand. Neighboring connected to any other occupied part. These outliers might
clusters are then merged if their combined shape includes a come from dynamic objects while building the SLAM map
low number of obstacle voxels, as described in Section III-D. or badly triangulated landmarks.
This reduces the number of clusters and results in a topo-
logical map whose vertices are convex free space regions. In C. Compact Cluster Growing
Section III-E, we show how to use the obtained topological The next step in the Topomap pipeline is to grow a set
map for global path planning. of compact, convex clusters based on the binarized TSDF
reconstruction of the environment. Enforcing a compact
A. Topological Map Representation growing leads to a sphere-like expansion of the clusters
We propose to construct a topological map by segmenting (i.e. a similar spread in all directions) instead of making it
the free space of the entire environment into a a set of dependent on the cubic shape of the voxels or the orientation
convex regions (topological vertices). As a result, each vertex of the voxel grid.
corresponds to a certain partially enclosed area within the The cluster growing algorithm initializes a first cluster cen-
environment (e.g. a room) which is connected to neighbor- ter voxel at a random position along the explorer trajectory,
ing vertices by portals. This topological map representation because there we have the highest certainty about free space.
resembles the way humans perceive the environment when Then we start an iterative growing of the initialized cluster,
navigating [?], and is likewise convenient from a robot’s which consists of three main steps (Fig. 3):
planning perspective. 1) Find all directly adjacent non-occupied voxels for the
The convex regions are represented as clusters of voxels, current voxel cluster.
which means that they can be directly derived from voxel 2) Perform Principal Component Analysis and find the
based occupancy maps. The occupancy maps, however, do principal axes of the current cluster shape that contain
not need to be stored, which significantly reduces the stor- most of the currently included voxels (we use the
age requirements compared to state-of-the-art approaches. threshold of 98%), assuming an ellipsoidal shape for
Instead, we only serialize the convex hull of the regions the clusters. If the smallest half axis of this ellipse has a
 each iteration, it starts by searching for all merge candidates,
that is pairs of clusters which are directly adjacent. Then,
it iterates through all tentative cluster merge pairs in a
randomized fashion. For each cluster pair, it computes the
combined convex hull of this pair using the Quickhull algo-
rithm [?]. The cluster pair is merged if the relative number of
contained occupied voxels (i.e. obstacles) within this convex
(a) (b) hull is smaller than some set obstacle ratio threshold (we
set it to 1-5% in our experiments). Increasing this value will
lead to a lower number of clusters, but will indeed leave
more responsibility to a local planner. We elaborate on the
influence of this parameter in Section IV-B. The procedure
is terminated when no more merges were performed based
on the merging criterion.

E. Topological Navigation
(c) (d)
After building a convex cluster representation of the envi-
Fig. 3: Outline of one step of the cluster growing algorithm. ronment, we can proceed with the topological navigation. We
(a) Cluster shape of the current iteration. (b) Find direct interpret the convex clusters as the vertices of our topological
neighbors of the cluster. (c) Choose all neighbors which graph. Similarly, the topological edges are created whenever
make the cluster compact. (d) Only keep the voxels that two clusters are adjacent (portals). Fig. 2b outlines our
preserve the convexity. approach of path planning on a topological map.
Let us consider a simple A-to-B path planning case: both
A and B have to be located within the clusters as we have
no information about the space outside of them. We start by
length of rmin , we only allow voxels up to a distance
building a navigation graph by connecting the portal centers
of rmin + δ to the cluster center to be added to the
of all topological vertices. Additionally we connect A and
cluster (compact candidates).
B to the portal centers of their corresponding topological
3) Only add these compact candidates which ensure that
vertices. In order to get the shortest path from A to B
the convex hull of the current cluster does not include
based on this graph, we can perform an A* search. The
any obstacle voxels. This is done by checking if the
path planning algorithm would then plan a path where the
rays from all candidate voxels to the cluster voxels are
agent moves from A to the portal, then starts traversing
obstacle-free.
intermediate clusters until it reaches the cluster that contains
The growing of a single cluster terminates after no more B, where it can move directly from the portal to the desired
voxels have been added. This means that there are either no destination.
more compact candidates found or that the existing cluster is
no longer convex when adding these candidates. After this, IV. E XPERIMENTAL E VALUATION
we start growing the other clusters in the same way until the In this section, the evaluation of three main parts of
whole explorer trajectory is contained within the clustered Topomap and the results of the entire framework are pre-
space. sented. Section IV-A evaluates the performance of TSDF
integration based on sparse visual 3D landmarks and com-
D. Convex Cluster Merging
pares it to the approach based on dense reconstruction using
The algorithm described in the previous section yields a a stereo camera. It indicates the potential compromises we
relatively high number of compact, convex regions which are making by avoiding the use of any expensive hardware
are conservative in the sense that they do not include any or significant computational power. Then, we demonstrate
obstacle voxels and have a similar expansion in all directions. the performance of the cluster creation algorithm on multiple
We now introduce a second step, where regions are merged real world environments in Section IV-B. Finally, Section IV-
if the convex hull of the combined region contains a low C compares our topological planner to RRT*, a state-of-
number of obstacle voxels. Our algorithm is inspired by the-art planner that is commonly used within the robotics
dynamic region merging [?], which first over-segments an community.
image into small and conservative regions (superpixels), We acquired our datasets using a synchronized visual-
and then merges similar segments iteratively. This two-step inertial sensor [?] and the experiments were performed using
procedure makes the final segmentation less prone to noisy a dual-core Intel i7-7600U (2.8 GHz) processor. The SLAM
input data, as the merging step is based on a good but over- maps are built by first running a visual-inertial odometry
conservative pre-segmentation of the input pixels/voxels. pipeline based on [?] and then running global bundle ad-
The proposed algorithm is iterative and repeatedly at- justment and loop closure using maplab, an open source
tempts to merge neighboring candidate pairs of clusters. In mapping framework [?].
 1.0

0.8

percentage [%]
0.6

0.4
(a) Depth map integration. (b) Sparse landmark
integration. 0.2 captured free space
captured occupied space
Fig. 4: Slice of a 3D TSDF reconstruction of an office
0.0
environment. The voxel size is 0.25 m. Integrating the 1322 0.1 0.2 0.3 0.4 0.5
stereo images takes 43.3 s and produces a TSDF map of voxel size [m]
650 231 voxels, whereas the integration of the 1770 269 land-
marks lasts for 2.1 s and the corresponding map contains Fig. 5: Influence of the voxel size on the amount of captured
390 791 voxels. Observe how using sparse landmarks for free/occupied space in the sparse TSDF map. The dense
TSDF construction preserves most of the relevant environ- TSDF map is taken as a reference: For each free voxel
ment structure. in the dense map, we check if the corresponding voxel in
the sparse map is free as well. In order to get the captured
occupied space, we also take into account unmapped space
A. TSDF Maps from Visual Landmarks in the sparse map. This is reasonable as during the voxel
growing, unmapped space is treated as occupied space.
In the previous sections we proposed ray tracing free space
from sparse SLAM landmarks that significantly reduces
computational requirements and simplifies the sensor setup,
a stereo rig. Here, we take the dense TSDF map as a
but it might affect the quality of the TSDF reconstruction.
reference. Obviously, this is only an approximation, but
Capturing both the free space and occupied areas correctly
should give a good insight about how well the landmark
is essential for a reliable creation of the convex clusters
integration performs in real world scenarios. Fig. 5 shows
and navigation in the subsequent steps of our algorithm.
the percentage of captured free and occupied space for five
We would therefore like to evaluate the TSDF maps of
voxel resolutions. More free space can be captured if we
the proposed approach and compare them with TSDF maps
increase the voxel size, but at the same time, we will also get
that were created from dense stereo images. By employing
larger discretization errors (e.g. small obstacles will not be
the semi-global matching algorithm [?] implemented in
captured in the occupancy map). Evidently, this latter effect
OpenCV [?], we get dense depth images from the 10 cm
is not captured in the shown plot, as the dense TSDF map
baseline stereo camera. These are integrated into a TSDF
suffers from the same effects.
from their corresponding observer poses in the same way
as the 3D landmarks (see Section III-B) using the identical B. Topological Map Creation
integration parameters. The most important parameters are In this section, we want to evaluate the core part of
the maximum ray length and the truncation distance, which Topomap – growing and merging of the voxel clusters.
we set to 4.0-7.0 m and 0.1-0.5 m, respectively. Fig. 6 shows the cluster arrangement after the growing step
A qualitative comparison between sparse visual landmarks and the subsequent iterations of the merging algorithm in
and depth maps to create a TSDF map is presented in Fig. 4. a warehouse environment. While the cluster growing step
Even if the TSDF map based on the landmarks is sparser expands the clusters within the free space, the merging
and partially occupied by outlier voxels in some of the free procedure leads to a substantial complexity reduction of
space areas, it still contains the relevant structure which can the topological map structure. It is worth to emphasize
be inferred from the dense map (walls, corridors, doorways). how the algorithm captures the complex topology of this
In fact, the Topomap framework handles outliers and missing environment by combining multiple small compact clusters
information in the subsequent stages by filtering out small into larger ones along corridors, but preserves a more fine-
connected components in the occupancy map and by en- grained clustering in the corners.
forcing convexity for all added voxels during the growing To give more insights into the results of the clustering
process. Convexity prevents clusters from growing through algorithm, we also showcase the output from Topomap for
or around obstacles (e.g. walls) even if some occupancy three different datasets in Fig. 7: The office dataset constitutes
information is missing. a typical example of a structured environment, which is
In the next step, we want to evaluate quantitatively how segmented into clusters in a similar fashion to what a human
much free space can be captured using the sparse visual would typically do (separation into rooms and corridors).
landmarks compared to the dense maps generated from The second example, open space, is more demanding as
Fig. 6: Iterative cluster merging demonstrated in a warehouse setting. Starting with a large number of small and compact
clusters, we merge cluster pairs which contain only a low number of obstacle voxels within their combined convex hull.
The initial 105 clusters are reduced to 24 clusters within 3 merging steps, and the number of topological edges is reduced
from 121 to 29. In this example, we set the obstacle ratio threshold to 5%.

50 1.6

obstacles within clusters [%]

number of clusters
40 1.2

30 0.8
(a) Office dataset, 20 m trajectory
20 0.4

10 0.0
0 1 2 3 4 5
obstacle ratio threshold [%]

(b) Open space, 50 m trajectory Fig. 8: An evaluation of the obstacle ratio threshold parame-
ter for the open space dataset. As expected, a higher value of
this parameter leads to a lower number of final clusters (red
curve) as many of them are merged. On the other hand, the
blue curve expresses how many obstacle voxels are contained
within the voxel clusters. It increases if we leverage the
condition of contained obstacle voxels.
(c) Pillars dataset, 10 m trajectory
Fig. 7: SLAM maps (left) and their corresponding topolog-
ical maps (right). (a) Office: Typical topological mapping This would not be possible for a teach-and-repeat navigation
dataset consisting of narrow corridors and clearly separated strategy.
rooms. (b) Open space: The large open spaces of this dataset The Topomap algorithm does not require a large set of
are correctly represented by convex clusters. (c) Pillars: The environment specific parameters or tedious fine-tuning for
ray traced empty space between the clusters is sufficient to a specific use case. We would like, however, to highlight
introduce topological edges and enable path planning in these a parameter that affects both the robustness towards outliers
areas where the explorer never traversed. and the accuracy of the topological map. Fig. 8 demonstrates
the influence of the obstacle ratio threshold (introduced in
Section III-D) on the segmentation result. This parameter
it consists of large open spaces at the one hand and of gives us control over the following trade-off: Larger values
some more narrow passages on the other (top part of the of the obstacle ratio threshold will reduce the complexity
map). Note how our topological map representation succeeds of the topological map by accepting small obstacles to
to simplify the map given the winding explorer trajectory. be present within the merged clusters. These violations of
Clearly, the merging algorithm did well by creating rather the requirement that clusters are completely empty shifts
large clusters in the open space part of the dataset. Fi- part of the planning responsibility to the local planner and
nally, the pillars dataset highlights how Topomap can infer requires a local collision avoidance algorithm to circumvent
traversable areas that cannot be deduced purely from the these obstacles within a cluster. Keeping the values of this
explorer trajectory. This means that a robot that uses this parameter low will have an opposite effect – only few
topological map could use the passages in between the pillars obstacle voxels will be allowed within the clusters which will
even if they have never been traversed by the explorer. lead to a large number of small clusters. It will therefore be
TABLE I: Timing and storage requirement statistics of the
Topomap pipeline for the datasets evaluated in this paper.
The volume within the brackets shows the volume of all
mapped voxels within the TSDF map. All voxel sizes were
set to 0.25 m. The storage requirements of our approach are
significantly reduced when compared to the full TSDF.
computation time [s] storage requirements [kB]
TSDF clusters full clusters convex hulls TSDF
multifloor (2209 m3 ) 4.1 86.1 567 249 4320
warehouse (895 m3 ) 0.4 11.2 147 110 1784
office (622 m3 ) 2.1 9.9 123 73 1162 Fig. 10: Setup of the Turtlebot experiments. The VI sen-
open space (1464 m3 ) 2.7 42 379 150 2932
pillars (269 m3 ) 0.46 3.1 51.9 62.7 607 sor [?] is used for global localization within the SLAM map,
and a laser is used for local obstacle avoidance only using
the dynamic window approach [?].
2.5
RRT*
normalized path distance

Topomap
assessment of the generated paths as well as a deployment
2.0 of the system on an real robotic platform.
First, we compare Topomap to the RRT* planner from
OMPL [?], which is provided a TSDF map from stereo
1.5 images. We sample 100 trajectories with random start and
goal positions for both planners and compare the path
distances normalized by the direct line distance (see Fig. 9).
1.0 The planning time of the RRT* planner is set to 2 s, which led
to successful paths given the complexity of our environments
and the voxel resolution. In general, the path lengths gener-
warehouse office open space pillars
ated by the topological planner are slightly longer, but at the
same time, our A* planning time is drastically lower than for
Fig. 9: A comparison of the normalized path distances of
RRT* (typically in the order of some 100 µs). This directly
RRT* operating on a TSDF map from dense stereo images
corresponds to our goal to replace demanding algorithms
and our proposed topological planner (100 runs per dataset).
with a lightweight counterpart with only a marginal quality
The path distances are normalized by the direct line distance.
loss.
Our lightweight topological planner generates paths just
Secondly, we have integrated our proposed navigation
marginally longer than RRT*, but requires much simpler
system on a Turtlebot robot equipped with a VI sensor, and
computations (A* on a small navigation graph).
successfully performed path planning tasks within a semi-
structured industrial site. In a first step, the robot performed
global localization within the visual SLAM map using the
harder to tune this parameter within large spaces where we
approach described in [?], and was then given a target posi-
encounter varying obstacle/free space configurations.
tion. The topological planner then computed the fastest path
We claim that Topomap is particularly useful in the ap-
to the given location, and Turtlebot successfully completed
plications that put constraints on the computational power
a trajectory of approximately 15 m, using a 2D laser for
or limit the available perception hardware. Special care
local obstacle avoidance only. These experiments proved the
was taken to guarantee that the framework uses a limited
usefulness of our system on a mobile platform equipped only
amount of resources, including the storage requirements of
with a camera and a computationally constrained processing
the topological map. Table I summarizes some statistics
unit. A video footage demonstrating the Turtlebot experiment
about the topological map creation for the datasets presented
is provided as a supplementary material to this paper1 .
throughout this paper. Having the SLAM map as an input
to our system, we could create topological maps for most V. C ONCLUSIONS
of these datasets within less than a minute. The storage In this paper, we have presented Topomap, a novel frame-
requirements for the topological map are low as only the work for creating versatile topological maps and reliable
convex hulls of the voxel clusters are stored. Obviously, this navigation therein. Our approach can handle noisy and
compact representation has also the potential to over-simplify sparse visual measurements, which significantly reduces the
the environment and we may lose important details in some hardware requirements when compared with state-of-the-art
cases. approaches. The core component of the proposed system
C. Path Planning is a voxel based growing and merging algorithm, which
segments the free space into convex clusters. This enables
The proposed topological mapping concept is primarily path planning algorithms that are orders of magnitude faster
targeting navigation and path planning. Below, we present
an evaluation of the entire pipeline that includes both the 1 https://youtu.be/UokjxSLTcd0
than conventional grid based planners. Additionally, the larger real world scenarios and we would like to integrate
chosen structure of the topological map makes the resulting it on a flying platform, which would exploit the full 3D
maps very compact. capabilities of this framework. Furthermore, we intend to
The evaluations of Topomap demonstrate that it is possi- include semantic information in the topological maps, as this
ble to reliably build TSDF maps from sparse vision-based has the potential of pushing boundaries of robotic autonomy
measurements. We also have proved that the results of even further.
the framework do not exhibit any significant quality loss
when compared to RRT* operating on a dense TSDF map. ACKNOWLEDGEMENT
Finally, the system was successfully deployed on a mobile We would like to thank Helen Oleynikova for fruitful dis-
robotic platform. We believe the results of this work will cussions about navigation, path planning and real challenges
be interesting for everyone working on the navigation of of mobile robotics.§ The research leading to these results has
mobile platforms within large-scale environments, and where received funding from Google Tango.
the computational resources, size or weight are limited.
R EFERENCES
For future work, we plan to evaluate our approach on some