Lanelet2: A high-definition map framework for the future of

automated driving
Fabian Poggenhans1 , Jan-Hendrik Pauls2 , Johannes Janosovits2 , Stefan Orf1 , Maximilian Naumann1 ,
Florian Kuhnt1 and Matthias Mayr1

Abstract— Although accurate and comprehensive maps are For highly automated vehicles, not only the lane itself is
indispensable for highly automated driving, especially in com- relevant. Information about the complex environment of a
plex urban scenarios, there are hardly any publications in which vehicle such as bicycle lanes and sidewalks must also be
requirements for these maps are discussed. In our opinion,
such maps must meet high demands in terms of accuracy, made available and should therefore be provided by a map.
completeness, verifiability and extensibility, so that the resulting As a result, many parts of the processing chain of a highly
complexity can only be handled by an enclosing, carefully automated vehicle depend on precise map information, but
designed software framework. In this paper we therefore intro- each with a different focus (see Fig. 1). The elements of the
duce the open-source map framework Lanelet2 implemented in road on which the map is based and their exact position,
C++ and explain the underlying concept. The goal of Lanelet2
is not only to be usable for typical, isolated applications such such as markings, traffic signs and curbs, are relevant for
as localization or motion planning, but for various potential localizing as well as validating the map. During behavior
applications of maps for highly automated driving. On the generation, however, the relevance of these elements for the
basis of both abstract and real examples we show the concrete vehicle is important, e.g. which traffic light is valid for a
structure of Lanelet2 maps and its use for automated driving. particular lane. For the prediction of other road users it
must be clear which traffic rules apply to them and where
I. INTRODUCTION
they can move to. This is especially true for pedestrians,
There is no doubt that maps are essential for highly auto- whose behavior is difficult to predict with sensors alone [5].
mated driving. High-definition maps, which provide accurate Detailed routing requires knowledge of all lanes and where
information about the surroundings of a vehicle, were an they lead. It must be clear where lane changes are possible
essential component in all major automated driving projects or even mandatory. Other requirements are special situations
(cf. [1]–[4]) so far. The reason for this is not only that it such as emergency or parking maneuvers. For this purpose,
compensates for the inadequacies of the sensors, but also the map must provide information beyond the actual lanes.
that maps are important for providing information about For the simulation of automated vehicles, maps must be very
regions that cannot be observed by sensors, whether due to detailed to be able to test under the most realistic conditions.
occlusion or insufficient sensor range. In addition, a correct In addition, high quality standards must be met: maps must
interpretation of the observed data is required, which copes be complete and accurate, up-to-date and verifiable within a
with the uncertainties of the sensors. Maps provide the few centimeters to ensure safe automated driving.
reliable information needed to understand a scene correctly. In view of the increasing complexity of HAD maps, it is
Finally, maps allow to transfer knowledge from previous also clear that it is not sufficient to simply store the relevant
journeys and thus represent an additional level of redundancy. information in a map. In order to be able to consistently
With increasing progress in the field of automated driving, interpret the information and to determine its implications
the maps have to be refined more and more and at the same for the ego vehicle and other road users, good software
time meet higher and higher quality requirements. While support in form of a special software framework is necessary
the coarse road course is sufficient for navigation devices to ensure a consistent view of the map. Ideally, such a
and the position and number of lanes can be sufficient for framework should provide simple interfaces for the various
partially automated driving on highways and most country different perspectives from which the data contained in a
roads, maps for the city center must provide considerably map can be viewed.
more information. The term ”high-definition” (HD) map is Given the high demand for highly accurate maps for
often misleadingly used for the latter two, but it is not many different applications, it is all the more surprising that
precisely defined. In this paper we therefore use the term there are hardly any publications discussing requirements for
lane-level accurate map for the first, while maps that enable maps for HAD. Accordingly, there are no map frameworks
safe highly automated driving (HAD) even in complex city that meet the demand. Commercial map providers claim
center scenarios will be referred to as HAD maps. to already provide HAD map data, but we will show that
these maps do not meet the requirements. In addition to the
1 FZI Research Center for Information Technology, 76131 Karlsruhe, Ger-
completeness of the map information, an easy evaluation of
many {poggenhans, orf, naumann, kuhnt}@fzi.de the data during the journey is also decisive. We would like
2 Institute of Measurement and Control Systems, Karlsruhe Institute
of Technology, 76131 Karlsruhe, Germany {jan-hendrik.pauls, to deal with both aspects in the following pages which form
johannes.janosovits}@kit.edu the basis of the map framework Lanelet2.
 in which these formats are used. Examples include the Uni-
fied Map [11], which was developed for trajectory planning.
Here, moving objects, road boundaries and traffic rules, such
as traffic lights, are uniformly represented as obstacles that
restrict vehicle movement.
Another example can be found in [12]. The aim of this
publication is to show uncertainties in the representation
of a map created by the sensors while driving. Similar to
the already mentioned map formats, the road is represented
here as a projection of all elements onto the center line.
Intersections are not covered.
Most of the remaining publications where new map for-
Fig. 1: Applications and required interaction with HAD maps
mats are introduced deal with localization. Most likely be-
cause accurate localization requires very detailed information
about observable landmarks. The map formats mentioned
II. RELATED WORKS
above do not provide sufficient support for this. An out-
A. Existing Map Formats and Frameworks standing example is [13], because it not only introduces a
Most available map formats used for automated driving map representation, but also formulates requirements that
have their origins in the time before this became a popular a map format suitable for HAD must meet. In addition to
topic. These include, for example, the maps of commercial landmarks suitable for localization, the main requirements
providers TomTom and Here [6], [7], but also the free listed are information on the geometry and direction of travel
mapping project OpenStreetMap (OSM) [8]. Due to the of the lanes, especially in intersections, information on height
increased requirements, providers are inevitably pursuing a profiles and the resulting occlusions, and information on the
top-down approach: roads are represented by an imaginary prediction of other road users.
center line, as is already used in navigation devices. More The map framework Liblanelet [14] follows a very differ-
information about lanes, for example, is added by adding ent approach. For the first time used for Bertha-Benz drive
certain attributes to this center line. For example, the po- [1], this is a map format that was specially developed for
sition of traffic lanes and the form of the road border are HAD, especially for motion planning. It is based on atomic
also stored. As requirements increase, additional attributes lane sections, Lanelets, which together form the road through
are added. This leads to an extremely complex, implicit their neighborhood relationships. This format made a simple,
representation of the information, since, for example, the explicit representation of the lanes possible. Traffic rules
absolute position of the roadsides can only be determined are represented by so-called Regulatory Elements, which
indirectly by offsetting the center line left and right with the represent, for example, right of way, traffic lights and stop
width of the road. This complexity is further increased in lines. Liblanelet also offers a library in C++ with interfaces
intersections, as there is no clear center line. In OSM as an for routing, motion planning and prediction of other vehicles.
open format, the quality of the data varies greatly [9], as these However, the format has certain weaknesses related to the
are created and maintained by volunteers. However, suitable, fact that it was designed for use with a specific, previously
freely available editors exist to supplement existing maps. known route. For example, lane changes are only possible at
For the commercial map providers, no editors or software predefined points. Overtaking is therefore also not possible.
libraries are publicly available. Furthermore, many other fields of application described in
The OpenDRIVE [10] framework occupies a similar po- Fig. 1 are not supported, such as localization or special ma-
sition. This format was originally developed for a uniform neuvers. In this paper we therefore present a generalization
description of the road and its surroundings in driving of this map format that eliminates these weaknesses.
simulators, but is now also used for HAD. The information is
represented very similarly to the formats already mentioned. B. Usage of Maps
Although the specification is open, there are no freely To find out how maps are used in literature and what
available libraries for interpreting and processing the data. requirements have been placed on them, we have evaluated
The above-mentioned map formats with lane-level accu- 32 publications of the last five years from various research
rate information have also been successfully used for driving directions in HAD that explicitly state the use of high-
on motorways. An example of this is [4]. However, the resolution maps. Most of them (13) dealt with the topic
authors themselves conclude that “there is still a lot of work of localization, five with the topic of map generation, four
to be done, especially in the area of validation/certification described a system architecture, four dealt with motion
and the generation of large-scale digital maps”. planning and four with prediction. One dealt with scene
Due to the shortcomings of existing map formats, map understanding, one with behavior generation and detection
representations have been introduced and used in recent of map deviations. Most approaches (13) used a self-created
years that are specially tailored for a particular application or map format, especially for localization methods, where the
publication. Consequently, there are few further publications map was often created using the same technique. This
already indicates that there is a certain lack of suitable to move within their lanes or change lanes if traffic regula-
map formats. Most of the other publications used OSM (7), tions permit this. Similarly to behavior generation, the set of
although in three cases the format was extended to provide possible maneuvers must be determined. For this purpose,
more information. In five approaches Liblanelet was used. however, the rules must be taken into account depending on
They dealt with motion planning, localization and prediction the type of road user. For example, special traffic rules often
of other road users. Two approaches used OpenDRIVE, one apply to buses, which allow them to use additional turn-offs
maps from HERE. In the other four cases, the map format or separate lanes. A map should therefore contain traffic rules
was not described. for any type of road user, not only for the ego vehicle. This
This small investigation suggests that maps are indeed includes in particular pedestrians, cyclists or trams. These do
an important component in many applications. However, the not move on lanes, but a map must be able to tell where, for
fact that mostly self-created map formats are used for this example, pedestrians can be expected and in which directions
or existing ones are adapted, clearly indicates a deficiency. they can move.
The work that is spent in the ever new construction of
own, specialized map formats could be used much more B. Lane and Environment
meaningfully – for example by contributing to a common, Precise knowledge of the lane geometry is crucial for path
public library. Therefore, in the following section we will planning. Knowing the center line alone is not sufficient, as
give some thought to the requirements that a map framework the trajectory should be able to be adjusted depending on
for highly automated driving must meet and how this can be the speed, for example in curves for comfortable driving. In
solved with Lanelet2. the city center, obstacles such as parking vehicles protruding
into the lane often have to be avoided without endangering
III. DESIGN CONSIDERATIONS oncoming traffic. Since the localization and the actuators of a
As already mentioned in Section I, HAD places many de- vehicle are also not arbitrarily precise, it must also be known
mands on maps and how information is represented in them. which distance to the center line can still be tolerated without
In this section we will present essential requirements and leaving the lane. As already suggested in [14], it therefore
our proposed solutions by looking at common applications makes sense to map the lanes completely with their left and
in automated driving that make use of maps and identifying right borders.
their requirements. These applications can be divided into An often forgotten problem, however, is that highly au-
three groups (see Fig. 1): Applications that focus more on tomated driving does not only involve normal driving on a
the road network, applications that need detailed information public road, but also special maneuvers. An example of this
about lanes and their environment, and applications that need is parking. To do this, appropriate parking facilities must first
direct access to the individual physical elements of a map. be known, but also extensive information must be known
Each of these three groups has different requirements. for reaching a parking space. For example, which areas can
be used on the way to a parking space or which areas
A. Road Network are suitable for maneuvering without obstructing oncoming
One of the most obvious applications is routing. This is traffic. Another example are emergency situations where the
of course already a standard feature of many map formats, vehicle has to evade another vehicle or a safe position has
but HAD requires greater accuracy. It is not enough to know to be reached after a sensor failure. For this purpose, map
which roads make up a route. It must also be clear which data may not be limited to the lanes themselves, but must
lanes within the road can be used and which, for example, also contain detailed information about the surroundings of
lead to turning lanes or are reserved for other road users a road.
such as buses. Previous knowledge of the course of the
individual lanes can also help to minimize the number of C. Physical Elements
lane changes required during a journey. In addition, it must As already mentioned in Section I, localization is a typical
be clearly visible where lane changes are possible. Especially case where observable elements are important. For this
on urban roads, and complex intersections, lane changes are purpose, a map must contain elements that can be observed
often temporarily forbidden in one or both directions. Also, it by as many sensors as possible in order to be usable for
must be easy to find alternative lanes or routes if the selected vehicles with different sensor setups. At the same time a
lane is blocked. sufficient density of such elements must be offered so that
In behavior generation, a choice is made between different an exact localization is possible everywhere. Elements such
maneuvers, such as overtaking, merging or braking maneu- as markings, crash barriers and roadsides can be used for
vers in front of traffic lights and pedestrian crossings. This this, as they can be found on many roads and are also a
requires precise knowledge of right of way rules in order to basic component of corridors, so that a map offers inherently
generate behavior that is consistent with traffic regulations. consistent information for localization and motion planning.
Therefore, an exact knowledge of a chosen route is required, A major challenge for maps is however to ensure that
together with the traffic rules that apply to it. the data is up to date. The more details a map contains,
A similar task is the prediction of other road users. Under the more likely it becomes that a map needs to be updated.
normal conditions it can be assumed that they either continue Today’s map formats often do not contain the underlying
data, but only an abstract, pre-processed form. However, if The requirements for maps for HAD are constantly chang-
the environment changes, it is no longer possible to trace ing as new algorithms and processes are being developed.
which parts of the map are affected. For example, removing At the same time, the map formats must be valid for
a speed limit sign may mean that this speed limit no longer many different countries with differing traffic rules. It must
applies to the following road section. But there may also be therefore be ensured that the map format is expandable and
another sign (for example on the other side of the road) that flexible so that it can be adapted to changing requirements.
still enforces this speed limit. If only the derived information This applies in particular to the traffic rules contained therein.
(the speed limit) is stored on the map, it is no longer possible The high degree of complexity of the resulting maps
to determine whether the traffic sign was the source of the requires suitable software interfaces to ensure a uniform,
traffic rule or not. The map should therefore not only contain consistent interpretation. We have described above that the
the derived information, but also allow conclusions to be representation of the elements of a map should be separate
drawn as to where this information came from. from their interpretation. This requirement also applies to the
The same applies to lanes: Structural changes to a road software implementation of such a framework. This makes
can only be detected in time by change detection algorithms it possible to interpret the traffic rules stored in a map
if it is known what limited the lane. In a subsequent step, dynamically, depending on the road user and her abilities.
maps could also be updated automatically by deriving the In the following sections we would therefore like to present
new, correct lanes from the changed lanes that a vehicle has how we designed Lanelet2 to follow these principles.
observed. In summary, we consider the following principles to be
Besides local changes, there is another challenge to ensure indispensable for future HD map formats for HAD:
the map’s correctness over time: Physical elements need to • verifiability of all information in the map while driving
be stored and referenced in a way that is robust against by associating it to observable objects,
environmental changes such as continental drift. The obvious • coverage of all potentially passable areas, even off the
solution is using a locally fixed reference frame instead of road,
using global coordinates (like GPS). • the interactions between the individual lanes and regions
on the map must be identifiable and comprehensible. It
D. Consequences for Maps must be possible to find out between which lanes lane
changes are possible or where conflicts can arise due to
We call the detectable elements mentioned above the crossing lanes,
physical layer of a map. It consists of real, observable • information on areas used by other road users and the
elements, such as markings, curbstones, traffic lights, etc. rules applicable to them,
All other elements, such as lanes, are only associations of • separation between sources of traffic rules and their
these elements to a more abstract representation. Likewise, implication on road users,
traffic rules are associations between the source of a rule, • expandability/modularity,
such as traffic lights, and the lanes to which they apply. This • modifiability to easily perform updates.
representation allows the information on a map to be traced.
We call this layer the relational layer of a map. IV. ARCHITECTURE OF LANELET2
This results in further advantages: For traffic rules it The map framework Lanelet2 has been developed with the
becomes transparent, to which type of road user they refer. requirements of Section III in mind. It is an extension and
For this purpose, it must only be comprehensible whether generalization of the map format Liblanelet [14] developed
the element associated with a rule is also valid for this for the Bertha-Benz journey. As indicated in the previous sec-
road user. Accurate, map consistent localization can also tion, the map is divided into a physical layer, which contains
be ensured, since elements of the physical layer can be the usually real observable elements and a relational layer,
detected by definition. By matching these detections with in which the elements of the physical layer are connected to
the map, the position within it can be determined or at least lanes, areas and traffic rules. A third layer results implicitly
validated. Furthermore, the information can also be used to from the contexts and neighbourhood relationships of the
simulate and therefore validate highly automated vehicles on relational layer: the topological layer. Here, the elements of
real routes. Ideally, sensor observations can also be simulated the relational layer are combined to a network of potentially
based on this. passable regions depending on the road user and situation.
However, not all elements of a map can be associated It would be conceivable and possible to derive the elements
to observable objects, as implicit rules also occur in traffic. of the relational layer from the physical layer (as humans
At intersections, for example, there are traffic lanes whose do), but currently no methods exist to make this possible.
borders are not defined by markings. Nevertheless, human However, such a procedure is at least made possible by this
traffic participants move within an imaginary corridor, which presentation.
must therefore also be part of a map. Still, elements in a map The format is based on the format known from Liblanelet
based on such purely conventional behavioral patterns must and was designed to be representable on the XML-based
be avoided as much as possible, as they are difficult to verify OSM data format, for which editors and viewers are publicly
and behavioral patterns can change, too. available. However, we consider the actual data format of a
map to be irrelevant and interchangeable as long as it is lanelet, but also that the topological relationships with other
ensured that this format can be transferred to the internal lanelets does not change. A lanelet is defined by exactly one
representation without loss. A conversion to other formats line string as the left and exactly one as the right border. In
would therefore be easy. addition, it may have several regulatory elements expressing
We assume that all elements in the map can be described traffic rules applying to this lanelet. Lanelets can also overlap
predominantly by a projection onto a flat ground plane. Usu- or intersect. The type of the left or right border expresses
ally, this requirement is met for all elements in the vicinity whether, for example, lane changes to an adjacent lanelet are
of a road. Nevertheless, height information is important, for possible. Successive lanelets share the end points of the left
example, to determine the height profile of a road [13] or and right border. Within a lanelet, movement in the opposite
to prevent a crossing bridge from being misinterpreted as an direction can be allowed, so that the left border becomes the
intersection. Therefore, a height coordinate is also required. right (and vice versa).
For storing map data, the most important requirement is Areas are sections of the map in which undirected or
correctness. This means not only immutability when only no movement is possible. Examples include parking areas,
reading the map, but also robustness against influences like squares, green spaces or buildings. They are defined by
continental drift. Thus, a locally fixed reference system like one or more linestrings, which together form a closed outer
ETRS89 in Europe is used with lossless geographic coordi- border, and can have one or more line strings, which together
nates (latitude/longitude). When loading the map, geographic define several inner borders and thus holes within the area.
coordinates are transformed into a local, metric coordinate Similar to lanelets, these can also have regulatory elements.
system in order to be able to perform efficient calculations. Regulatory elements (short: regElem) define traffic rules,
For this working representation, the UTM (Universal Trans- such as speed limits, priority rules or traffic lights. A
verse Mercator) coordinate system offers itself. regulatory element is always referenced by one or more
lanelets or areas for which they apply. Because there are
A. Lanelet Primitives very different types of traffic rules, the exact structure of
A Lanelet2 map consists of five elements: Points and line a regulatory element can be very different. They usually
strings belonging to the physical layer and lanelets, areas reference an element that defines the rule (e.g. a traffic sign)
and regulatory elements belonging to the relational layer. and, if necessary, an element that cancels the rule (e.g. a
All elements have in common that they are identified by sign at the end of a speed zone). In addition, stop lines
a unique ID (this is useful for an efficient construction of can be referenced, for example. The attributes of a rule
the topological layer) and that attributes in the form of key- element specify what kind of rule it is. RegElems can also
value pairs can be assigned to them. Some of these attributes be dynamic, which means that a rule is only valid based
are fixed, but additional attributes can be used to enhance on a condition. Examples include digital speed indicators or
the map. With the exception of areas, all primitives were traffic rules that are only valid at certain times of the day.
also already part of Liblanelet. However, in particular, the
definition of lanelets has been modified and generalized to B. Modules
meet the aforementioned requirements. As explained in Section III-D, it is important to separate
Points are the basic element of the map. Single points the representation of the elements of a map from their
represent, for example, vertical structures such as poles. interpretation. Lanelet2 is therefore based on a consistent
Usually, however, they are part of line strings. Each point modularization of the individual tasks. This also ensures
is described by its three-dimensional position in metric the extensibility of the framework. Representation is solved
coordinates. Points are the only primitives that actually have by the core module, which contains the above-mentioned
position information. All other primitives are directly or primitives. There is the traffic rules module for interpretation,
indirectly composed of points. with which the traffic rules can be queried depending on
Linestrings are an ordered array of two or more points the road user and country. For example, it decides which
between which linear interpolation takes place. It is used to maximum speed may be driven on a lanelet and whether
represent the shape of elements in the map. Examples are a lane change between two lanelets is possible. Based on
road markings, curbs, facades, fences, etc. Linestrings can these modules, there are further modules specialized for
also be virtual, for example, if they form an implicit border of the individual applications, such as for routing, matching
a lane. Linestrings were chosen as a form of representation, and access to the individual physical elements. In addition,
because they can be used to describe any one-dimensional there are other modules that provide interfaces for common
form, if necessary through high discretizations. Compared software frameworks, such as Python, ROS [16] or for
to splines, they can be calculated efficiently and can be reading and writing to certain file formats.
used to represent sharp corners. Viable solutions exist for In summary, Lanelet2 is separated into the following
the problem of non-linear differentiability (see [15]). modules:
Lanelets define an atomic section of the map in which Core: The core module contains the basic primitives and
directed motion takes place. Examples of this are normal layers described above. In addition, geometry functionalities
lanes, but also pedestrian crossings and rails. Atomic means are included to calculate center lines, distances and overlaps,
that currently valid traffic rules do not change within a for example.
 1 Right of Way Traffic Light
Ref_Line: Refers: 2
2 4 5 a 6 3 Right of Way: B Ref_Line:
Yield: E
2

A C F
7 8 9 A C E
B D G I
10 11 12 J
13 b 14 E H B D F
15 16
17 c Speed Limit 18 K L
19 Refers:
Traffic Light
1
Routing Graphs Refers: 3
Ref_Line:
Car Emergency Vehicle
Traffic Light
Refers: 1
G H
A C F A C F Ref_Line: 3
Topological Layer
B D G B D G

E H b E H

Areas Lanelets Reg. Elements A C E

J I
Linestrings
1 2 3 4 5 6 7 8 18 19

Physical Layer
Points B D F
P1 P2 P3 P5 P6 P7 P8 P9

K L
Fig. 2: Map example for a highway road (enclosed by
guardrails) and the resulting map structure. Lanelets have
capital letters, areas lowercase letters and linestrings have
G H
numbers.
Fig. 3: Map example for an intersection and resulting routing
graph for normal vehicles. Conflicting lanelets are shown in
Traffic Rules: This serves to interpret the rules contained
red in the graph.
in the map depending on the type of road user and country. It
also determines whether lane changes are possible or whether
a certain road user is permitted to enter a lanelet. V. CASE STUDIES
Physical: This module allows direct, comfortable access In this section we would like to illustrate with some exam-
to the elements of the physical layer. ples how a lanelet map can be constructed. For this purpose
Routing: With the help of traffic rules, routing graphs can we first show the basic concept by means of some typical
be set up to determine the exact route to be driven, including scenarios and then present more complex real examples to
possible lane changes, or to predict possible routes and demonstrate the applicability on them.
points of conflict for other road users. It is also possible to
A. Scenarios
construct maneuverable zones by combining adjacent areas
and lanelets. Fig. 2 shows a Lanelet2 map example for one direction of
Matching: This module is used to assign lanelets to road a motorway, which is extended by one lane. Since this lane is
users or to determine possible positions on the map based added to the side, it can only be reached by changing lanes.
on specific observations of the sensors. The map also contains several areas: one for the emergency
lane, and two for the green strip between the road border
Projection: Contains functionality to convert global lati-
and the guardrail. The latter may seem somewhat pedantic,
tude/longitude coordinates to local, metric coordinates.
but are intended to serve as an example for areas that can
IO: The input/output module contains functions for read- be included in emergency maneuvers, for example.
ing and writing maps from various map formats, in particular Two routing graphs a shown that might result from this
the OSM format. scenario: One for a normal vehicle and one for an emergency
Validity: Module that searches and reports typical mapping vehicle. Because emergency vehicles are allowed to use the
errors in a map. emergency lane b and might do a lane switch from F to G,
ROS: A connection to the Robot Operating System (ROS) the routing graph contains more edges.
which is often used in demonstrators. Note also the division of the lanelets E and H. This is
Python: For using the above modules in Python. necessary, as with the addition of the continuous line no lane
Fig. 4: Complex intersection in the east of Karlsruhe, Ger-
many and corresponding map. Lane markings are visualized
in blue, curbs in gray, road borders in green and virtual
Fig. 5: Small roundabout in an urban area in Karlsruhe and
lines in black. Areas for pedestrians are shown in gray, green
corresponding map. In addition to the above, walls are shown
stripes in green.
in black, traffic islands as red areas and parking spaces as
Aerial image: c CNES Distribution Airbus DS, HERE.
blue areas.
Aerial image: c Stadt Karlsruhe | Liegenschaftsamt
change from G to F is possible. If E and H were one single
lanelet, this would allow the conclusion that a lane change
from the former H to C would be possible, since these are Because the intersection is controlled by traffic lights, the
now connected via D. The restriction that lanelets may only lanelets B, E and H each have a regElem that references the
have one right edge forces that such a situation cannot arise, corresponding traffic light and the corresponding stop line.
since the splitting of the lanelets D and G forces that also Because vehicles turning left from E have to yield to the
the right edge is separated, which propagates further to the vehicles going straight from B, the lanelet E and B both
lanelets E and H. have a right of way regElem that references B as the lanelet
Fig. 3 is an example of an intersection. Since the inter- with right of way and E as lanelet that has to yield. The
section does not contain road markings for some of the right of way regElems always reference lanelets right at the
turning lanes, the lanelets G, H and I partly have one entrance of the intersection where vehicles can be assigned
virtual line as edge. Furthermore, the routing graph contains unambiguously to one single lanelet. If lanelets intersect
relationships between conflicting lanelets, so that potentially without a right of way regElem, both lanelets have equal
critical situations can be determined. priority and the first of two vehicles has the right of way.
B. Real World Example ACKNOWLEDGEMENT
A more complex real-world example for an intersection
in Karlsruhe, Germany is shown in Fig. 4. The map also This publication was written in the framework of the
includes footpaths and tram tracks that cross the intersection Profilregion Mobilitätssysteme Karlsruhe, which is funded
from right to bottom. The routing graph for this section is by the Ministry of Science, Research and the Arts and
too complex and is therefore not displayed. the Ministry of Economic Affairs, Labour and Housing in
Fig. 5 shows the map of a roundabout in the east of Baden-Württemberg and as a national High Performance
Karlsruhe. It lies in the middle of roads that can be driven on Center by the Fraunhofer-Gesellschaft.
in both directions and have no dividing middle line. These R EFERENCES
roads are represented by a single lanelet each, but the routing
graph contains two nodes, one for each drivable direction. [1] J. Ziegler, P. Bender, and M. Schreiber, et al., “Making bertha drive –
an autonomous journey on a historic route,” IEEE Intell. Transp. Syst.
VI. CONCLUSION AND FUTURE WORK Mag., vol. 6, no. 2, pp. 8–20, 2014.
[2] J. M. Wille, F. Saust, and M. Maurer, “Stadtpilot: Driving au-
In the past pages we have shown what a framework for tonomously on braunschweig’s inner ring road,” in 2010 IEEE In-
HAD can look like and demonstrated that it is applicable to telligent Vehicles Symposium, pp. 506–511, June 2010.
[3] Ö. Ş. Taş, N. O. Salscheider, and F. Poggenhans, et. al, “Making bertha
practical situations. Maps for HAD have so far received little cooperate – team annieway’s entry to the 2016 grand cooperative
attention in scientific publications. However, we believe that driving challenge,” IEEE Transactions on Intelligent Transportation
this topic is too important for the future of HAD to be left Systems, vol. 19, pp. 1262–1276, April 2018.
entirely to commercial map suppliers. [4] M. Aeberhard, S. Rauch, and M. Bahram, et al., “Experience, results
and lessons learned from automated driving on germany’s highways,”
An important element that maps for highly automated IEEE Intelligent Transportation Systems Magazine, vol. 7, pp. 42–57,
driving have so far lacked is the traceability of the informa- Spring 2015.
tion stored in them. This principle requires a fundamentally [5] E. Rehder and H. Kloeden, “Goal-directed pedestrian prediction,” in
2015 IEEE International Conference on Computer Vision Workshop
different structure, so that even very abstract components of (ICCVW), pp. 139–147, Dec 2015.
maps, such as the road network, can be traced back to their [6] “Here hd live map - on the road towards autonomous driving.”
basic origins, such as the border of a traffic lane. This is the https://www.here.com/en/products-services/here-
automotive-suite/highly-automated-driving/
only way to ensure the validity and updatability of a map. here-hd-live-map, 2018. Accessed: 2018-08-20.
We also propagate the change from a map structure that [7] “Tomtom hd map with roaddna.” automotive.tomtom.com/
focuses solely on the representation of information to an automotive-solutions/automated-driving/hd-map-
application-oriented structure that provides the respective roaddna, 2018. Accessed: 2018-08-20.
[8] M. Haklay and P. Weber, “Openstreetmap: User-generated street
functionalities in HAD with exactly the information they maps,” IEEE Pervasive Computing, vol. 7, pp. 12–18, Oct 2008.
need through clear interfaces. This makes it possible to store [9] P. Neis, D. Zielstra, and A. Zipf, “The street network evolution of
significantly more complex information and relations on a crowdsourced maps: Openstreetmap in germany 20072011,” Future
Internet, vol. 4, no. 1, pp. 1–21, 2012.
map without making its simple use impossible. A good map [10] “Opendrive.” www.opendrive.org/, 2018. Accessed: 2018-04-
for HAD therefore does not only consist of a certain format, 11.
but it is a comprehensive software framework. [11] I. Shim, J. Choi, and S. Shin, et al., “An autonomous driving system
Since the applications that depend on a highly accurate for unknown environments using a unified map,” IEEE Transactions
on Intelligent Transportation Systems, vol. 16, pp. 1999–2013, Aug
map are still the object of research, we need a discourse 2015.
on how such a framework should be structured and which [12] F. Dierkes, M. Raaijmakers, and M. T. Schmidt, et al., “Towards a
functionalities it must support. For this the map framework multi-hypothesis road representation for automated driving,” in 2015
IEEE 18th International Conference on Intelligent Transportation
Lanelet2 should represent a starting point and support these Systems, pp. 2497–2504, Sept 2015.
developments through its expandability and flexibility. [13] A. Schindler, Vehicle Self-Localization Using High-Precision Digital
We will make the map framework Lanelet2 available Maps. PhD thesis, University of Passau, Faculty of Computer Science
and Mathematics, 2013.
under the BSD license on Github1 and will continue to add
[14] P. Bender, J. Ziegler, and C. Stiller, “Lanelets: Efficient map repre-
increasingly more functionality. We hope for feedback and sentation for autonomous driving,” in 2014 IEEE Intelligent Vehicles
contributions. Symposium Proceedings, pp. 420–425, June 2014.
Lanelet2 -based maps for large parts of Karlsruhe, Ger- [15] E. Héry, S. Masi, P. Xu, and P. Bonnifait, “Map-based curvilinear
coordinates for autonomous vehicles,” in 2017 IEEE 20th International
many, are currently in preparation and will also be published. Conference on Intelligent Transportation Systems (ITSC), pp. 1–7, Oct
2017.
1 https://github.com/fzi-forschungszentrum- [16] M. Quigley, K. Conley, and B. P. Gerkey, et al., “Ros: an open-source
informatik/lanelet2 robot operating system,” 2009.