Disabled, but at What Cost?
An Examination of Wheelchair Routing Algorithms
University of Bremen
University of Bremen
Platforms like Google Maps or Bing Maps are used by a
large number of users to find the shortest path to their
destinations. While these services mainly focus on
supporting drivers and pedestrians, first services exist that
support wheelchair users. Routing algorithms for
wheelchair users try to avoid obstacles like stairs or
bollards and optimize on criteria like surface properties and
slope of the route. In this study, we undertake the first
controlled examination of wheelchair routing approaches.
By analyzing three routing platforms, including two
wheelchair routing algorithms and three pedestrian routing
algorithms, across fifteen major cities in Germany, our
results highlight that the routes for wheelchair users are
significantly longer and partially also more complex than
those for pedestrians. In addition, we show that today’s
pedestrian routing algorithms also output very diverse
Accessibility; HCI; Disability; Wheelchair Users;
Pedestrian Navigation; Routing; City Planning.
ACM Classification Keywords
H.5.m. Information interfaces and presentation (e.g., HCI):
INTRODUCTION & RELATED WORK
Besides a large number of unreported cases, more than 65
million people in the world need a wheelchair on a daily
basis . In Germany, as per the Federal Statistical Office
[17,], 1.5 million people need a wheelchair every day. This
is 2.7 percent of the whole population of Germany.
Various mobile technologies, apps and web platforms can
support wheelchair users to master their daily lives. For
example, Wheelmap (http://wheelmap.org) provides crowd-
sourced accessibility information for buildings, points of
interests (POIs) and restaurants based on OpenStreetMap
The Wheelmap service uses a traffic light metaphor to
indicate if a building is easily accessible or not. Google also
recently announced their decision to crowd-source
accessibility information for buildings  to be integrated
into Google Maps.
Besides accessibility information, obstacles along a route
are crucial for wheelchair users. While some obstacles are
obvious (e.g. steps or pillars), some others are often only
considered by disabled people (e.g. slope of sidewalks or
different surface properties like sand or cobblestone, which
are difficult to overcome in a wheelchair). Various
approaches analyzed the accessibility of sidewalks with
sensors and technical equipment such as depth cameras or
acceleration sensors [3,4,5,6,9,12,16,19,20].
MobileHCI '18, September 3–6, 2018, Barcelona, Spain
© 2018 Copyright is held by the owner/author(s). Publication rights
licensed to ACM.
ACM ISBN 978-1-4503-5898-9/18/09…$15.00
Figure 1: Routes for pedestrians (purple, green, orange) and
wheelchair users (blue, yellow) calculated by different routing
platforms for an origin-destination pair in Frankfurt,
Germany. Base map © OSM.
Besides these (semi-) automated approaches, Hara et al. 
explored the use of Amazon Mechanical Turk (AMT)
workers for classifying the accessibility of sidewalks for
wheelchair users. They compared the differences in the
annotation of the AMT workers and the experts (wheelchair
users) and found that even untrained personnel could easily
identify problems in the image of sidewalks.
Today, first routing services exist that help to guide
wheelchair users from an origin to a destination [10, 11,
17]. These services take numerous variables into account to
provide routes that are easy for wheelchair users to follow.
Most commonly, they avoid obstacles and barriers such as
stairs or bollards, utilize the surface properties, the slope,
and the height of the pavement edges to calculate routes
particularly suited for wheelchair users. Karimi et al. 
provide a good overview of requirements and components
needed for a routing service that can assist disabled people.
In this study, we undertake the first controlled examination
of these wheelchair routing services and report on the
following three contributions:
• We found that significant differences exist
between routes for pedestrian and wheelchair users
within same areas. For example, in Frankfurt,
Germany, routes (between 1.5 and 2.0 km straight-
line distance between origin and destination) for
wheelchair users were on average nearly double
the length of those for pedestrians and were also
more complex (e.g. in terms of # of turns).
• Furthermore, a notable difference exists between
cities. While we observed a large difference in
Frankfurt, the difference between routes for
pedestrian and wheelchair users differed by just 1
% in Hamburg (again for routes between 1.5 and
2.0 km straight-line distance between origin and
destination and routes calculated with
OpenRouteService). Therefore, our results provide
a novel way to benchmark cities with regard to the
• We also noticed that pedestrian routing algorithms
calculate very different routes in terms of their
complexity even though they provide routes of
similar length. For example, Google Maps tries to
minimize the number of turns to reduce the
complexity of the calculated routes even so this
In our study, we investigated three different routing
platforms, namely Google Maps
(https://www.openrouteservice.org) and Routino
(https://www.routino.org). While Google Maps uses
proprietary data, OpenRouteService (ORS) and Routino
both rely on data from OpenStreetMap (OSM)
(https://www.openstreetmap.de/) for their calculations and
both offer dedicated routing algorithms for wheelchair
users. Google Maps provide routes for different modes of
locomotion but not for wheelchair users.
The ORS algorithms use default settings to calculate the
routes for pedestrians (ORS_Ped) and wheelchair users
(ORS_Wheel). ORS_Ped assumes a speed of 6 km/h and
ORS_Wheel uses 8 km/h. And while in the ORS_Wheel the
whole route has to be paved, the surface is not taken into
account in ORS_Ped.
Similarly, the ORS_Wheel algorithm avoids steps, while
ORS_Ped does not take this feature into consideration.
Additionally, ORS_Wheel takes the incline of route
segments (max. 6 % incline) and the height of sloped curbs
(max. height of 6 cm) into account. A detailed description
of the ORS routing algorithms can be found here .
In contrast to the ORS algorithms, the Routino algorithm
for pedestrians (Rou_Ped) and Routino for wheelchair users
(Rou_Wheel) both assume the speed of 4 km/h for their
calculations. While in the Rou_Wheel 90 % of the route has
to be paved, only 50 % needs to be paved in Rou_Ped.
Similar to ORS_Wheel, Rou_Wheel avoids steps. The
detailed description of the Routino algorithms can be found
Given the fact that Google uses proprietary algorithm and
data, we cannot provide detailed information on the
implementation of the Google_Ped algorithm. The only
available information is that the average speed is assumed
to be 5 km/h to calculate the time of travel.
Table 1 provides a comparison of the main criteria used by
the algorithms analyzed in this paper. Figure 1 shows an
example of different routes calculated by the five
algorithms used in our study for one origin-destination pair
in Frankfurt, Germany.
To evaluate the 5 routing algorithms (Google_Ped,
ORS_Ped, ORS_Wheel, Rou_Ped, Rou_Wheel), we
extended a framework developed by Johnson et al.  and
integrated all five routing algorithms. Johnson et al. used
their framework to investigate externalities that arise with
three common approaches to the fastest path option
(“beauty”, “safety” and “simplicity”) for car-based routing
algorithms across four cities around the globe, namely
London, Manila, San Francisco and New York. They used
Table 1: Overview of the main criteria of the algorithms.
around 1000 origin-destination pairs for each city to
compare the effects of the different routing options. For the
origin and destination pairs, they used information about
the most common pathways of taxi companies and
generated random points in the city.
In order to compare the outcome provided by the different
routing platforms, we had to identify a set of representative
origin-destination pairs for all fifteen cities. As no public
data was available for wheelchair routes in those cities, we
calculated origin-destination pairs between POIs and public
restrooms for disabled people. As confirmed by wheelchair
users, these pairs describe a set of typical routes.
We selected the 15 biggest German cities namely, Berlin
(BER), Bremen (HB), Cologne (COL), Dortmund (DOR),
Dresden (DRE), Düsseldorf (DUE), Erfurt (ERF), Essen
(ESS), Frankfurt (FRA), Hamburg (HH), Hanover (HAN),
Leipzig (LEI), Munich (MUN), Nuremberg (NUR) and
Stuttgart (STU). One of the cities is the hometown of one of
the co-authors, who also uses a wheelchair, which offers the
possibility to use his knowledge to compare the quality of
the calculated routes. To generate the origin-destination
pairs we first gathered data on the location of public
restrooms for disabled people from www.myhandicap.de.
Then, we used the Google Maps API to identify POIs in a
radius of 2 km around these public restrooms. The POIs,
e.g. parks or restaurants, are places from where wheelchair
users would usually need to drive to a restroom. The
Google Maps API provides up to 200 POIs around each
location. Using this process, we generated 267.421 origin-
destination pairs for all 15 cities. We extracted 2715 pairs
for every city with the same average straight-line distance
and same variance to be comparable across all 15 cities. For
further analysis, we grouped these pairs into four classes of
routes that have a straight-line distance of 0.0-0.5 km, 0.5-
1.0 km, 1.0-1.5 km and 1.5-2.0 km.
In order to derive evidence on the statistical differences
between the routing alternatives per city and the summary
of them (in terms of length and number of turns), we used a
one-way ANOVA with Bonferroni correction. The
significance threshold was set to p<0.05.
Figure 2a summarizes the average length of the routes of all
five routing algorithms across the fifteen cities (40.725
origin-destination pairs for all cities, 2715 routes per city).
There is a significant difference between all pairs of routing
algorithms with the exception of the pair ORS_Ped and
Rou_Ped, which is somehow expected as both pedestrian
routing algorithms are operating on OSM data. We further
calculated the effect size of each pair with a significant
difference to analyze the impact of it. The results of this
analysis showed that the effect sizes were negligible.
Figure 2b shows the average route length grouped into
classes by the straight-line distance between origin and
destination. The chart illustrates that the difference of the
length between the pedestrian and the wheelchair
algorithms grow with the length of the straight-line
Figure 2: Average route length a) across all cities b) grouped in distance bins and c) in Frankfurt and d) Hanover.
For the classes 0-0.5 km, 0.5-1.0 km and 1.0-1.5 km there
are no significant differences, but for the class 1.5-2.0 km
all differences between routing algorithms are significant
(again with the exception of the pair ORS_Ped and
Rou_Ped). To summarize, Figures 2a and 2b show that the
routes for wheelchair users are indeed longer than those for
pedestrians. Moreover, this difference increases as the
straight-line distance between origin and destination point
grows. Google_Ped generates the shorter routes compared
to all other 4 routing algorithms. Routino_Wheel generates
longer route lengths than the pedestrian algorithms in all
We observe a continuous increase in the gap between the
average wheelchair and pedestrian route lengths, with the
rise of the straight-line measure between origin and
destination. We also found that Google_Ped is a good
representative of the other pedestrian algorithms, and for
straight-line distances (average of all cities) between 0.0-
0.5 km, the difference in route length of Google_Ped and
Rou_Wheel is approximately 50m. For the next class, this
difference is 120 m and for 1.0-1.5 km it is 250m. For the
last class between 1.5 km and 2.0 km the difference
increases to 300m.
Wheelchair vs. Pedestrian Routing across Cities
We expected that all three pedestrian routing algorithms
(Google_Ped, ORS_Ped, and Rou_Ped) generate shorter
routes as both algorithms for wheelchair users (ORS_Wheel
and Rou_Wheel). Therefore, we decided not only to
compare the groups of pedestrian and wheelchair
algorithms in general, but also between cities. We further
analyzed the differences between the five algorithms across
the 15 cities.
As can be seen in figure 2c in FRA, for the straight-line
distance class of 1.5-2.0 km, there is a significant average
difference of 1,5 km between Google_Ped and Rou_Wheel
as well as 1,6 km between Google_Ped and ORS_Wheel.
As can be seen in figure 2d in HAN, the differences are
around 100 m on average between Google_Ped and
ORS_Wheel and 120 m between Google_Ped and
Rou_Wheel, but still significant. Frankfurt was a
representative city for a rather high difference between the
algorithms, whereas Hanover was an example for a city
with rather small differences between the five routing
When comparing other pairs of algorithms, we found
similar trends. For example, while the average difference of
ORS_Ped and ORS_Wheel for 1.5-2.0 km origin-
destination pairs is around 100 m in HAN, it is around 2.0
km in FRA. We observe similar patterns between Rou_Ped
and Rou_Wheel. In the group of the fifteen cities, FRA is
special because, the route length of the wheelchair
algorithms is much longer than in the other cities.
Finally, we examined the complexity of the routes that were
calculated by the five routing algorithms. We analyzed if
there is a significant difference between the pedestrian and
wheelchair algorithms in terms of the overall number of
turns for each route.
Figure 3 shows the averages number of turns for the cities
HAM, LEI, DOR and FRA. Here it can be seen that the
complexity is not always higher for wheelchair users than
for pedestrians. For example, in DOR the ORS_Ped
algorithm is more complex than ORS_Wheel for routes
between 1.0-1.5 km, but for routes between 1.5-2.0 km the
ORS_Wheel algorithm indicates a higher number of turns.
The Rou_Ped and the Rou_Wheel perform in a similar way
in FRA for the equivalent route lengths. Instead in HAM
the higher complexity does not switch between the
pedestrian and wheelchair routes, but it is different between
the providers. Here the ORS algorithms show always that
the routes for wheelchair users are more complex, whereas
the Rou algorithms indicate that the pedestrians have more
turns on their routes.
We inferred that the difference in the number of turns
depends strongly on the provider (Google, ORS, Rou). In
comparison to Google_Ped, the ORS_Ped algorithm require
more turns for the 1.5-2.0 km class in all cities; HH (1.9),
LEI (14.2), DOR (14.1) and FRA (13.5). The Rou_Ped
seems to generate the routes not optimizing for a low
number of turns and complexity. It includes twice as many
turns as the ORS_Ped and many times more turns than
Google_Ped. For the straight-line distance section of 1.5 km
and 2.0 km the difference in the number of turns between
Google_Ped and Rou_Ped is in HH 46.5, in LEI 40.9, in
DOR 39.3 and in FRA 37.8.
Wheelchair Routing Differences
As can be seen in figure 4, the fifteen analyzed cities are
slightly different in terms of wheelchair accessibility. For
every of the 15 cities the figure shows the averaged route
length of all routes generated for wheelchair users
(ORS_Wheel and Rou_Wheel). Thus, it can be seen in
which cities the route lengths are similar and in which one
of the algorithms perform different. On the other hand, the
average route lengths of ORS_Wheel are sometimes similar
to those of Rou_Wheel but are often shorter even though
both use the same data e.g about the surface.
If we look at the pedestrian algorithms (Google_Ped,
ORS_Ped, and Rou_Ped), it is obvious that the average
length of the routes is very similar. Only for higher straight-
line distances in FRA, Rou_Ped generates longer routes
than the other pedestrian algorithm. For the 1.5-2.0 km
class, the average route length is approximately 0.15 km
longer. Although the average route lengths of the pedestrian
algorithms are very similar, the complexity and therefor the
number of turns of the routes are different (Figure 3). It can
thus be concluded that the route length is similar, but the
way is different.
The results of our analysis show that the accessibility of
cities can differ significantly.
Societal Impact of Wheelchair Routing
The results for FRA illustrate a weak implementation of
accessibility standards, which has a direct impact on lives
of many wheelchair users. This is accentuated by the long
routes generated by ORS_Wheel and Rou_Wheel. The
difference between Rou_Ped and Rou_Wheel includes
obstacles that wheelchair users have to avoid. A detailed
survey of such obstacles could be useful not only to
wheelchair users but also other stakeholders.
Complexity of Routing Algorithms
Trying to reach a destination from a given origin point can
be very difficult if a turn is missed. Figure 3 shows that
Google_Ped is trying to reduce the complexity of the route
by minimizing the number of turns to prevent wrong turns.
Compared to Google_Ped the algorithms of Routino
(Rou_Ped and Rou_Wheel) produce more complex routes
and therefore have more decision points, where users can
make mistakes. In addition to that, the ORS algorithms
(ORS_Ped and ORS_Wheel) generate the double number of
turns as the Google algorithm. Overall, the complexity of a
route is more depending on the provider than the modality
(pedestrians and wheelchair users).
To generate origin destination pairs, we used only randomly
generated pairs (between public restrooms for disabled
people and POIs). It would be interesting to include more
realistic data into our framework, but no such data exists
that captures typical routes of wheelchair users across
multiple cities (similar to the taxi data used by Johnson et
al. ) according to the Federal Statistical Office of
Germany as well as the building authorities and
departments of town planning of the 10 largest cities in
Germany. This data would not only be useful to be plugged
in our framework, but also for more general planning
The fifteen cities examined in our study are scattered over
Germany, hence their geographical topology can differ due
to various causes. However, we could still inspect
differences in routing algorithms and analyze the results in
respect to the wheelchair accessibility issue.
In general, the more spatial data and attributes are available
to calculate the route, the better the algorithms of the
routing platforms can perform. The data to be used by the
pedestrian routing algorithms is mostly complete and
requires less attributes information. However, this is not the
case with the algorithms for wheelchair users that have to
use incomplete data for different parts of the city collected
mostly by volunteers. Therefore, missing or incomplete
information can also lead to failures in wheelchair
navigation, producing longer and/or inaccessible routes.
Figure 3: Average number of turns across the five routing algorithms for four cities grouped by straight-line distance.
By analyzing three pedestrian routing platforms, including
two wheelchair routing algorithms, across fifteen major
cities in Germany we show that the routes for wheelchair
users are significantly longer and partially also more
complex than those for pedestrians. As this could be due to
missing attribute information the wheelchair routing
algorithms rely now, we manually investigated those cases
in the city of Bremen. We found out that, even so attribute
information is missing, many barriers still exist that could
be removed by decision makers to minimize route lengths
for wheelchair users. In addition, we as technologist can
help to collect missing attribute information to improve the
route generation of wheelchair routing algorithms.
Automatic and semi-automatic approaches, similar to the
ones proposed by [2,9,16,20], could be used to fill this gap.
We would like to thank Ankit Kariryaa and Reuben
Kirkham for their valuable comments and input on this
paper. Furthermore, we thank Isaac Johnson for the help
adapting his framework and Brent Hecht for his support in
the beginning of this study. This work is also partially
supported by the project “InWi – Inklusion in der
Wissenschaft” and by the Volkswagen Stiftung through a
1. Sasha Blair-Goldensohn. 2017. A better world for
wheels on Google Maps.
2. Kotaro Hara, Vicki Le, and Jon Froehlich. 2013.
Combining crowdsourcing and google street view to
identify street-level accessibility problems. In
Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems (CHI '13). ACM, New
York, NY, USA, 631-640. DOI:
3. Catherine Holloway & Nick Tyler (2013) A micro-
level approach to measuring the accessibility of
footways for wheelchair users using the Capability
Model, Transportation Planning and Technology, 36:7,
636-649, DOI: 10.1080/03081060.2013.845434
4. Yusuke Iwasawa, Koya Nagamine, Yutaka Matsuo,
and Ikuko Eguchi Yairi. 2015. Road Sensing: Personal
Sensing and Machine Learning for Development of
Large Scale Accessibility Map. In Proceedings of the
17th International ACM SIGACCESS Conference on
Computers & Accessibility (ASSETS '15). ACM, New
York, NY, USA, 335-336. DOI:
5. Yusuke Iwasawa, Kouya Nagamine, Ikuko Eguchi
Yairi, Yutaka Matsuo, Toward an Automatic Road
Accessibility Information Collecting and Sharing
Based on Human Behavior Sensing Technologies of
Wheelchair Users, In Procedia Computer Science,
Volume 63, 2015, Pages 74-81, ISSN 1877-0509,
6. Nor Haslina Ja’afar, Asiah Abdul Rahim, Nur Amirah
Abd. Samad & Che Raiskandar Che Rahim. 2017.
Sidewalk accessibility at Meleka’s traditional street for
people with disabilities. In Planning Malaysia: Journal
Figure 4: Average length of routes for wheelchair users.
of the Malaysian Institute of Planners. Volume 15
Volume 1 (2017), 389 – 396. Malaysian Institute of
Planners. Petaling Jaya, Selangor Darul Ehsan. DOI:
7. I. Johnson, J. Henderson, C. Perry, J. Schöning, and B.
Hecht. 2017. Beautiful…but at What Cost?: An
Examination of Externalities in Geographic Vehicle
Routing. Proc. ACM Interact. Mob. Wearable
Ubiquitous Technol. 1, 2, Article 15 (June 2017), 21
pages. DOI: https://doi.org/10.1145/3090080
8. Hassan A. Karimi, Lei Zhang & Jessica G. Benner
(2014) Personalized accessibility map (PAM): a novel
assisted wayfinding approach for people with
disabilities, Annals of GIS, 20:2, 99-108, DOI:
9. Reuben Kirkham, Romeo Ebassa, Kyle Montague,
Kellie Morrissey, Vasilis Vlachokyriakos, Sebastian
Weise, and Patrick Olivier. 2017. WheelieMap: an
exploratory system for qualitative reports of
inaccessibility in the built environment. In Proceedings
of the 19th International Conference on Human-
Computer Interaction with Mobile Devices and
Services (MobileHCI '17). ACM, New York, NY,
USA, Article 38, 12 pages. DOI:
10. N. P. Kozievitch, L. D. A. Almeida, R. Dutra da Silva
and R. Minetto, "An alternative and smarter route
planner for wheelchair users: Exploring open data,"
2016 5th International Conference on Smart Cities and
Green ICT Systems (SMARTGREENS), Rome, 2016,
11. Kirill Kulakov, Anton Shabaev, and Irina Shabalina.
2015. The route planning services approach for people
with disability. In Proceedings of the 17th Conference
of Open Innovations Association FRUCT (FRUCT'17),
Sergey Balandin, Tatiana Tyutina, and Ulia Trifonova
(Eds.). FRUCT Oy, Helsinki, Finland, Finland, Pages
14, 7 pages. DOI:
12. Q. Mourcou, A. Fleury, P. Dupuy, B. Diot, C. Franco
and N. Vuillerme, "Wegoto: A Smartphone-based
approach to assess and improve accessibility for
wheelchair users," 2013 35th Annual International
Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), Osaka, 2013, pp. 1194-1197.
13. Newdisability. Wheelchair Statistics: How Many
Wheelchair Users Are There?
14. ORS. OpenRouteService-docs
15. Routino. Routino: Algorithm.
16. Nina Runge, Pavel Samsonov, Donald Degraen, and
Johannes Schöning. 2016. No more Autobahn!: Scenic
Route Generation Using Googles Street View.
In Proceedings of the 21st International Conference on
Intelligent User Interfaces (IUI '16). ACM, New York,
NY, USA, 147-151. DOI:
17. A. S. Sobeck, H. J. Miller. 2006. U-Access: a web-
baesd system for routing pedestrians of differing
abilities. Journal of Geographical Systems 8, 2: 269-
18. Statistisches Bundesamt -
19. Thomas Stockx, Brent Hecht, and Johannes Schöning.
2014. SubwayPS: towards smartphone positioning in
underground public transportation systems.
In Proceedings of the 22nd ACM SIGSPATIAL
International Conference on Advances in Geographic
Information Systems (SIGSPATIAL '14). ACM, New
York, NY, USA, 93-102. DOI:
20. Alexander Zipf, Amin Mobasheri, Adam Rousell and
Stefan Hahmann. 2016. Crowdsourcing for individual
needs – the case of routing and navigation for mobility-
impaired persons. In: European Handbook of
Crowdsourced Geographic Information. Pp.325–337.
London: Ubiquity Press.