a

r
X
i
v
:
1
0
1
1
.
4
4
6
5
v
1


[
c
s
.
D
S
]


1
9

N
o
v

2
0
1
0
Compressed Transmission of Route Descriptions

Gernot Veit Batz, Robert Geisberger, Dennis Luxen, and Peter Sanders
{batz,geisberger,luxen,sanders}@kit.edu
November 22, 2010
Abstract
We present two methods to compress the description of a route in a road network, i.e., of a path in a
directed graph. The rst method represents a path by a sequence of via edges. The subpaths between
the via edges have to be unique shortest paths. Instead of via edges also via nodes can be used, though
this requires some simple preprocessing. The second method uses contraction hierarchies to replace
subpaths of the original path by shortcuts. The two methods can be combined with each other. Also,
we propose the application to mobile server based routing: We compute the route on a server which
has access to the latest information about congestions for example. Then we transmit the computed
route to the car using some mobile radio communication. There, we apply the compression to save
costs and transmission time. If the compression works well, we can transmit routes even when the
bandwidth is low. Although we have not evaluated our ideas with realistic data yet, they are quite
promising.
Keywords. server-based mobile route planning, GPS-based car navigation, hybrid scenario, data
compression, mobile communication, cellular phone network, graph algorithms, shortest path.
1 Introduction
Today GPS-based car navigation is quite common. There, routes can be computed either by a mobile
device located in the car or by a server system located in a computing center. The latter can be useful to
take temporary congestions or the latest changes of the road network or of the POI data into account. But
then, a description of the route has to be transmitted to the car using some mobile radio communication
like the cellular phone network for example. Of course, good compression rates are desireable as they
help to save costs and transmission time. Especially on countryside highways bandwidths can be low.
In this paper we describe two approaches for compressing a route: In the rst approach, we represent
a route by giving only a few road segments which we call via edges. Equivalently, one can also use via
nodes, though this requires some simple preprocessing of the road network. Via edges and via nodes
require that the mobile device located in the car has passably up-to-date map data and the capability to
compute optimal routes with respect to this data. In the second approach, contraction hierarchies [10]
are used to encode a route using shortcuts. In this case the mobile device in the car has to support
contraction hierarchies too. The two approaches can also be combined.
2 Preliminaries
In route planning a road network is usually represented by a weighted directed graph G= (V, E, c) where
the nodes represent junctions and the edges represent road segments; c : E C assigns a weight to every

Partially supported by DFG grant SA 933/5-1

1
edge in E. Here, we only consider the simplest case of time-independent edge weights, where C =R
>0
holds. But of course, our method can be applied to more general scenarios, where the server performs
time-dependent route planning [2, 3] for example.
For C = R
>0
optimal routes can be computed using Dijkstras algorithm. Alternatively, one of the
numerous recent and much more efcient methods can be used [5]. Routes are actually paths in G. Here,
we formalize a path as a sequence (u
1
, u
2
), (u
2
, u
3
), . . . , (u
n
, u
n+1
) with (u
i
, u
i+1
) E for 1 i n. Of
course, two paths P := (u
1
, u
2
), . . . , (u
n
, u
n+1
) and P

:= (v
1
, v
2
), . . . , (v
m
, v
m+1
) with u
n+1
= v
1
can
be concatenated we write PP

for the concatenated path. Shortest paths are not necessary unique
in G, i.e., there may be more than one shortest path between the respective start and destination nodes.
But if there is a unique shortest path from a node u to a node v, then we denote it by SP(u, v).
3 Compression of Paths Using Via Edges
3.1 Representing a Path by Via Edges
Given a path P = (u
1
, u
2
), (u
2
, u
3
), . . . , (u
n
, u
n+1
) in G, which is not necessarily a shortest path. Let
Q = (u
i
1
, u
i
1
+1
), (u
i
2
, u
i
2
+1
), . . . , (u
i
k
, u
i
k
+1
) be a subsequence of P with the property that the shortest
paths from u
1
to u
i
1
, from u
i
k
+1
to u
n
, and from u
i
j
+1
to u
i
j+1
are unique for all 1 j < k. If
P = SP(u
1
, u
i
1
) (u
i
1
, u
i
1
+1
) SP(u
i
1
+1
, u
i
2
) (u
i
2
, u
i
2
+1
) SP(u
i
k
+1
, u
n+1
)
holds, then P is completely determined by Q. We call Q a representation of P by via edges in this case.
There, (u
i
1
, u
i
1
+1
), . . . , (u
i
k
, u
i
k
+1
) are the via edges.
Corollary 1. If Q is a representation of a path P by via edges, then |Q| |P| holds.
Corollary 1 means that Q is a compressed representation of P. At best, P itself is a unique shortest
path in G and we can represent P by the empty sequence Q = if the start and the destination of P are
known. Given a representation Q = (u
i
1
, u
i
1
+1
), . . . , (u
i
k
, u
i
k
+1
) of a path P by via edges the path P can
be reconstructed very easily: We only have to perform a one-to-one shortest path query between all pairs
of nodes u
i
j
+1
and u
i
j+1
for j {1, . . . , k 1}, between the start node and u
i
1
, as well as between u
i
k
+1
and the destination node.
3.2 Representing Paths by Via Nodes
Analogous to via edges, several paths can also be represented by via nodes. Then, a path P is represented
as a sequence of nodes lying on P. Again, the subpaths between the via nodes have to be unique shortest
paths. However, this fails in some cases. Consider, for example, a path that consists of one single
edge and that is not a unique shortest path. With via nodes we could not represent this path. However,
this problem can solved if we apply a simple preprocessing to G: We just have to replace all edges
in (u, v) E, that are not unique shortest paths, by a path (u, w
new
), (w
new
, v), where w
new
has to be
newly added to V. Also, we have to set c((u, w
new
)) := c((w
new
, v)) := c((u, v))/2. Having applied this
preprocessing all shortest paths in G can be represented using via nodes. Although the results in the
work are described in terms of via edges, they can also be used with via nodes if this preprocessing
is applied to the underlying graph before.
3.3 Computing a Minimal Representation by Via Edges
A representation Q of a path P by via edges is minimal if for all other such representations Q

holds
|Q| |Q

|. Algorithm 1 yields such a minimal representation of a path P by via edges. It performs

O(|P|) one-to-one shortest path queries in G.
2
Algorithm 1: Computes a minimal representation of a path by via edges.
1 function viaEdges((u
1
, u
2
), . . . , (u
n
, u
n+1
) : path) : sequence of E
2 i := 1 : N
3 Q := : sequence of E
4 for j := 1 to n do
5 if (u
i
, u
i+1
), . . . , (u
j
, u
j+1
) is unique shortest path from u
i
to u
j+1
in G then continue
6 Q.append((u
j
, u
j+1
))
7 i := j +1
8 return Q
Algorithm 2: Computes the same as Algorithm 1 but runs faster when few via edges are needed.
1 function maxPrexSP( j, k : N) : N
2 (, r) := ( j 1, k)
3 while true do
4 if +1 = r then return
5 m := ( +r)/2
6 if (u
j
, u
j+1
), . . . , (u
m
, u
m+1
) is unique shortest path from u
j
to u
m+1
in G then := m
7 else r := m
8 function viaEdges((u
1
, u
2
), . . . , (u
n
, u
n+1
) : path) : sequence of E
9 i := 1 : N
10 Q := : sequence of E
11 while i n do
12 p := maxPrexSP(i, n)
13 if p < n then Q.append((u
p+1
, u
p+2
))
14 i := p+2
15 return Q
Algorithm 2 does in principle the same as Algorithm 1 because both algorithms compute the maxi-
mum p {i 1, . . . , n} such that the subpath (u
i
, u
i+1
), . . . , (u
p
, u
p+1
) of P is a unique shortest path in
G (there, p = i 1 means that no such subpath of P starting at u
i
exists). Then, both algorithms append
the edge (u
p+1
, u
p+2
) to the result sequence Q. In case of Algorithm 2 we delegate the computation
of p to the auxiliary function maxPrexSP (cf. Lines 1 to 7) which works similar to binary search and
performs only O(log|P|) one-to-one shortest path queries hence. As a consequence Algorithm 2 com-
putes the same result as Algorithm 1 but needs only O(|Q| log|P|) one-to-one shortest path queries for
Q being the resulting representation by via edges. This means that Algorithm 2 should run faster than
Algorithm 1 when Q contains only few edges. If we replace the function maxPrexSP from Algorithm 2
by another variant (cf. Algorithm 3), we can even reduce the number of one-to-one shortest path queries
to O(|Q| log(|P|/|Q|)).
4 Applying Contraction Hierarchies
A contraction hierarchy (CH) [10, 7] is a hierarchical representation of a road network, where the nodes
are ordered by importance. More important nodes are on a higher level of the CH than less important
nodes. The CH is obtained from the original road network G in a preprocessing step. It contains shortcut
3
Algorithm 3: Reduces the asymptotic running time of Algorithm 2 to O(|Q| log(|P|/|Q|)).
1 function maxPrexSP( j, k : N) : N
2 h := 0
3 while true do
4 if (u
j
, u
j+1
), . . . , (u
j+h1
, u
j+h
) is unique shortest path from u
j
to u
j+h
in G then h := 2h
5 else break
6 (, r) := ( j +h/2 1, k)
7 while true do
8 if +1 = r then return
9 m := ( +r)/2
10 if (u
j
, u
j+1
), . . . , (u
m
, u
m+1
) is unique shortest path from u
j
to u
m+1
in G then := m
11 else r := m
edges that leap valleys, i.e., subpaths that go through lower levels. More precisely, every shortcut edge
represents a path of the form (v, u), (u, w) such that the level of u is lower than the level of v and w.
Note, that (v, u) and (u, w) can be shortcuts themselves.
4.1 Compressing a Path Using Contraction Hierarchies
Originally, CHs are used for the very fast computation of shortest paths. Here, we also use them
to represent paths, that are not necessary shortest ones, in a space efcient way: Given a path P =
(u
1
, u
2
), (u
2
, u
3
), . . . , (u
n
, u
n+1
) and a CH we compress P by replacing subpaths of P with shortcuts
present in the CH. To do so, we contract the path P as far as possible. Algorithm 4 shows this method
in more detail. Note, that even if a shortcut (v, w) is present in the CH, it is not necessarily the case
that this shortcut represents the path (v, u
i
j
), (u
i
j
, w). This is because parallel edges arise during the
construction of the CH and parallel edges are thrown away during preprocessing.
Applying Algorithm 4 to a path P we get a path P

which contains shortcuts. If we repeatedly replace

all shortcuts in P

by the path they represent until no shortcuts are left, then we surely end up with the
original path P. Hence, P is completely determined by P

and we call P

a representation of P using a
CH. Note, that P

as generated by Algorithm 4 is minimal with respect to this property. More precisely,

for every other path P

, that contains shortcuts from the given CH and that completely determines P, we
have |P

| |P

|. According to the following corollary such a representation is also a way to compress a

path:
Corollary 2. Let P be a path in G. Let P

be the result of Algorithm 4 when applied to P with some CH

given. Then |P

| |P| holds.
The sorting of the nodes (cf. Line 3) can be done in O(|P| log|P|) time. Let d be the maximum
number of incoming or outgoing edges of a node in the CH. Then, the Lines 4 to 6 take O(d|P|) time.
That means, Algorithm 2 needs O(d|P| +|P| log|P|) time in total.
4.2 Combining Via Edges and Contraction Hierarchies
The representation by via edges and the representation using a CH can be combined quite easily. First,
apply Algorithm 4 to the given path P and then compute a representation Q

of the resulting path P

by via edges. The unpacking of Q

is straightforward then: Just reconstruct P

from Q

as described in
Section 3.3 and then recursively unpack all shortcuts in P

completely.
4
Algorithm 4: Replaces subpaths of a path by shortcuts of a CH.
1 function compressWithCH((u
1
, u
2
), . . . , (u
n
, u
n+1
) : path) : path
2 R := (u
1
, u
2
), . . . , (u
n
, u
n+1
) : path
3 sort u
2
, . . . , u
n
according to increasing level yielding u
i
1
, . . . , u
i
n1
4 for j := 1 to n1 do
5 if a shortcut (v, w) representing the subpath (v, u
i
j
), (u
i
j
, w) of R is present in the CH
then
6 replace (v, u
i
j
), (u
i
j
, w) by (v, w) in R
7 return R
Note, that |P

| |P| holds. We even expect that in many cases we have |P

| |P|. But then, we have

to perform much less one-to-one shortest path queries than if we represent P by via edges directly. But
one-to-one shortest path queries are one of the expensive operations in this setup. So, the combination
of both methods may not only provide better compression but also better running times than if via edges
were used alone.
5 Application
5.1 A Hybrid Scenario
We propose a hybrid scenario of server based mobile routing, where start, destination, and maybe some
additional requirements are send from the car over an available wireless communication network to a
server. The message containing this query will be very small. Unfortunately, this will most probably not
be the case for the resulting route, which may consist of hundreds of road segments and must be send
from the server to the car.
If routes consist mainly of subpaths which are unique shortest paths, then the sizes of the corre-
sponding messages can be reduced greatly if we represent the routes by via edges. Further compression
is possible by using via nodes which can be represented somewhat more compactly than edges but
require the preprocessing explained in Section 3.2.
The route has to be reconstructed in the car. This requires a mobile device present in the car, which is
equipped with map data which is not too far away from the current state of the road network and which
is able to compute optimal routes with respect to this data. However, the map data must be present in the
car anyway as driving directions have to be generated in the car in this setting. Of course, the server has
to know the map data present in the car exactly. To achieve this, the server could store every version of
the map data which has ever been delivered to any cars mobile device. The mobile device has to send the
unique ID of its map data to the server together with every query. Note, that outdated map data requires
that additional information is send from the server to the car if necessary. If this happens only rarely,
this is no problem. However, updates of the cars map data may be necessary from time to time.
The compression of routes using via edges or via nodes only makes sense, when the compressed rep-
resentation needs signicantly less space than the route itself. In this case the mobile device has to per-
form only few one-to-one shortest path computations to uncompress the route. However, with Dijkstras
algorithm this would still be to slow. This holds even more for mobile devices, which usually have little
main memory and slow ash storage. However, the current speedup techniques for shortest-paths com-
putation [5] can solve this problem. Especially, we recommend contraction hierarchies (CHs) [10, 7],
which were used to compress routes in Section 4. CHs are a very effective speedup technique, that has
already successfully been adapted to the needs of mobile devices [11, 12]. Mobile contraction hierar-
chies need on average 0.1 sec to answer a random shortest-path query on a data set of the European road
5
network [11]. So, if a route is represented by 10 via edges, for example, then we have to perform 11
one-to-one shortest path queries and the whole route can be constructed within 1.1 sec or even less
as the paths between the via edges are most likely not that long. Furthermore, to start driving, we only
need the rst part of the route quickly which requires only one shortest-path query.
Note, that on server systems CHs run much faster than on mobile divices: For the European road
network an average one-to-one shortest path query need less than 1 msec. As an example consider a
route, that consists of 700 edges and that can be represented by 10 via edges.
1
Then, Algorithm 2
performs up to 11 log690 104 one-to-one shortest path queries at less than 1 msec each. That means,
that the server needs less than 0.1 sec for compressing the complete route in this example.
As an alternative to via edges we can compress the route using a CH. In this case, the server needs
not only to know the map data present in the car but also the exact structure of the CH present in the
car. Also, we can combine both techniques as described in Section 4. However, which of the two
methods works better and whether the combination brings a signicant better compression rate within
less running time must be evaluated experimentally.
5.2 The Benet of Server-Based Mobile Route Planning
Server-Based mobile route planning enables several interesting applications. If the server has detailed
statistical data about the usual trafc volume of many streets for every day of the week and for every
time of day, then the server can compute a time-dependent optimal route [2, 3]. Or, the server could
compute a route which is optimal depending on a certain parameter which can be chosen freely by the
driver of the car (we call this exible route planning [8]). An example of such a parameter is the tradeoff
between travel time and fuel consumption. It is not at all trivial to make time-dependent and exible
route planning run efciently on a mobile device without communicating with a server. Also, the server
could support ride sharing [9]. Alternative routes are also interesting [1, 4].
6 Conclusions and Future Work
We have presented two methods to compress the description of routes, that is of paths. The rst method
represents a path, which is a sequence of edges, by a subsequence of via edges such that the missing
parts are unique shortest paths. Alternatively, via nodes can be used instead of via edges. Then, the path
requires less space but the road network requires more space. The second method replaces subpaths of
the original path by shortcuts that are taken from a contraction hierarchy. In contrast to the rst method,
the second method performs no one-to-one shortest path queries. As this is a relatively expensive oper-
ation, we expect that compression and uncompression take less time for the second method. However,
the two methods can also be combined. This way we hope to achieve better compression within less
running time than if rst method were used alone.
We also described a possible scenario: The routes are computed on a server, which enables us to
take the latest data about congestions and changes of the road network into account. Then we have to
transmit a description of the route to the car using some mobile communication network. As a route can
consists of hundreds of edges, the accordant message can get quite large. This is where we apply the
compression methods presented in this work. We hope to reduce the size of the message signicantly
not only to save costs but also to make our approach applicable even when bandwidths are low. We have
not evaluated these ideas with realistic data yet, but we hope to do this in the not too far future.
If the routes are computed on a server rather than in the car, many more interesting applications
are possible. These are time-dependent routeplanning [2, 3], exible route planning [8], support for
ride-sharing [9], and alternative routes [1, 4] for example.
1
Note, for comparison, that an average route in the European road network has about 1 400 edges. But most real-live routes
should be shorter.
6
References
[1] Ittai Abraham, Daniel Delling, Andrew V. Goldberg, and Renato F. Werneck. Alternative Routes in Road
Networks. In Festa [6], pages 2334.
[2] Gernot Veit Batz, Daniel Delling, Peter Sanders, and Christian Vetter. Time-Dependent Contraction Hier-
archies. In Proceedings of the 11th Workshop on Algorithm Engineering and Experiments (ALENEX09),
pages 97105. SIAM, April 2009.
[3] Gernot Veit Batz, Robert Geisberger, Sabine Neubauer, and Peter Sanders. Time-Dependent Contraction
Hierarchies and Approximation. In Festa [6], pages 166177.
[4] Jonathan Dees, Robert Geisberger, Peter Sanders, and Roland Bader. Dening and Computing Alternative
Routes in Road Networks. Technical report, ITI Sanders, Faculty of Informatics, Karlsruhe Institute of
Technology, 2010.
[5] Daniel Delling, Peter Sanders, Dominik Schultes, and Dorothea Wagner. Engineering Route Planning Al-
gorithms. In J urgen Lerner, Dorothea Wagner, and Katharina A. Zweig, editors, Algorithmics of Large and
Complex Networks, volume 5515 of Lecture Notes in Computer Science, pages 117139. Springer, 2009.
[6] Paola Festa, editor. Proceedings of the 9th International Symposium on Experimental Algorithms (SEA10),
volume 6049 of Lecture Notes in Computer Science. Springer, May 2010.
[7] Robert Geisberger. Contraction Hierarchies. Diploma thesis, Universit at Karlsruhe (TH), Fakult at f ur Infor-
matik, 2008. http://algo2.iti.uni-karlsruhe.de/documents/routeplanning/geisberger_dipl.pdf.
[8] Robert Geisberger, Moritz Kobitzsch, and Peter Sanders. Route Planning with Flexible Objective Functions.
In Proceedings of the 12th Workshop on Algorithm Engineering and Experiments (ALENEX10), pages 124
137. SIAM, 2010.
[9] Robert Geisberger, Dennis Luxen, Peter Sanders, Sabine Neubauer, and Lars Volker. Fast Detour Computa-
tion for Ride Sharing. In Proceedings of the 10th Workshop on Algorithmic Approaches for Transportation
Modeling, Optimization, and Systems (ATMOS10), OpenAccess Series in Informatics (OASIcs), 2010.
[10] Robert Geisberger, Peter Sanders, Dominik Schultes, and Daniel Delling. Contraction Hierarchies: Faster
and Simpler Hierarchical Routing in Road Networks. In Catherine C. McGeoch, editor, Proceedings of the
7th Workshop on Experimental Algorithms (WEA08), volume 5038 of Lecture Notes in Computer Science,
pages 319333. Springer, June 2008.
[11] Peter Sanders, Dominik Schultes, and Christian Vetter. Mobile Route Planning. In Proceedings of the 16th
Annual European Symposium on Algorithms (ESA08), volume 5193 of Lecture Notes in Computer Science,
pages 732743. Springer, September 2008.
[12] Christian Vetter. Fast and Exact Mobile Navigation with Open-
StreetMap Data. Diploma thesis, Karlsruhe Institute of Technology, 2010.
http://algo2.iti.kit.edu/download/christian_vetter_thesis.pdf.
7