To Re-Route, or not to Re-Route: Impact of Real-Time

Re-Routing in Urban Road Networks

Mohamed Amine Falek, Antoine Gallais, Cristel Pelsser, Sébastien Julien,
Fabrice Theoleyre

To cite this version:

Mohamed Amine Falek, Antoine Gallais, Cristel Pelsser, Sébastien Julien, Fabrice Theoleyre. To Re-
Route, or not to Re-Route: Impact of Real-Time Re-Routing in Urban Road Networks. Journal of
Intelligent Transportation Systems: Technology, Planning, and Operations, Taylor & Francis: STM,
Behavioural Science and Public Health Titles, 2021, �10.1080/15472450.2020.1807345�. �hal-02614875�

HAL Id: hal-02614875

https://hal.archives-ouvertes.fr/hal-02614875
Submitted on 21 May 2020

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
To Re-Route, or not to Re-Route:
Impact of Real-Time Re-Routing in Urban Road Networks

Amine M. Faleka,b , Antoine Gallaisb,c , Cristel Pelsserb , Sebastien Juliena , Fabrice

Theoleyreb
a
Technology & Strategy group, 4 rue de Dublin, 67300 Schiltigheim, France.
b
ICube Lab, CNRS / University of Strasbourg, Pole API, Boulevard Sebastien Brant, 67412
Illkirch Cedex, France.
c
LAMIH lab, CNRS / Polytechnic University Hauts-de-France, Le Mont Houy, 59313
Valenciennes Cedex 9, France.

ARTICLE HISTORY
Compiled May 21, 2020

ABSTRACT
Route planning represents a major challenge with a substantial impact on safety,
economy, and even climate. An ever-growing urban population caused a significant
increase in commuting times, therefore, stressing the prominence of efficient real-
time route planning. In essence, the goal is to compute the fastest route to reach the
target location in a realistic environment where traffic conditions are time-evolving.
Consequently, a large volume of traffic data is potentially required and the route
continuously updated. We thereby address the re-routing problem to answer ques-
tions such as when, how often, and where is re-routing worthwhile. We base our
study on a real dataset, comprising the travel times of the road segments of New
York, London, and Chicago, collected over three months. By exploiting this dataset,
we implement an optimal algorithm, able to mimic ideal predictions of road segment
speeds in the network. Thereby, allowing us to compute the lower bound of travel-
time to serve as a reference against other routing techniques. Mainly, we quantify the
achieved travel-time gain of a static, no re-routing, and continuous re-routing strate-
gies. Surprisingly, we find that traffic conditions are sufficiently stable for short time
windows, and re-routing a vehicle is very seldom useful when exploiting accurate
statistics at departure time. Typically, real-time re-routing should only be triggered
during rush hours, for long routes, passing through well-identified road segments.

KEYWORDS
road networks; route planning; real-time data; vehicle re-routing; traffic prediction

1. Introduction

Road network traffic congestion is a well known common cause of anger and frustration.
From an economic point of view, congestion has an even more drastic impact. The 2015
urban mobility scorecard (Schrank et al., 2015) reports that urban Americans traveled
an extra 6.9 billion hours and had to purchase an extra 3.1 billion gallons of fuel. To
reliably arrive on time, travelers had to plan 48 minutes for a trip that should last
only 20 minutes in light traffic. The estimated cost of congestion is steeply increasing

Amine M. Falek. Email: [email protected] - [email protected]

This work has been done when A. Gallais was employed by the University of Strasbourg
 since reported in 1982, with a cost of $42 billion against a total of $160 billion in
2015. While solutions exist to control traffic lights dynamically (Li and Sun, 2019), it
is highly probable that congestion would still occur.
Route planning algorithms (Bast et al., 2016) compute a route, which consists of
a sequence of road segments between a departure and arrival locations, and a travel-
time, denoting its associated end-to-end cost. Historically, mobile navigation devices
were used to guide the driver by autonomously computing a suitable route. With the
wide adoption of smartphones, though, many applications nowadays benefit from the
cloud where all computations are performed (Li et al., 2017).
Route planning requires traffic information (estimated speed, vehicle counts) gath-
ered from various sensors and probes within the network. Some mobile navigation
devices rely on the median recorded travel-times of a set of monitored road segments
to predict congestion. Smartphones and vehicles which usually embed GPS devices
are also exploited by regularly relaying traffic measurements to the cloud so that a
real-time model of the road network can be constructed (Ahsani et al., 2019). Real-
time traffic information is particularly valuable to reduce the cost of urban freight
transportation (Flamini et al., 2018).
Prediction as a service aims at predicting the travel time of each road segment, based
on past measurements (Wu et al., 2016). That way, a vehicle would use an alternative,
faster route if congestion has been predicted along its primary route. Uncertainty may
also be considered to compute a route with a given latest time of arrival (Lee et al.,
2019). While long-term prediction algorithms exist in the literature (Chen et al., 2019),
short-term predictions are still challenging. Traffic jams are complicated to predict (Hu
et al., 2017), and their impact on travel time depends on various local factors (e.g.,
speed limit, type of street, neighboring local roads).
While vehicle re-routing may reduce the impact of congestion, relying on real-time
data can be computationally expensive. Computing a route, directly on a small em-
bedded device is inconvenient as traffic data must be downloaded first, which may
correspond to a large volume of data when the target location is geographically far
away. Practically, this means that hundreds of weights have to be collected periodi-
cally, even if the target is in the same urban area. This represents a massive amount of
data when considering a large collection of users. Alternatively, the route computation
can be executed directly in the cloud: centralized servers reply in real-time to queries.
Thereby continuously re-evaluating the route of each vehicle until it reaches its des-
tination. Consequently, the load in the cloud is roughly proportional to the sampling
rate used to refresh the real-time data. In particular, it becomes challenging to exploit
route planning algorithms that rely on pre-processed data because of the continuous
traffic changes measured on the network.
In this paper, we quantify the benefit of using dynamic shortest path algorithms
that re-route vehicles when congestion increases. Indeed, alternative paths may provide
a faster route, but can only be identified when using real-time information. For this
purpose, we exploit a real dataset comprising the travel times of the road segments of
several cities (i.e., New-York, London, Chicago, and Cologne). To compare simulations
with these real datasets, we also used the TAPAS (simulated) dataset (Uppoor et al.,
2014). The contributions of this paper are as follows:
(1) we quantify the travel-time gain when using real-time data to re-route vehicles, com-
pared to a no re-routing approach that prohibits diverting from the path evaluated
at the departure. Our dataset highlights that re-routing is very seldom required real
conditions, even during rush hours;

2
(2) we provide an optimal algorithm to compute a lower bound for the travel time between
any pair of locations. The algorithm replays the real dataset to mimic ideal predictions.
The gain of using ideal predictions is below 10% compared with a continuous re-routing
solution, even for the worst route;
(3) we compare the impact of re-routing based on two types of datasets (a real dataset
vs. the simulated TAPAS dataset). Surprisingly, and contrary to current belief, we
obtain very different behaviors when studying re-routing strategies. Simulations tend
to exacerbate the randomness of travel time. Thus, simulations seem to not accurately
capture the complexity of urban road network dynamics, proving the relevance of
exploiting real datasets;
The remainder of this paper is organized as follows. Section 2 provides the related
work. We then present the routing strategies used in our evaluation in section 3, and
detail our models and assumptions. We explain our methodology in section 3.3, as well
as the detailed description of our datasets. Section 4 presents our comparative analysis
of the travel time gain of each routing strategy. We conclude and discuss future work
in section 5.

2. Related Work

Road networks are usually abstracted through graph data structures where ver-
tices and edges represent physical intersections and road segments, respectively. Edge
weights are labeled with various metrics, such as distances or travel times required to
join two vertices along a given edge.

2.1. Re-routing in road networks

The field of shortest path algorithms flourished in the last decade, with continuous
improvement in execution times. Pre-processing consists of transforming the initial
graph to reply efficiently to route queries (Bast et al., 2016). However, using real-time
information implies that the pre-processing phase has to be re-executed.
Gmira et al. (2019) propose to construct a delivery plan related to the Dynamic
Vehicle Routing Problem. Their solution collects speed values in real-time, and update
the path for a vehicle only if it becomes infeasible. However, they do not investigate
the sub-optimality cost, i.e., a route is not updated if no constraint is violated, even
if its travel time is not the smallest one.
Understanding when to re-route vehicles is critical to reducing the re-computation
cost. Pan et al. (2013) propose an infrastructure-based approach: when congestion is
detected, the system asks nearby vehicles to re-compute their shortest route. Conges-
tion threshold, as well as the set of vehicles to re-route, impact the efficiency of this
proposal significantly.
In the Dynamic Shortest Path problem (DSP), a re-routing algorithm tries to update
the shortest path to handle multiple edge weight updates (Chan and Yang, 2009). Such
an approach is much more efficient than re-executing the shortest path algorithm from
scratch.
While these approaches aim to devise the most accurate strategy to trigger route
computation, the objective of this paper is to instead quantify the gain in
travel-time when authorizing redirections or exploiting ideal predictions.

3
2.2. Real-time data sources
Collecting travel-time measurements of each road segment in real-time is a challeng-
ing objective. We may use the GPS trajectories of a collection of vehicles to deduce
the specific travel time of each road segment (Duan et al., 2018; Sanaullah et al.,
2016). For this purpose, each trajectory has to be mapped to a set of road segments
while minimizing the mismatch ratio (Falek et al., 2018). Ladino et al. (2016) pro-
pose to merge heterogeneous data sources (e.g., cameras, induction loops), to create a
combined, more accurate dataset. Imputation techniques (Chen et al., 2018) help to
reconstruct missing data (e.g., a road segment is not crossed for short time-period).
Recent approaches propose to adopt a synthetic model. Typically, the SUMO simu-
lator (Behrisch et al., 2011) is used to generate the mobility pattern of a large number
of devices, using realistic urban points of interest. Travel and Activity PAtterns Simu-
lation (TAPAS) was used to generate the well-known TAPAS-Cologne dataset (Uppoor
et al., 2014). However, to the best of our knowledge, no study has been conducted to
verify the accuracy of these simulations. In (Kamga et al., 2011), the authors, use
simulated data from VISTA (Ziliaskopoulos et al., 1999) to determine the impact of
incidents on travel times. Other traffic simulation software such as Dynameq (Dy-
nameq, 2020) and TRANSIMS (Barrett et al., 2002) provide simulated datasets of the
travel times in a road network.
In this paper, we highlight in Sections 4 and 4.4 the differences obtained between
the different approaches when using real (measurement-based) datasets vs.
the TAPAS dataset.

2.3. Traffic prediction

Recently, traffic prediction has received much attention in order to provide prediction
as a service (Liebig et al., 2017). These techniques try to consider the inherent charac-
teristics of road networks (e.g., flow conservation) to predict future trends accurately.
Predictions rely on computationally intensive techniques such as, e.g., bee colony opti-
mization (Dell’Orco et al., 2016), or spatiotemporal random field (Liebig et al., 2017).
Alternatively, Wang et al. (2019) propose to predict end-to-end travel times directly
but limiting its practical interest for route computation.
However, congestion is highly variable: Coifman and Mallika (2007) highlight that
48% of the congestion is difficult to predict. Li et al. (2014) concludes that accidents
are impossible to predict and even complicated to detect early.
Since existing prediction techniques are partly inaccurate and computationally in-
tensive, we aim in this paper to quantify its gain: if individual travel times for
each road segment can be ideally predicted, is the end-to-end travel time gain sig-
nificant?

2.4. Evaluating routing performance on dynamic road networks

Smith et al. (2014) evaluate the impact of accidents on traffic congestion using a
vehicular simulation and highlight the need for using actual traffic conditions to predict
the travel time.
Wang et al. (2013) provide a performance analysis of different route planning algo-
rithms (i.e., Dijkstra, static A∗ , dynamic Dijkstra, and dynamic A∗ ) in smart cities.
Their experiments use the TAPAS-Cologne dataset (Uppoor et al., 2014), which was
built by generating traffic demand using TAPAS and Gawron’s algorithm for traffic

4
 assignment. They consider a rush hour during a weekday to evaluate the impact of
traffic jams.
McArdle et al. (2012) attempt to simulate traffic in the Greater Dublin region.
Typically, each vehicle selects its destination according to a radiation model, to model
probabilities of interactions between different regions.
Unfortunately, all these performance comparisons use simulations with synthetic
models. The objective of our paper is instead to use real datasets, with the
actual travel time for each segment, at any instant, during very long periods
(i.e., a few months).

3. Methodology

A route planning solution identifies a path to follow for a vehicle. Most solutions can
be classified into:
(1) embedded devices help to compute a route, sometimes after having downloaded
real-time data about the congestion of the road network (Wang et al., 2015);
(2) cloud-based infrastructures receive a collection of queries from the vehicles that they
have to handle in real-time (Li et al., 2017). Continuously reconsidering the route be-
cause of real-time data is also expensive in a cloud serving a large number of customers,
where additional resources have to be provisioned.
Each route planning strategy has to solve individual queries, returning a route with
minimal travel time. Formally, we define a query as q = (s, d, ts )|s, d ∈ V and ts ∈ T ,
where s, d and ts represent the departure location, the destination location and the
departure time respectively. Likewise, a route is defined as an ordered list of vertices
(road segments) in the graph.
In this paper, we consider the following strategies, ordered by the volume of resources
(bandwidth and computation) they require:
(1) static: we do not have any knowledge of the actual traffic conditions (i.e., no data is
exchanged);
(2) no re-routing: we know the travel time of each road segment precisely just before
the vehicle leaves its starting point. For the sake of limiting computational cost, the
vehicle does not reconsider its decision after departure;
(3) continuous re-routing: we have continuous access to the most recent real-time data.
A vehicle may be redirected to a different (shorter) route as soon as traffic conditions
change;
(4) prediction based routes: we are able to predict the traffic conditions perfectly.
Consequently, we can identify the shortest ideal route, which constitutes our lower
bound;

3.1. Assumptions and Model

A road network is defined by a list of road segments. Each of them consists of a set
of consecutive coordinates (latitude and longitude) that define its shape. We use a
dynamic directed graph GT = (V, E, WT ) to represent the road network, where v ∈ V
is a vertex representing a physical intersection of two or more road segments, eij ∈ E
a directed edge from vertex i to j, and wij (t) ∈ WT the weight assigned to eij , as
a function of time t ∈ T . Table 1 contains definitions of the recurring keywords and

5
 Table 1.: Keywords & symbols definitions.

Keywords & Definition

symbols
eij Directed edge from vertex i to j
dij Length [m] of eij
sij (t) Speed [m/s] of eij at time t
wij (t) Travel time [s] of eij at time t
T Number of timeslots in a given dataset
δt = 60 sec Sampling rate of the datasets
∆t ≥ δt Variable sampling rate
ri = {u, v, .., w} Defines a road segment
R = q(s, d, ts ) Route in reply to the query from vertex s to d at timeslot ts
R = {u, v, w, .., z} A route is an ordered set of vertices
T T ime[q or R] Travel time of a query q (or a route R)
algo Designates one of four algorithms: static/no re-
routing/continuous re-routing/ideal
T T ime[q, algo] Travel time of q using algorithm algo
t≥0∈R Represents the time in seconds
0 ≤ tk ≤ T ∈ N Timeslot index of a real-time data update

symbols used throughout the paper.

For the sake of clarification, we use the term dynamic to refer to a graph whose
structure remains the same (no edges are deleted or inserted), but edge weights change
over the period T . We consider road networks with traffic congestion, but we neglect
the new roads that may appear.
We exploit a dataset where the speed of each road segment is monitored periodically
and synchronously. Thus, we denote by timeslot the discretized time, during which the
speeds for all the road segments remain unchanged. An edge weight wij (t) = dij /sij (t)
represents the travel time in seconds required to join the vertices i and j, where dij
and sij (t) represent the length of the edge eij and its corresponding speed at time t
respectively.
A road segment ri = {u, v, w, .., z} ∈ V consists of an ordered list of vertices in the
graph. When the speed on the road segment ri changes, it affects the weights of all the
edges associated with that road segment. Hence, in our implementation, every edge in
the adjacency list points to a specific road segment in the road segments map.

3.2. Route Planning Strategies

We now describe in more detail the route computation algorithms we use in the rest of
the paper to evaluate the importance of real-time speed data along with road segments.
Real-time information implies that vehicles may change their route when congestion
occurs. Practically, drivers may be progressively aware of current incidents because
they use different data sources (Kucharski and Gentile, 2019). We neglect here the
mutual impact of their decisions, i.e., travel-time may increase if all the drivers take
the same decision. We model in the rest of this section different families of routing
algorithms, taking into consideration traffic information.
Each strategy corresponds to a given travel time formulation. The here presented

6
 route planning strategies (i.e., static, no re-routing, continuous re-routing, and ideal
prediction based) respectively deal with travel times that are either time-independent
(i.e., no knowledge of traffic conditions), evolution-independent (i.e., knowledge of
traffic conditions at the initial computation time only), time-aware (i.e., knowledge of
initial traffic conditions and regular updates throughout the journey) or ideal (perfect
forecast of traffic conditions).

3.2.1. Static Route Planning

We assume here the system has no knowledge about traffic conditions (e.g., an embed-
ded device disconnected from the Internet). It represents our worst but thrifty strategy
and provides a baseline for comparison. Thus, it will use the maximum speed for each
road segment to compute the fastest route. The weights are time-independent, and
hence wij represents the travel time from vertex i to j, where sij is the speed limit in
this particular case. For a given query q(s, d, ts ), the departure time ts is meaningless
since the algorithm would always return the same route for a departure from s toward
the destination d. We use Dijkstra’s algorithm (Dijkstra, 1959) to compute the route
with the shortest travel time.

3.2.2. No re-Routing Route Planning

The system has here a complete knowledge of the travel times but never reconsid-
ers its decision. It mimics an embedded navigation system that is disconnected after
departure or cloud infrastructure that executes the query only once. Thus, we apply
the same strategy as previously, just executing Dijkstra’s algorithm, with the travel
time of each road segment at departure. This way, we can quantify the gain of using
up-to-date information before departure.

3.2.3. Continuous re-Routing Route Planning

We continuously reconsider the routing decision by trying to compute a better route
with shorter travel time, modeling the approach proposed by Chen et al. (2010). This
strategy helps to bypass congested areas when they appear, but it also consumes
more resources. If the computation is delocalized, the device has to retrieve all the
travel times periodically for its area. In a cloud, this means that the query has to be
re-executed continuously, consuming computational resources.
Let ∆t be the sampling rate of the real-time traffic data. The route planning algo-
rithm will re-compute the optimal route toward the destination at each crossroad. We
use a dynamic version of Dijkstra’s algorithm as detailed in algorithm 1:
(1) we use a time-dependent graph model, where the weight of each edge is time-variant.
When executing Dijkstra’s algorithm, we pick the most recent weights;
(2) the route is reconsidered when a new data sample occurs, i.e., it corresponds to an
update of speed data. In that case, the shortest route to the destination from the
current position is computed (line 4);
(3) we traverse the graph, following the current route (lines 7-17). We have to verify if we
can reach the next crossroad before the next speed sample occurs (line 11);
(4) if we cannot, we have also to consider the upcoming speed updates to reflect the actual
travel time (line 14).
(5) when a new sample occurs, we set the upcoming crossroad as the new position (line
18) from which we re-evaluate the route (step 1).

7
 Algorithm 1: Fastest Route with continuous re-Routing.
Data: departure vertex (s), destination vertex (d), departure time (ts ), sampling rate (∆t)
Result: shortest route as an ordered list of vertices (route)
/* Set current position to vertex s and current time to ts */
1 here ← s;
2 tnow ← ts ;
3 do
/* updates the route, from here to the destination */
4 route ← dijkstra(here, d, tnow ) ;
/* true if a new data sample occurs */
5 isN ewSample ← f alse;
/* traverse the route until a new data sample occurs */
6 while !isN ewSample do
/* next edge in shortest route */
7 evw ← getN extEdge(route) ;
8 lg ← getLength(evw ) /* get length of edge evw */
9 while lg > 0 do
/* maximum distance that can be covered by next speed change */
10 lgmax ← speed(evw , tnow ) ∗ (∆t − tnow mod ∆t);
11 if lg < lgmax then /* crossroad reached before next sample */
12 lg ← 0 ;
lg
13 tnow ← tnow + getSpeed(e ,t )
;
vw now

14 else /* crossroad not reached before next sample */

15 isN ewSample ← true;
16 lg ← lg − lgmax ;
lgmax
17 tnow ← tnow + getSpeed(e ,t
;
vw now )

18 here ← getHeadV ertex(evw );

19 while here 6= d;
/* we reached d */
20 return (route)

This version accommodates any sampling rate. This way, we can compare the impact
of the data accuracy on the travel time.

3.2.4. Ideal Prediction Based Route Planning

As mentioned earlier, traffic congestion forecasting is a technique used by most ad-
vanced route planning algorithms. By incorporating predictions, one could predict
recurrent traffic jams, and thus, plan the route accordingly at departure time.
We propose here to model such prediction-based routing algorithm (Liu, 2017).
More precisely, our goal is to quantify the maximal benefit attainable when using a
perfect forecast. Thus, we use an ideal prediction algorithm to compute the maximum
gain achieved by any prediction-based routing algorithm. To do so, we depart a vehicle
in the past by replaying the recorded measurements a posteriori.

We propose to use Algorithm 2:

(1) we first insert the departure vertex into the set of settled vertices (line 1) as the triplet
(vertex, parent, arrival time). We insert its neighboring vertices (i.e., head vertices of
outgoing edges) into a priority queue keeping the vertex with earlier arrival time at
the head;
(2) at each iteration, we poll a vertex v from the queue and insert it into the settled set
(line 5);

8
 Algorithm 2: Optimal Route with ideal predicion routing.
Data: departure vertex (s), destination vertex (d), departure time (ts )
Result: shortest route as an ordered list of vertices (route)
/* settled is a list containing all vertices for which the shortest route was found */
1 settled ← {(s, s, ts )}; /* (the settled vertex, its parent, arrival time) */
/* priority queue holds vertices that we either did not visit yet (∞) or for which we
did not find the shortest path */
2 queue ← {(v, −1, ∞)|v 6= s ∈ V };
3 route ← {d}; /* insert destination into route initially */
4 do
/* extract vertex v with smallest arrival time from the queue and insert it into the
settled list with departure time tv */
5 settled.add((v, u, tv ) ← queue.poll());
/* iterate over all outgoing edges from v */
6 for e ∈ getOutgoingEdges(v) do
7 w ← getHeadV ertex(evw );
8 tw ← tv ; /* initialize arrival time at w */
9 lg ← getLength(evw ); /* get length of evw as lg */
/* compute travel time of edge evw */
10 while lg > 0 do
/* maximum distance which can be covered by next speed change */
11 lgmax ← speed(evw , tw ) ∗ (∆t − tw mod ∆t);
12 if lg < lgmax then /* w is reached before next sample */
13 lg ← 0 ;
lg
14 tw ← tw + getSpeed(e ,t )
;
vw w

15 else /* next sample occurs before reaching w */

16 lg ← lg − lgmax ;
lgmax
17 tw ← tw + getSpeed(e ;
vw ,tw )

18 queue.push((w, v, tw ));
19 while d 6∈ settled;
/* browse parent vertices starting at d until s is reached */
20 parent ← getP arentV ertex(d);
21 while parent 6= s do
22 route.add(parent); /* insert into head of route */
23 parent ← getP arentV ertex(parent);
24 return (route)

(3) we compute the travel time required to reach each of its neighbors w (lines 7-18).
In particular, we traverse each edgevw , and we update its speed when a new sample
occurs before reaching the next crossroad (lines 15-17);
(4) we re-insert w into the queue with its corresponding parent vertex v and the updated
arrival time tw (line 18);
(5) when the destination vertex is settled, we construct the route by browsing backward
all the parent vertices starting at the destination until we reach the departure vertex
(lines 20-23).

3.3. Evaluation Workflow

To fairly compare different route planning strategies, we need to compute enough
routes using each strategy to cover most roads in the road network. To do so, we
generate a massive number of queries and solve each of them with our four routing
strategies. Hence, for a given route R = q(s, d, ts ), by varying the departure time ts ,
we can track the changes of R as a vehicle experiences congestion along the route from
s to d, at different times of the day.

9
 Figure 1.: Workflow for the stretch factor quantification of not using real-time data.

We insist on the fact that we do not simulate the deployment of multiple vehicles
on the road network at once. Indeed, we do not have any means to incorporate the
added congestion due to those vehicles, as one might expect in a mobility simulation
tool. Instead, we consider each query as representing a probe vehicle, to measure the
impact of the measured traffic (from our datasets) on its route (and not the inverse).
We apply here the following workflow to quantify the interest in exploiting real-time
data (Figure 1):
(1) we use real vs. simulated (TAPAS) datasets, and emulate different sampling rates
(from 1 to 30 min) by subsampling the datasets;
(2) we randomly select 1000 pairs of source and destination vertices in each road network;
(3) for every (source, destination) pair, we generate a broad set of queries at different
timestamps (every 1 min for the simulation and every 10 min on the real dataset);
(4) for each query, we execute each route planning strategy to extract the route to follow;
(5) we then emulate a vehicle moving along the route returned by the algorithm in the
previous step. This way, we can accurately evaluate the actual end-to-end travel time
for each strategy at different times of the day.
Finally, we use the previous results to compare the static, no re-routing, continuous
re-routing and ideal prediction strategies, detailed in sections 3.2.1, 3.2.2, 3.2.3 and
3.2.4 respectively.
We ran all our experiments on the High-Performance Computing (HPC) at the
University of Strasbourg. We generated thousands of jobs that were executed on Intel
Xeon Sandy-Bridge nodes with 16 cores and 64GO of RAM. The combined computa-
tion effort required approximately 50,000 CPU-hours.

3.4. Datasets
To evaluate the impact of the different route planning strategies, we use real travel
times for a broad set of road segments. Additionally, we also use a simulated dataset
for comparison purposes. We rely upon:

10
 0/1 1/1 1/1 0/1 1/1 1/1 1/1 1/ 1

s1 d1
0/1 2/3 2/3 div. routes 1/2 1/2 0/1 0/1 0/1 0/1 0/1
total routes
= 2/3

s2 d2

1/1 1/1 1/1 1/1 1/1 1/1 1/1

s3 d3
0/1 1/1 1/1 0/1 0/ 1
Route using continuous re-routing Route using ideal prediction Diverging vertex
Diverging route segments

Figure 2.: Diagram illustrating the process of computing the divergence ratio of a cell.

(1) the simulated TAPAS dataset: (Uppoor et al., 2014) focuses on a small German
City (Cologne), during a weekday (24 hours). It relies on the simulator SUMO to
simulate the traffic from a large set of emulated vehicles, pseudorandomly selecting
pairs of sources and destinations (e.g., home, office). We run the simulation using
SUMO to generate a dump file specifying the speed along every road segment at 1-sec
intervals. We subsample the dataset to get the same sampling rate used in the real
dataset.
(2) a real dataset: we use the dataset of three major cities (New York, London, and
Chicago) for which HERE (HERE Technologies, 2018) provides a fine-grain estimation
of the speeds for the most important road segments. We collected three months (i.e.,
from September 21st to December 18th 2017) worth of data at a sampling rate of
1 min, which allows an accurate estimation of the actual travel time experienced by
users. Additionally, we also use the city of Cologne, for the same geographical area
and during the same time window as the TAPAS dataset for a fair comparison.

3.5. Metrics for the Performance Evaluation

We now detail the metrics we used to compare the different routing strategies.

3.5.1. Identification of congestion

To identify the rush hour period, we compute the Congestion Factor (CF ) of each
query as the ratio of the optimal travel time (using the ideal routing strategy) to the
free-flow travel time. In the TAPAS dataset, the free-flow speed of a road segment
corresponds to its speed limit. In the real dataset, the free-flow speed was provided
as the average measured speed during low-volume periods and depends on the road
characteristics such as lane width.
Hence, given a query R = q(s, d, ts ), the congestion factor of the end-to-end route
R is defined as:

T T ime[R, ideal]
CF [R] = (1)
T T ime[R, f reef low]

11
 To analyze more clearly the behavior of each routing strategy, we need to classify the
routes according to their congestion level. For this purpose, we define for each route its
traffic flow (TF) metric, which represents the congestion of its most congested road
segment. For each route R = q(s, d, ts ), we measure the level of congestion of each
road segment ri ∈ R. More precisely, the congestion level corresponds to the relative
speed decrease compared with free-flow conditions:
 
speed(ri , t)
T F [r] = max (2)
speedf f (ri ) ri ∈R

where speed(ri , t) corresponds to the speed at time t for the road segment ri , and
speedf f (ri ) to its speed in ideal conditions (i.e., free flow). T F = 1 corresponds to
free-flow conditions, and T F = 0 corresponds to a complete halt of all vehicles on one
of its road segments.

3.5.2. Travel Time Stretch and Gain Factors

We will use the stretch factor in travel time to compare the different strategies. The
ideal strategy provides, by definition, the lowest end-to-end travel time and represents
our lower bound. The stretch factor for a query q(s, d, ts ) is defined as:

T T ime[q, algo]
SF [q] = (3)
T T ime[q, ideal])

where T T ime[q, algo] represents the end-to-end travel time for the route returned by
the algorithm algo (static/no re-routing/continuous re-routing/ideal) for the query q.
Notice that SF [q] ≥ 1, and hence, the higher the stretch factor gets, the worse is the
performance of algorithm algo compared to the ideal.
To specifically focus on the gain achieved by re-routing vehicles after their depar-
ture, we also compute the gain factor. It corresponds to the relative gain in travel
time through the path selected by the prediction-based and the redirecting strategies
compared with the static one (i.e., without redirection after the departure).

3.5.3. Identification of divergences

We focus on the continuous re-routing strategy to precisely determine where a vehicle
is practically re-routed because the congestion has changed. Practically, we identify
the divergence vertices, i.e., geographical locations where the next edge is different
with or without rerouting. Formally, given two routes R1 and R2 that share at least
the same first and last vertices, we define their diverging vertices as the set {j ∈ V |
∃i, k, w ∈ V | (eij , ejk ) ∈ R1 ∧ (eij , ejw ) ∈ R2 ∧ k 6= w}.
We construct a grid overlay where each cell is a 10 × 10 square meters. We consider
a large number of route queries (pairs of sources/destinations). We execute the route
computation for each query at different instants covering to cover the whole dataset.
For each cell celli , we consider all the routes that cross this cell. We compute its
divergence ratio (0 ≤ div[celli ] ≤ 1) as the ratio of the number of diverging routes to
the number of total routes. A route is diverging if the routes with and without the
re-routing strategy are different for at least one instant of the dataset. Hence, a cell
with a high divergence ratio means that more vehicles are re-routed when crossing this
cell.

12
 6
Congestion Factor
5
4
3
2
1
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time [hour]

Figure 3.: Congestion Factor distribution in Cologne simulation, Germany, over a

period of 24 hours.

Some cells may be traversed by only a few routes, leading to statistically meaningless
results. Thus, we only consider cells traversed by a significant number of routes (300
in this case).

4. Results

In a road network with no congestion, all routing strategies should return optimal
routes. Hence, our first goal is to distinguish the critical parts of the dataset by iden-
tifying the rush hours in each city. The remainder of the paper will solely focus on
evaluating the presented routing strategies during rush hours only.
We also provide a visual, interactive interface to showcase a sample of the NYC
dataset and obtained results at http://its-icube.com/. Each query is represented
individually, to visually represent the temporal characteristics of each route all along
the week. In particular, we can identify the diverging vertices, as well as the evolution
of the congestion.

4.1. Absolute Travel Time and Rush Hours

Figure 3 illustrates the CF in the Cologne TAPAS dataset. As expected, we identify
the rush hours in the morning (6:00-9:00) and afternoon (15:00-19:00) of the working
day. The optimal travel time is almost six times longer than the free-flow travel time
during these two periods.
Figure 4 illustrates the congestion factor for the experimental datasets. We focused
uniquely on Thursday, Friday, Saturday, and Sunday to provide readable charts (the
remaining weekdays display similar characteristics to Thursday). We can derive the
following observations:

(1) some queries benefit from travel times smaller than the free-flow. They correspond to
very short distances (a few hundred meters) when streets are empty;
(2) we observe the rise in congestion starting at 6:00 and intensifying in the afternoon
with a peak at 17:00.
(3) we can easily make a distinction between working days and weekends, which are much
less congested.

Similar observations hold for London and Chicago.

13
 4.0 Week Day
Thursday
3.5 Friday
Saturday
3.0
Congestion Factor
Sunday
2.5
2.0
1.5
1.0
free-flow
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time [Hour]

Figure 4.: Congestion Factor distribution in New York, based on weekday and daytime
over a period of 3 months.

3.0
Algorithm
2.5 static
no re-routing
Stretch Factor

2.0 continuous re-routing

1.5

1.0
Cologne_sim Cologne New York London Chicago

Figure 5.: Stretch factor of the static, no re-routing and continuous re-routing algo-
rithms on both the TAPAS (left) and the real (right) datasets.

4.2. Travel Time Stretch and Gain Factors

We now compare the achieved travel time for each routing strategy. In particular, Fig. 5
illustrates the stretch factor for the TAPAS dataset (left) vs. the real datasets (right).
The stretch factor denotes the travel time increase compared with the shortest path,
with an ideal dataset (i.e., with ideal predictions). We clipped the plot at SF = 3, as
it reaches ≈ 6 in the simulation.
The static routing strategy provides, as expected, the longest travel time. However,
the distribution is significantly different between the simulation and the real dataset.
With TAPAS, 28% of the queries have an SF greater than 1.5. In contrast, we only
report 1.6% of the queries in Cologne and approximately 4% in New York, London, and
Chicago. In Cologne, the SF using either the no re-routing or continuous re-routing
strategies yields optimal travel times most of the time. Only ≈ 12% of the queries
are characterized with a stretch factor greater than 1. In TAPAS, though, more than
75% of the queries have a SF > 1 even when using continuous re-routing. Moreover,
for several queries in the simulation, travel time is worse when using continuous re-
routing rather than no re-routing. The high variability of congestion profoundly affects
re-routing in the simulation as a vehicle might engage in a seemingly faster alternative
route only to become more congested than the initial route. In New York, London,
and Chicago, approximately 75% of the queries benefit from ideal travel times when
using continuous re-routing.
To further comprehend when continuous re-routing is advantageous, we pinpoint

14
 0.5
Algorithm
0.4 continuous re-routing gain
Travel Time Gain Factor ideal routing gain
0.3
0.2
0.1
0.0
0.1
0.2 n:15384 n:13598 n:11002 n:9608 n:9214 n:6258 n:3186 n:3034 n:2260 n:920
10 20 30 40 50 60 70 80 90 100
Traffic Flow [%]

Figure 6.: Travel time gain of continuous re-routing and ideal routing algorithms over
the no re-routing algorithm as function of traffic flow in New York

the travel time gain it offers (compared to no re-routing) based on the congestion level
(Fig. 6). We normalize the travel time difference of using continuous re-routing (and
ideal routing for reference) over the no re-routing strategy.
As expected, re-routing is relevant only when congestion occurs. The gain is, how-
ever, often negligible when the actual speed is close to 50% the speed limit. For very
congested routes (T F ≤ 10%), we reduce the travel time significantly by re-routing
the vehicle. Exploiting ideal predictions allows the gain to be even higher, to choose a
better route: the best route is selected, before the congestion’s increase.
Conclusion: continuous re-routing significantly reduces travel time for most
queries. In New York City, we can reduce by 15% (1st quartile), which seems to justify
the usage of a smart route planning strategy. However, predictions only marginally
decrease the travel time for all the measured datasets. The traffic conditions seem to
evolve smoothly, and redirecting the vehicle when the congestion occurs appears as a
sufficient strategy. Clearly, continuous re-routing is essential to re-route around highly
congested road segments. When the road network is less congested (T F ≥ 40%), the
gain factor quickly converges to zero.

4.3. Impact of Sampling Rate on Travel Time

For the continuous re-routing algorithm, the sampling rate ∆t impacts both travel and
execution times. For fine-grained values, the algorithm is aware of the latest changes
in traffic congestion and can re-route accordingly. However, it may also re-compute
unnecessarily the routes, increasing execution time significantly. Our goal is thus to
determine the right sampling rate for a good travel/execution time trade-off.
By varying the sampling rate ∆t from 1 to 30 minutes, we re-compute the queries
we generated using continuous re-routing. We only consider queries with a travel time
T T ime[q] ≥ 60min to make sure re-routing has potentially occurred for the largest
sampling rate ∆t = 30.
Figure 7a illustrates the distribution of the stretch factor as a function of the sam-
pling rate. Obviously, a higher sampling rate means inaccurate speed values. Thus,
the route planning strategy will take sub-optimal decisions. However, the re-routing
strategy performs well, even for average sampling rates. Surprisingly, in New York, for
instance, the road network variations can be efficiently handled even if measurements
are reported every 10 minutes.

15
 2.0
Road Network
Stretch Factor 1.8 New York
London
1.6 Chicago

1.4

1.2

1.0
1 2 3 4 5 10 15 20 25 30
Sampling Rate [min]

(a) Travel time stretch factor in New York, London and Chicago

2.25

Execution Time [ms]

1500
2.00
Stretch Factor

1.75 1000
1.50
500
1.25
1.00 0
1 2 3 4 5 10 15 20 25 30 1 2 3 4 5 10 15 20 25 30
Sampling rate [min] Sampling Rate [min]
(b) Cologne simulation stretch factor (c) New York execution time

Figure 7.: Impact of sampling rate on travel time in Cologne simulation, New York and
London (a, c and d) and execution time in Cologne simulation (b) using the continuous
re-routing algorithm.

Figure 7c illustrates the execution time for New York. At ∆t = 10min, the exe-
cution time drops from ≈ 500ms to less than 50ms, which provides a good trade-off
between travel time and execution time. Of course, there exists a myriad of algorith-
mic techniques in the literature, capable of significantly reducing the execution time.
Considering that ∆t dictates the number of times we have to re-compute the route,
we are only interested here in the rate of the change of execution time as ∆t increases.
Figure 7b represents the execution time for the TAPAS dataset. In complete contra-
diction with the real dataset results, the stretch factor distribution is almost the same
regardless of the sampling rate value. Again, we observe that congestion changes too
fast throughout the whole road network (even at ∆t = 1min), causing the algorithm
to inevitably re-route through highly congested road segments.
Conclusion: we can accommodate average sampling rates when using real-time
data. Five minutes provide enough accuracy (for all our road networks) to identify the
best routes while reducing the computational or bandwidth cost.

4.4. Route Divergence Patterns

Figure 8 illustrates the distribution of the diverging vertices in each road network,
as defined in section 3.5.3. A divergence vertex corresponds to a crossroad where
at least one vehicle has been re-routed to reduce travel time (i.e., traffic congestion
has changed since its departure). The divergence ratio counts the ratio of routes
(source/destination) for which the path diverges when crossing the cell, with and
without the re-routing strategy.
The road networks in the real dataset exhibit a small set of diverging vertices with

16
 1.0
Vertex Divergence Ratio
0.8

0.6

0.4

0.2

0.0
n:6575 n:108 n:6771 n:4677 n:5902
Cologne_sim Cologne New York London Chicago

Figure 8.: Diverging vertices divergence ratio distribution.

5 km 5 km 10 km

(a) Cologne simulation (b) Cologne (c) New York

Figure 9.: Diplaying the divergence ratio level (proportional to the red shade intensity).
The yellow dots depict diverging vertices with a divergence ratio ≥ 0.5.

a high divergence ratio, typically almost all diverging vertices have a divergence ratio
≤ 0.5. This means that only 50% of the routes that cross these cells are re-routed at
least once during the whole duration of the dataset.
Inversely, the simulated dataset (with TAPAS) exhibits a significant number of
diverging vertices. TAPAS generates pseudorandomly the traffic and estimates the
level of congestion, and the speed for each road segment. It seems that TAPAS exhibits
a very different pattern compared with the real datasets.
Figure 9 summarizes the obtained divergence patterns for both the TAPAS and
real datasets. We only represent here New-York City since other real datasets behave
similarly. Each graph corresponds to a heat map (a red cell corresponds to a cell with
a high divergence ratio). We also highlighted the divergence vertices with a divergence
ratio ≥ 0.5 (yellow dots).
Remark: while we identify many cells with a high divergence ratio for all the
graphs, we can, however, still make a distinction between the TAPAS and the real
datasets. In the TAPAS dataset of Cologne, the divergence seems present everywhere.
We assume that this behavior comes from the fact that the source/destination of each
vehicle is picked pseudorandomly, and the shortest route is used. Individual routing
strategies seem more complex, and the trips seem to follow a uniform distribution.
Conclusion: many cells exhibit a large divergence ratio, and we cannot directly use
this metric to trigger the route re-computation efficiently. However, we identified only a
small number of diverging vertices. Thus, we would be able to execute the computation
only when the route crosses these specific points, making the computation much more

17
efficient. We would reduce the processing load without increasing the end-to-end travel
time.

5. Conclusion & Perspectives

Modern intelligent systems guide vehicles through the fastest routes to avoid congested
areas. However, they need to exploit real-time data, where the speed of each road
segment has to be known precisely. Even worse, this real-time feature has a cost, in
bandwidth (data collection), and computation (re-computation of the route to the
destination). Interestingly, the no re-routing strategy provides close to ideal travel
times most of the time. Using the continuous re-routing strategy can further improve
travel time by avoiding very congested areas when they appear.
We also used a simulated dataset (TAPAS) that leads to very different results con-
cerning the travel time and the re-rerouting gain. TAPAS seems not able to capture
the road network characteristics accurately, and particularly its dynamics. This obser-
vation speaks in favor of working with real datasets to model realistic environments.
In future work, we plan to enhance the use of real-time data in our route planning
strategies. We considered here that each vehicle triggers a route computation after
reaching each crossroad (if the speed has changed). We may improve the re-routing
strategy by triggering a computation only when the routes may diverge. We identified
specific crossroads where vehicles have a higher probability of being re-routed. We
only considered selfish decisions in this study, i.e., each vehicle decides by itself when
it should re-compute its route to the destination. To reduce the congestion, we may
consider a more collective objective, redirecting some of the vehicles . Thus, we expect
to propose a strategy that could also consider the gain in travel time, depending on
the characteristics of the route and of the local area.

References

Ahsani, V., Amin-Naseri, M., Knickerbocker, S., & Sharma, A. (2019). Quantita-
tive analysis of probe data characteristics: Coverage, speed bias and congestion
detection precision. Journal of Intelligent Transportation Systems, 23(2):103–119.
doi:10.1080/15472450.2018.1502667.
Barrett, C., Bisset, K., Jacob, R., Konjevod, G., & Marathe, M. (2002). Classical and
contemporary shortest path problems in road networks: Implementation and exper-
imental analysis of the transims router. In: European Symposium on Algorithms
(ESA), pages 126–138. doi:10.1007/3-540-45749-6 15.
Bast, H., Delling, D., Goldberg, A., Müller-Hannemann, M., Pajor, T., Sanders, P.,
Wagner, D., & Werneck, R. F. (2016). Algorithm engineering, chapter Route Plan-
ning in Transportation Networks, pages 19–80. Springer.
Behrisch, M., Bieker, L., Erdmann, J., & Krajzewicz, D. (2011). Sumo–simulation
of urban mobility. In: The Third International Conference on Advances in System
Simulation (SIMUL 2011), Barcelona, Spain, volume 42.
Chan, E. P. F. & Yang, Y. (2009). Shortest path tree computation in dynamic graphs.
IEEE Transactions on Computers, 58(4):541–557. doi:10.1109/TC.2008.198.
Chen, C., Jiao, S., Zhang, S., Liu, W., Feng, L., & Wang, Y. (2018). Trip-
imputor: Real-time imputing taxi trip purpose leveraging multi-sourced urban

18
 data. IEEE Transactions on Intelligent Transportation Systems, 19(10):3292–3304.
doi:10.1109/TITS.2017.2771231.
Chen, C.-M., Liang, C.-C., & Chu, C.-P. (2019). Long-term travel time prediction
using gradient boosting. Journal of Intelligent Transportation Systems, 0(0):1–16.
doi:10.1080/15472450.2018.1542304.
Chen, Y., Bell, M. G. H., & Bogenberger, K. (2010). Risk-averse autonomous route
guidance by a constrained a* search. Journal of Intelligent Transportation Systems,
14(3):188–196. doi:10.1080/15472450.2010.484753.
Coifman, B. A. & Mallika, R. (2007). Distributed surveillance on freeways emphasiz-
ing incident detection and verification. Transportation research part A: policy and
practice, 41(8):750–767. doi:10.1016/j.tra.2006.12.001.
Dell’Orco, M., Marinelli, M., & Silgu, M. A. (2016). Bee colony optimization
for innovative travel time estimation, based on a mesoscopic traffic assignment
model. Transportation Research Part C: Emerging Technologies, 66:48 – 60.
doi:10.1016/j.trc.2015.10.001, Advanced Network Traffic Management: From dy-
namic state estimation to traffic control.
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische
mathematik, 1(1):269–271. doi:10.1007/BF01386390.
Duan, Z., Yang, Y., Zhang, K., Ni, Y., & Bajgain, S. (2018). Improved deep hy-
brid networks for urban traffic flow prediction using trajectory data. IEEE Access,
6:31820–31827. doi:10.1109/ACCESS.2018.2845863.
Dynameq (2020). Traffic simulation software your city can plan on. https://www.
inrosoftware.com/en/products/dynameq/.
Falek, M., Pelsser, C., Gallais, A., Julien, S., & Theoleyre, F. (2018). Unambiguous,
real-time and accurate map matching for multiple sensing sources. In: International
Conference on Wireless and Mobile Computing, Networking and Communications
(WiMob). IEE. doi:10.1109/WiMOB.2018.8589103.
Flamini, M., Nigro, M., & Pacciarelli, D. (2018). The value of real-time traffic infor-
mation in urban freight distribution. Journal of Intelligent Transportation Systems,
22(1):26–39. doi:10.1080/15472450.2017.1309530.
Gmira, M., Gendreau, M., Lodi, A., & Potvin, J.-Y. (2019). Managing in real-time a
vehicle routing plan with time-dependent travel times on a road network. Technical
Report 2019-4, CIRRELT.
HERE Technologies (2018). Real-time traffic. https://www.here.com. [Online; ac-
cessed May 2019].
Hu, W., Yan, L., Wang, H., Du, B., & Tao, D. (2017). Real-time traffic jams prediction
inspired by biham, middleton and levine (bml) model. Information Sciences, 381:209
– 228. doi:10.1016/j.ins.2016.11.023.
Kamga, C., Mouskos, K., & Paaswell, R. (2011). A methodology to esti-
mate travel time using dynamic traffic assignment (DTA) under incident con-
ditions. Transportation Research Part C: Emerging Technologies, 19:1215–1224.
doi:10.1016/j.trc.2011.02.004.
Kucharski, R. & Gentile, G. (2019). Simulation of rerouting phenomena in dynamic
traffic assignment with the information comply model. Transportation Research
Part B: Methodological, 126:414 – 441. doi:10.1016/j.trb.2018.12.001.
Ladino, A., Kibangou, A., Fourati, H., & de Wit, C. C. (2016). Travel time forecast-
ing from clustered time series via optimal fusion strategy. In: European Control
Conference (ECC), pages 2234–2239. doi:10.1109/ECC.2016.7810623.
Lee, H., Choi, S., Jung, H., Park, B. B., & Son, S. H. (2019). A route guidance system
considering travel time unreliability. Journal of Intelligent Transportation Systems,

19
 23(3):282–299. doi:10.1080/15472450.2018.1542303.
Li, X. & Sun, J.-Q. (2019). Multi-objective optimal predictive control of signals in
urban traffic network. Journal of Intelligent Transportation Systems, 23(4):370–388.
doi:10.1080/15472450.2018.1504294.
Li, Y., Jin, D., Hui, P., Wang, Z., & Chen, S. (2014). Limits of predictability for large-
scale urban vehicular mobility. IEEE Transactions on Intelligent Transportation
Systems, 15(6):2671–2682. doi:10.1109/TITS.2014.2325395.
Li, Z., Kolmanovsky, I. V., Atkins, E. M., Lu, J., Filev, D. P., & Bai, Y. (2017).
Road disturbance estimation and cloud-aided comfort-based route planning. IEEE
Transactions on Cybernetics, 47(11):3879–3891. doi:10.1109/TCYB.2016.2587673.
Liebig, T., Piatkowski, N., Bockermann, C., & Morik, K. (2017). Dynamic route
planning with real-time traffic predictions. Information Systems, 64:258 – 265.
doi:10.1016/j.is.2016.01.007.
Liu, J. (2017). Lstm network: a deep learning approach for short-term traffic forecast.
IET Intelligent Transport Systems, 11:68–75(7). doi:10.1049/iet-its.2016.0208.
McArdle, G., Lawlor, A., Furey, E., & Pozdnoukhov, A. (2012). City-scale traffic
simulation from digital footprints. In: ACM SIGKDD International Workshop on
Urban Computing, pages 47–54. doi:10.1145/2346496.2346505.
Pan, J., Popa, I. S., Zeitouni, K., & Borcea, C. (2013). Proactive vehicular traf-
fic rerouting for lower travel time. IEEE Transactions on Vehicular Technology,
62(8):3551–3568. doi:10.1109/TVT.2013.2260422.
Sanaullah, I., Quddus, M., & Enoch, M. (2016). Developing travel time estimation
methods using sparse gps data. Journal of Intelligent Transportation Systems,
20(6):532–544. doi:10.1080/15472450.2016.1154764.
Schrank, D., Eisele, B., Lomax, T., & Bak, J. (2015). Scorecard, urban mobility.
Technical report, The Texas A&M Transportation Institute and Inrix.
Smith, D., Djahel, S., & Murphy, J. (2014). A SUMO based evaluation of
road incidents’ impact on traffic congestion level in smart cities. Proceedings -
Conference on Local Computer Networks, LCN, 2014-Novem(November):702–710.
doi:10.1109/LCNW.2014.6927724.
Uppoor, S., Trullols-Cruces, O., Fiore, M., & Barcelo-Ordinas, J. M. (2014). Genera-
tion and analysis of a large-scale urban vehicular mobility dataset. IEEE Transac-
tions on Mobile Computing, 13(5):1061–1075. doi:10.1109/TMC.2013.27.
Wang, C., Pan, J., Xu, H., Jia, J., & Meng, Z. (2015). An improved a* algorithm for
traffic navigation in real-time environment. In: International Conference on Robot,
Vision and Signal Processing (RVSP), pages 47–50. doi:10.1109/RVSP.2015.20.
Wang, H., Tang, X., Kuo, Y.-H., Kifer, D., & Li, Z. (2019). A simple baseline for
travel time estimation using large-scale trip data. ACM Trans. Intell. Syst. Technol.,
10(2):19:1–19:22. doi:10.1145/3293317.
Wang, S., Djahel, S., Mcmanis, J., Mckenna, C., & Murphy, L. (2013). Compre-
hensive Performance Analysis and Comparison of Vehicles Routing Algorithms in
Smart Cities. In: Global Information Infrastructure Symposium (GIIS), pages 1–8.
doi:10.1109/GIIS.2013.6684365.
Wu, Y.-J., Chen, F., Lu, C.-T., & Yang, S. (2016). Urban traffic flow prediction
using a spatio-temporal random effects model. Journal of Intelligent Transportation
Systems, 20(3):282–293. doi:10.1080/15472450.2015.1072050.
Ziliaskopoulos, A., Waller, S., & Barrett, C. (1999). VISTA, Visual Interactive Sys-
tem for Transportation Algorithms. Technical report, UC Berkeley Transportation
Library.

20