Analysis of Urban Traffic Incidents Through Road Network Features

Abstract

Road traffic prediction is crucial for transport operators. Traffic operators use traffic simulators with different precision levels (from microscopic to macroscopic simulators) to capture the complex nature of mobility, especially in urban environments. Predicting the impact of traffic incidents (e.g., accidents, events and protests) is one of the major challenges faced by traffic operators due to their direct impact on traffic congestion with its negative effects on many aspects of our lives (economy, wellbeing, health, pollution, etc.). In this work, we analyse how we can characterise the impact of road incidents through features of the road network on a microscopic simulation platform as a benchmark for measuring incident impact. We confirm that the impact severity of a road incident varies between crowded and uncrowded roads. However, we show that features of the road where the incident happened on their own are not enough to infer the impact severity of the incident. By extending the characterisation of the incident to its surrounding region, we show that the impact of a road incident is also affected by its location and the characteristics of its neighbouring roads. Furthermore, we identify that the impact of road incidents spans beyond the surrounding area, thus requiring further features for an accurate prediction of road incidents.

Full text

Analysis of Urban Traffic Incidents
Through Road Network Features
Takfarinas Saber∗, Laurentiu Capatina†, and Anthony Ventresque∗
∗Lero, School of Computer Science, University College Dublin, Ireland
Email: {takfarinas.saber, anthony.ventresque}@ucd.ie
†AI Lab, Mantu, Nice, France
Email: lau.capatina@gmail.com
Abstract—Road traffic prediction is crucial for transport
operators. Traffic operators use traffic simulators with different
precision levels (from microscopic to macroscopic simulators)
to capture the complex nature of mobility, especially in urban
environments. Predicting the impact of traffic incidents (e.g.,
accidents, events and protests) is one of the major challenges
faced by traffic operators due to their direct impact on traffic
congestion with its negative effects on many aspects of our
lives (economy, wellbeing, health, pollution, etc.). In this work,
we analyse how we can characterise the impact of road
incidents through features of the road network on a microscopic
simulation platform as a benchmark for measuring incident
impact. We confirm that the impact severity of a road incident
varies between crowded and uncrowded roads. However, we
show that features of the road where the incident happened
on their own are not enough to infer the impact severity of
the incident. By extending the characterisation of the incident
to its surrounding region, we show that the impact of a road
incident is also affected by its location and the characteristics
of its neighbouring roads. Furthermore, we identify that the
impact of road incidents spans beyond the surrounding area,
thus requiring further features for an accurate prediction of
road incidents.
Keywords-Urban Traffic Simulation, Incident Impact, Road
Network Features.
I. INTRODUCTION
Traffic analysis is a domain with a particular interest
for towns and cities where traffic is an important issue.
Estimations suggest that almost 70% of the total global
population will live in cities by 2050 [1]. This is creating
major issues for city operators and planners, who have to
deal with complex and growing transportation networks–and
for citizens using the transportation networks for their daily
commute. For instance, in Dublin, 65% of daily commuters
used their private motorised transport, and 11% took the bus
in 2016, with an average commuting time of 32.2 minutes–
and increasing [2].
Many simulators have been proposed to capture the com-
plex nature of urban mobility and address its challenges [3],
[4]. These simulators aim at helping traffic operators and city
planners analyse the traffic under different conditions, learn
better traffic routing policies/strategies [5], evaluate/visualise
the effect of particular events (e.g., accident, protest), and
investigate the applicability and effectiveness of new gen-
eration of mobility technologies [6]. The literature outlines
three types of urban traffic simulators following their traf-
fic model [7]: (i) microscopic simulators, (ii) macroscopic
simulators, and (iii) mesoscopic simulators. Microscopic
simulators require more computation resources but at the
same time, they deliver results with the highest precision.
In general, their detailed models provide a more accurate
image of reality, making them the go-to simulators for a
large number of traffic operators.
Traffic incidents are one of the major challenges faced by
urban mobility as ∼70% of traffic congestions are caused by
accidents, and accidents cause large losses (e.g., accidents in
US cities cause losses exceeding 35 billion US dollars) [8].
The impact of traffic incidents on vehicles has already
been studied in various works. For instance, Lee et al. [9]
developed a traffic delay assessment tool for short-term
freeway road closures, Javid and Ramina [10] analyse travel
time variability resulting of traffic incidents, and He et
al. [11] study the impact of a road incident on travel speed.
Some other works attempt to predict the time it takes for
a traffic accident to be mitigated (e.g., [12]), while others
propose management systems to support traffic planners
in the event of a traffic incident (e.g., [13]). Despite the
numerous works in the area, the majority of them consider
freeway roads as their use-case with little clarity regarding
the extendability of their findings/tools to urban scenarios.
Zambrano-Martinez [14] modelled and characterised road
segments in the city of Valencia, Spain, with respect to travel
time during high traffic hours. However, this work is focused
on the road segments themselves and ignores the effects of
the surrounding region.
In this paper, we attempt to identify the role that differ-
ent road and region characteristics play in aggravating or
mitigating the impacts of traffic incidents. Particularly:
•We confirm that the impact severity of a traffic accident
varies between crowded and uncrowded roads.
•We classify roads based on their characteristics into
three classes (i.e., crowded, dynamic, and uncrowded)

and attempt to correlate roads of the same class with
the impact severity of incidents occurring on them. We
show that a classification based on road characteristics
is not enough on its own to infer the impact severity
of their incidents.
•We also expand the road network features including
the characteristics of the environment surrounding the
incidents and confirm that the impact is not only
influenced by the characteristics of the blocked road
but also by its location and by roads in it surrounding.
We also identify that the impact of a road incident goes
beyond the surrounding roads, and that other features
should be included if we would like to predict the
impact of a road with greater accuracy.
Identifying the impact of traffic incidents using
road/region characteristics will empower traffic planners
make faster and more accurate traffic predictions. It will
also provide the opportunity to improve the performance of
distributed traffic simulators (e.g., [15], [16]) by designing
better map partitioning techniques. For instance, Grammar-
Guided Genetic Programming [17], [18] can help create
partitioning algorithms that keep together regions that
are impacted by incidents–which will reduce network
synchronisation issues and significantly improve all tasks
that leverage such traffic simulators.
The remainder of this paper is organised as follows: Sec-
tion II reviews the background useful for the understanding
of this work. Section III details the experimental environ-
ment. Section IV confirms the existing difference in impact
severity between incidents in crowded and uncrowded roads.
Section V classifies roads based on their characteristics
and attempts to infer the incident impact severity based
on the classification. Section VI attempts to identify the
incident impact severity by considering the incident within
its surrounding region. Section VII concludes the paper.
II. BACKGROU ND
This section presents the background useful for the un-
derstanding of the paper in four parts: (i) types of traffic
simulators, (ii) SUMO; our microscopic simulator, (iii) ve-
hicle routing algorithms, and (iv) traffic incident simulation.
A. Types of Traffic Simulators
The literature outlines two types of simulations
(i.e., Discrete-Time Simulation and Time-Stepped
Simulation) [19], [20]. Urban traffic simulators can
be further divided based on their traffic models [7]:
•Microscopic Simulators: consider all the vehicles in
the simulation individually as agents: they all have
their attributes such as a specific behaviour, speed,
acceleration, length of the car, etc., and move on
their own while interacting with other vehicles. These
interactions between the agents lay on car-following
models. For instance, the Krauss model indicates with a
set of equations how the vehicles must move regarding
their dynamical attributes (e.g., their speed and their
position) and dynamical attributes of vehicles in front
of them.
•Macroscopic Simulators: do not have agents to repre-
sent the vehicles, but only flows of vehicles and aggre-
gated measures. in these simulators, we cannot track a
specific vehicle. We consider a number of vehicles that
compose a traffic flow object. All the vehicles in a flow
have the same behaviour including the same speed and
acceleration, and they do not interact with each other.
Traffic flow models lay on three main characteristics
that are the mean speed, the density (space occupied by
vehicles over the available space on a road segment),
and the flow. They define a relationship between these
characteristics using differential equations that describe
the traffic flows.
•Mesoscopic Simulators: manage the vehicles by keep-
ing them individually. But unlike the microscopic
model, there are fewer details as the vehicles are moni-
tored with a simplified car-following model. This model
is a hybrid/trade-off version of the microscopic and
macroscopic models. The characteristics of vehicles can
be not updated at each time step of the simulation but
only at a specific event like arriving on the next road.
B. Microscopic Simulator: SUMO
In this work, we use the simulator SUMO (Simulator
of Urban Mobility [21]). SUMO is an open-source, multi-
lane (it can handle roads with several lanes), multi-modal
(it can handle several means of transportation like cars,
bicycles, trains), microscopic traffic simulator. We chose
this simulator as it is open-source, has an important active
community, and has served in several studies (e.g., [15],
[16], [22], [23]).
SUMO offers several following-car models, has high
interoperability through the use of XML-data only, offers
a fast OpenGL graphical user interface, can handle different
right-of-way rules and traffic lights, etc.
In SUMO, simulations are time-stepped. Therefore the
time is divided into regular slices of times. The road network
is represented by a graph with edges as segments of roads
and nodes as intersections. Each edge is connected to only
two nodes and is unidirectional. An edge has multiple
attributes such as the number of lanes, the maximum speed,
the allowed vehicles types (e.g., buses, 4-passengers cars,
etc.). Nodes are characterised by their positions and can be
a representation of a right-of-way rule or a traffic light.
SUMO has an interface, TraCI (Traffic Control Interface)
which allows users to interact with the simulation through
a network port at run-time. Once users are connected to
the simulation, they can perform several commands such as:
trigger an event in the simulation, tell the simulator when to
process the next simulation step, add/remove new vehicles,

and have access to all the information of the network such
as the speed and position of every vehicle.
C. Routing Algorithms
The road network can be represented as a graph with
every vehicle aiming to minimise its time necessary to travel
from its origin to destination. The most commonly known
algorithm for this task is Dijkstra’s shortest path which
is exclusively based on the topology of the network. This
solution is eventually not depicting traffic routing accurately
as it omits the existing traffic (i.e., itineraries are calculated
without taking into account the potential existence of other
vehicles). Several works in the literature stress the necessity
for routing algorithms based on existing traffic load (e.g.,
[3], [24]).
Nagel [25] proposes a routing algorithm which takes into
account the traffic load by adapting vehicle itineraries step
by step until their convergence. This behaviour could be
seen as the way humans react when it comes to planning
a route. First of all, the k-shortest path algorithm calculates
the k shortest itineraries for every vehicle (which could be
seen as the different itineraries the driver needs to choose
from before starting his journey). Next, the algorithm has
two phases:
•The initialisation phase consists of launching a simu-
lation where each vehicle will choose a path not used
until now in a simulation. Consequently, there will be
a total of ksimulations executed in this phase. The
travel time will be saved after every simulation for
every chosen path.
•The daily execution phase consists of choosing the
shortest path in terms of travel time with a probability
of 95% and another one with a probability of 5%. The
number of vehicles which change their strategy will
decrease over time.
Through this process, the route of every vehicle is updated
and the chosen itinerary will be later used for the accident
measurements.
D. Simulation of Traffic Incidents
As a general definition, an incident is an unexpected event
which takes place in the road network. The presence of an
accident is often given as an example but others can exist
as well such as the presence of a street event, protest, etc.
In the rest of the paper, we will refer to an accident as an
incident.
In a network, the incident time can vary. For simplicity,
an incident will be represented in our simulations as a
blocked road with all its lanes (going in the same direction)
non-accessible for the whole duration. Vehicles arriving on
adjacent roads with a plan to pass through the blocked road
will have their itinerary updated by the simulator using the
shortest path from the vehicle’s position to its destination.
While the impact of an incident on vehicles’ travel time
was previously evaluated by Kamga et al. [26], vehicles were
given no or full knowledge about the incident. Therefore,
vehicles are either not deviating from their original path
(vehicles get stuck at the incident) or were informed about
the incident right at the moment when it appeared (vehicles
are rerouted upstream). These two scenarios have been
acknowledged by the authors as being extreme cases as, in
reality, only a small fraction of travellers are informed about
incidents as they happen. Consequently, we believe that it is
reasonable to assume that vehicles get informed about the
incident at the junction leading towards the affected route.
III. EXPERIMENTAL ENVIRONMENT
When defining an experimental setup, someone needs to
choose between either a real or a synthetic scenario. In our
case, we use synthetic scenarios as they have the advantage
of being simple while being configurable.
For our experiment, we designed three scenarios, each
with its road network (i.e., map) and vehicle itineraries (i.e.,
routes).
A. Road Networks
We create three grid road networks (as depicted in Fig-
ure 1) with the following properties:
•The size of the grid is 2km x2km for Scenario 1 and
2.4km x2.4km on the two others (i.e., Scenarios 2 and
3).
•In length, horizontal and vertical roads are 100m
and diagonal roads are 100√2m, whereas bridges are
200m.
•In Scenario 1, horizontal roads have 1 lane, whereas
vertical roads have two lanes. In Scenario 2, horizontal
and vertical roads have 2 lanes. In Scenario 3, hori-
zontal roads have 2 lanes, whereas vertical roads have
1 lane. In all scenarios, diagonal roads have 1 lane,
whereas bridges have 3 lanes.
•Each network has 6 or 7 randomly spread large junc-
tions.
•3, 5 and 7 bridges are present in the networks of
Scenarios 1, 2 and 7 respectively. Bridge positions are
determined using a normal distribution with Mean =
0.5and Sdev = 0.2.
•Traffic lights are placed on all junctions that are located
on every 300min the Y-axis.
•Junctions without traffic light are guided by the driv-
ing priority: bridges, vertical, horizontal, then diagonal
roads.
•The maximum allowed speed on a road is dependent
on the number of its lanes. One-lane roads are limited
to 13.9m/s, two-lane roads are limited to 22.2m/s and
three-lane roads are limited to 27.7m/s.

(a) Map in Scenario 1 (b) Map in Scenario 2 (c) Map in Scenario 3
Figure 1: Maps used in the experiments
B. Vehicle Itineraries
As stated above, every vehicle has a fixed itinerary before
launching a simulation. We generate vehicles and their
respective origin/destination as follows:
•The map of each scenario is split into four equally sized
zones (North West, North East, South West, and South
East).
•We select 1200 departures and 1200 arrivals from/to
each zone.
•We match the set of origins and destinations randomly
following a uniform distribution.
•We randomly generate the starting travel time of each
vehicle using a uniform distribution from the interval
[0,1500] seconds.
To route our vehicles, we use Nagel’s algorithm [25] with
each vehicle initially assigned 10 itineraries using the k-
shortest path, and the algorithm ran for 100 daily-execution
phases (i.e., converged itineraries).
IV. IMPAC T OF ROAD INCIDENTS ON VEHICLES
Our goal is to measure the impact of a road incident
on vehicles. We define two metrics to help us capture this
impact: delay and arrival time.
A. Delay
The delay of a vehicle vis the difference between its travel
time ttincident(v)in the event of an incident and its travel
time tt(v)in normal traffic conditions. Consequently, we
define the delay for every vehicle vas shown in Equation 1.
Delay(v) = ttincident(v)−tt(v)(1)
Figure 2 shows the cumulative percentage of vehicles
by their delay when introducing an incident in either a
crowded road (the road with the largest mean occupancy) or
an uncrowded road (road with the smallest non-zero mean
occupancy), from Scenario 1.
We can see from Figure 2 that the incident on the crowded
road has a more significant impact on the traffic than the
incident on a low frequented one. For instance, we notice
that almost 20% of vehicles have at least 2000s delay in
the case of an incident on a crowded road, whereas the
percentage is close to zero in the case of the low frequented
road.
We also observe that not all vehicles are delayed (a
percentage of vehicles improved their travel time). This
improved travel time is not as surprising. When a road is
blocked, there is less/no traffic flow on subsequent junctions.
Therefore, vehicles on the subsequent junctions are likely to
pass it faster.
B. Arrival Time
A second metric we evaluate in this paper is the arrival
time of vehicles to their destination.
Figure 3 shows the impact of an incident in a crowded
road from Scenario 1 in comparison to a low frequented
road from Scenario 1 (similar roads as for delay). In the
latter case, the impact is negligible as the ‘scattered points’
representing the arrival time of every vehicle in both cases
(with and without incident) almost overlap. However, in the
former one, there is a difference in arrival times. Considering
the arrival rate as the slope in every point of the simulation,
we can state that for the crowded road between 2,000s and
3,000s, the arrival rate decreased. This observation trans-
lates to the appearance of areas in the network where the
traffic is almost stuck for a certain time duration. After the
decongestion of these areas, the traffic resumes its normal
flow.
Our preliminary results show that there is a difference in
impact on the delays according to the traffic occupancy on
the blocked road.
V. EFF EC T OF RO AD CHARACTERISTICS
In a traffic environment, roads are more or less used
depending on several factors like if they are in a frequented
area or if they connect important points. Their capacity is
also significant as a one-lane road cannot take the traffic
of a three-lane road. A natural intuition would be that the
characteristics of an edge have a significant influence on the
outcome of an incident present on that edge. Consequently, a
classification of the roads is needed in case of normal traffic
(i.e., without the presence of an incident).
A. Road Clustering
To cluster the roads, we use two variables measured
during the incident-free simulation, i.e., the speed and
the occupancy of an edge at every simulation step. Since
variables evolve for every road in the network over the
simulation steps, we aggregate these results by measuring
their mean and standard deviation.
To categorise the roads, we use a straightforward classi-
fication method using unsupervised learning. The k-means
algorithm was used to obtain three clusters after normalising
the data over the four dimensions. The k-means algorithm
is effective in low dimension data. It is applicable to multi-
class problems and its only parameter is k (i.e., number of
clusters).

(a) incident on crowded road (b) incident on low frequented (uncrowded) road
Figure 2: Cumulative percentage of vehicles by their delay with traffic incident in either a crowded or an uncrowded road
(a) Incident on crowded road (b) Incident on a low frequented road
Figure 3: Arrived vehicles (in %) at every time step. Blue: in normal traffic conditions. Red: in presence of a traffic incident.
Figure 4 shows road clusters identified by k-means on
the 3 scenarios with four road features and k=3. Data in
Figure 4 is plotted in 3 dimensions for visualisation purposes
(we omit mean speed). From Figure 4, we can identify:
in yellow; a cluster of crowded roads (high mean road
occupancy percentage), in blue; a cluster of uncrowded roads
(low mean occupancy and low speed standard deviation), and
in red; dynamic roads (low mean occupancy, but medium
speed and occupancy standard deviation).
B. Delay in Crowded Roads
We aim to discover a correlation between the class of the
road and the delays of vehicles in case of an incident on
that road. By following a natural intuition, for roads with
the same class, we should obtain similar results.
Figure 5 shows the cumulative percentage of vehicles per
delay when an incident happens in a road that is classified
as crowded with the classification method in V-A. Results
are shown for two scenarios where the incident happening
in a different and randomly selected crowded road.
The road blocked due to an incident in Figure 5a has
Figure 4: K-means road classification into three categories;
every point is a road of the network

a more significant impact on the cumulative delays of
vehicles present in the network. One might argue that the
classification method is not accurate enough. However, we
found a similar trend between multiple crowded roads which
let us believe that features used in the classification alongside
number of clusters are not enough to correlate incident
impact and topology of roads.
We believe that we should take into consideration not
only the road itself with all its characteristics but also
the surrounding environment. The deviation of the traffic
goes into the surrounding region and we have to take into
account the capacity of absorption of the flow in this area.
An example would explain better this situation: an incident
blocks a crowded road and the traffic is spread on the
surrounding routes. If this region has other roads which
are only barely used, this is an important advantage in
comparison to another region which is highly saturated. The
impact on the delay is not the same in the two cases.
VI. EFF EC T OF REGION CHARACTERISTICS
In our paper, we have the particularity that an incident
exists in the network and we want to react according to
its potential impact. We have divided our maps using the
spatial partitioning in [27]. We select five 400x400 metre
regions of interest from each map (location of each region
could be seen in Figure 6), each with its particularity
(proximity to a bridge, middle of the map, corner of the map,
include important junctions, etc.). The goal is to observe
the impact of an incident inside every region separately.
More precisely, we measure the number of impacted vehicles
passing through each of the corresponding regions.
Due to the difference in scale between vehicles travel time
(and therefore, their delay), we move from Delay(v)to the
more ‘scaled’ DelayRatio(v)as shown in Equation 2.
DelayRatio(v) = ttincident(v)−tt(v)
tt(v)(2)
A small delay of a couple of seconds is negligible in no
matter what case, and a delay has to be compared to the
initial time spent in the traffic (being five minutes late does
not have the same meaning for a travel time of ten minutes
or two hours). In our work, we consider that a vehicle vis
impacted by the traffic incident if and only if its Delay Ratio
is larger than 0.2 (which corresponds to an increase of 20%
in its travel time).
A. Region Analysis
For every region in each scenario, we conduct multiple
simulations with an incident in one of its roads at a time.
In every simulation, we count the number of affected
vehicles, that are passing through the region covering the
road with the incident. We also compute the percentage of
affected vehicles passing through the region (out of number
of affected vehicles).
Table I reports the minimum, average, and maximum
number of vehicles passing through a region that are affected
by the introduced incident in a road of the same region,
in every scenario. It also reports, between parentheses, the
average, minimum and maximum percentage of vehicles
affected within each region where the incident happened,
out of the number of affected vehicles in each simulation.
We see in Table I that the percentage of impacted vehicles
varies a lot and it does not get closer to 100%. This means
that the incidents do not limit their impact to vehicles
passing through the regions of interest. The highest value
that we got, 72%, means that even in the most affected
case, more than a quarter of the vehicles were impacted
indirectly. On the other hand, there is also the case where
the number of impacted vehicles which cross the region is
negligible. Almost all of them were outside of it. These
observations support the hypothesis that the incident impact
is not influenced only by the blocked road.
After integrating our road classification from Section V-A
into this experiment (in each region, simulations were per-
formed for various types of roads according to the clas-
sification), we were unable to find correlations between
the road category and the incident impact. In other words,
acrowded road was not a guarantee that the number of
impacted vehicles in the region will be higher than in the
case of an uncrowded one. This result confirms again that
the classification based on road characteristics is not enough
to predict the impact of incidents in terms of delays.
When comparing the capacity of handling an incident in
different regions, we notice that the higher is the region’s
margin (i.e., difference between its capacity and occupancy),
the more capable it is to handle an incident (i.e., the impact
of an incident will be less important).
We also perform a region analysis while varying region
sizes between 300x300 metres and 400x400 metres. Table II
shows the average number of affected vehicles passing
through each region of interest after introducing an incident
on various routes of the same region.
We notice from Table II that the variation in region size
(at least between 300x300 and 400x400) has a noticeable
impact on the average number of affected vehicles in certain
regions (e.g., 4 and 5). However, it is always lower than
the maximum number of affected vehicles seen in Tables I.
We also notice that for other regions such as 2 and 3 from
Scenario 1, the increase in region size has a negligible
impact. This is mainly due to the fact that theses two regions
are peripheral, thus they do not affect too much the traffic
that is not departing from or arriving at them. Therefore,
the size of the region has an impact but this impact is less
important for peripheral regions.
VII. CONCLUSION
Road incidents are a major challenge for urban mobility
planners. In our paper we analysed the effect of road

(a) Crowded Road 1 (b) Crowded Road 2
Figure 5: Delay with and a traffic incident on one of two random roads classified as ‘crowded’
(a) Map in Scenario 1 (b) Map in Scenario 2 (c) Map in Scenario 3
Figure 6: Maps used in the experiments with highlighted regions of interest
Table I: Number and percentage (between parentheses) of impacted vehicles after introducing an incident
Region Scenario 1 Scenario 2 Scenario 3
Avg Min Max Avg Min Max Avg Min Max
1 186 (29.2) 110 (24) 443 ( (39) 41 (10) 31 (8) 59 (15) 32 (7.8) 26 (6) 40 (10)
2 6 (1.6) 3 (1) 8 (2) 104 (20.3) 57 (15) 217 (28) 217 (48.1) 178 (42) 259 (52)
3 98 (16.5) 59 (14) 199 (25) 189 (46.1) 0 (0) 334 (59) 101 (21.4) 42 (12) 212 (37)
4 717 (64.5) 301 (57) 1379 (72) 278 (53.0) 163 (45) 557 (63) 185 (37.7) 130 (31) 258 (42)
5 257 (45.9) 206 (36) 342 (56) 56 (13) 37 (10) 91 (18) 108 (24.3) 82 (21) 147 (31)
Table II: Average number of affected vehicles passing
through regions of interest after introducing an incident
Region Scenario 1 Scenario 2 Scenario 3
300x300 400x400 300x300 400x400 300x300 400x400
1 186 220 41 46 32 46
2 6 6 104 108 217 220
3 98 101 189 193 101 118
4 717 812 278 294 185 232
5 257 340 56 98 108 127
network characteristics on the severity of road accidents, we
particularly looked at the characteristics of roads and the
characteristics of their surrounding regions. We confirmed
that the impact severity of a traffic accident varies between
crowded and uncrowded roads. However, we found that
incidents on crowded roads (classified based occupancy and
travel speed) do not systematically lead to severe impacts,
which let us hypothesise that these road characteristics are
not enough on their own to infer the impact severity of
their incidents. After expanding the road network features to
include the characteristics of the incident’s surrounding, we
found that the impact of an incident is not only influenced
by the characteristics of the blocked road but also by its
location and by its neighbouring roads. More crucially, we
have shown that the impact of a road incident goes beyond
the surrounding roads. Therefore, features other than road
and region characteristics need to be included for a more
accurate prediction of road incidents. In the future, we would
like to analyse the influence of map partitioning algorithms
and their resulting maps on distributed traffic simulation.
ACK NOW LE DG EM EN T
This work was supported, in part, by Science Foundation
Ireland grant 13/RC/2094.

REFERENCES
[1] D. o. E. United Nations and P. D. Social Affairs, “World
urbanization prospects: The 2018 revision,” 2019.
[2] M. Sheng and B. Sharp, “Commuter’s transport mode prefer-
ences and social network effects in new zealand,” Journal of
Transport Economics and Policy, vol. 53, no. 1, pp. 19–46,
2019.
[3] T. Saber, A. Ventresque, and J. Murphy, “Rothar: Real-time
on-line traffic assignment with load estimation,” in DS-RT.
IEEE/ACM, 2013, pp. 79–86.
[4] T. Saber, C. Cachard, and A. Ventresque, “Ronin: a sumo in-
teroperable mesoscopic urban traffic simulator,” in SmartCity.
IEEE, 2020.
[5] T. Saber and S. Wang, “Evolving better rerouting surrogate
travel costs with grammar-guided genetic programming,” in
CEC. IEEE, 2020, pp. 1–8.
[6] M. Bendechache, T. Saber, G. Muntean, and I. Tal, “Ap-
plication of blockchain technology to 5G-enabled vehicular
networks: Survey and future direction,” in MCWN, 2020.
[7] S. Plakolb, G. Jäger, C. Hofer, and M. Füllsack, “Mesoscopic
urban-traffic simulation based on mobility behavior to calcu-
late NOx emissions caused by private motorized transport,”
Atmosphere, vol. 10, no. 6, p. 293, 2019.
[8] Q. Yang, Y. Fang, L. Zheng, X. Zhou, and B. Peng, “An
expressway traffic incident detection method based on convo-
lutional neural network and extreme gradient boosting,” 2020.
[9] C. Lee, D. A. Noyce, and X. Qin, “Development of traffic
delay assessment tool for short-term closures on urban free-
ways,” Transportation Research Record, vol. 2055, no. 1, pp.
39–48, 2008.
[10] R. J. Javid and R. J. Javid, “A framework for travel time
variability analysis using urban traffic incident data,” IATSS
research, vol. 42, no. 1, pp. 30–38, 2018.
[11] Y.-q. He, Y.-l. Rong, Z.-p. Liu, and S.-p. Du, “Traffic
influence degree of urban traffic accident based on speed
ratio,” Journal of highway and transportation research and
development (English edition), vol. 13, no. 3, pp. 96–102,
2019.
[12] L. Kuang, H. Yan, Y. Zhu, S. Tu, and X. Fan, “Predicting
duration of traffic accidents based on cost-sensitive bayesian
network and weighted k-nearest neighbor,” Journal of Intel-
ligent Transportation Systems, vol. 23, no. 2, pp. 161–174,
2019.
[13] M. Won, “Development of a traffic incident management
support system,” Ph.D. dissertation, University of Maryland,
College Park, 2019.
[14] J. L. Zambrano-Martinez, C. T. Calafate, D. Soler, J.-C. Cano,
and P. Manzoni, “Modeling and characterization of traffic
flows in urban environments,” Sensors, vol. 18, no. 7, p. 2020,
2018.
[15] Q. Bragard, A. Ventresque, and L. Murphy, “Self-balancing
decentralized distributed platform for urban traffic simula-
tion,” Transactions on Intelligent Transportation Systems,
vol. 18, no. 5, pp. 1190–1197, 2017.
[16] Q. Bragard, A. Ventresque, and L. Murphy, “dSUMO: To-
wards a distributed SUMO,” in SUMO conference, Berlin,
Germany, 2013.
[17] T. Saber, D. Fagan, D. Lynch, S. Kucera, H. Claussen,
and M. O’Neill, “Multi-level grammar genetic program-
ming for scheduling in heterogeneous networks,” in EuroGP.
Springer, 2018, pp. 118–134.
[18] T. Saber, D. Fagan, D. Lynch, S. Kucera, H. Claussen, and
M. O’Neill, “A multi-level grammar approach to grammar-
guided genetic programming: the case of scheduling in het-
erogeneous networks,” Genetic Programming and Evolvable
Machines, vol. 20, no. 2, pp. 245–283, 2019.
[19] M. Bendechache, S. Svorobej, P. T. Endo, M. N. Mario, M. E.
Ares, J. Byrne, and T. Lynn, “Modelling and simulation of
ElasticSearch using cloudsim,” in DS-RT. IEEE, 2019, pp.
1–8.
[20] M. Bendechache, S. Svorobej, P. Takako Endo, and T. Lynn,
“Simulating resource management across the cloud-to-thing
continuum: A survey and future directions,” Future Internet,
vol. 12, no. 6, p. 95, 2020.
[21] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz,
“SUMO – simulation of urban mobility: An overview,” in
SIMUL 2011. ThinkMind, October 2011.
[22] M. Shamim Akhter, N. Ahsan, S. J. S. Quaderi, M. A.
Al Forhad, S. H. Sumit, and M. R. Rahman, “A sumo based
simulation framework for intelligent traffic management sys-
tem,” Journal of Traffic and Logistics Engineering Vol, vol. 8,
no. 1, 2020.
[23] E. Maiorov, I. Ludan, J. Motta, and O. Saprykin, “Developing
a microscopic city model in sumo simulation system,” in
Journal of Physics: Conference Series, vol. 1368, no. 4. IOP
Publishing, 2019, p. 042081.
[24] D. Braess, A. Nagurney, and T. Wakolbinger, “On a paradox
of traffic planning,” Transportation science, vol. 39, no. 4,
pp. 446–450, 2005.
[25] K. Nagel, “Individual adaptation in a path-based simulation of
the freeway network of northrhine–westfalia,” International
Journal of Modern Physics C, vol. 7, no. 06, pp. 883–892,
1996.
[26] C. N. Kamga, K. C. Mouskos, and R. E. Paaswell, “A
methodology to estimate travel time using dynamic traffic
assignment (dta) under incident conditions,” Transportation
Research Part C: Emerging Technologies, vol. 19, no. 6, pp.
1215–1224, 2011.
[27] A. Ventresque, Q. Bragard, E. S. Liu, D. Nowak, L. Mur-
phy, G. Theodoropoulos, and Q. Liu, “SParTSim: A space
partitioning guided by road network for distributed traffic
simulations,” in DS-RT. IEEE/ACM, 2012, pp. 202–209.