Transportation Planning and Technology, April
June 2007

Vol. 30, Nos. 2
3, pp. 289
309

ARTICLE

A Microscopic Simulation Model

for Incident Modeling in Urban
Networks

YASER E. HAWAS
Department of Civil and Environmental Engineering, UAE University, Al-Ain, UAE

(Received 28 December 2005; Revised 19 February 2007; In final form 11 April 2007)

ABSTRACT This paper reviews the main modules of an integrated system for
incident management in real-time, . A core to such a system is a
microscopic simulator with extended abilities to model the temporal and spatial
evolution of specified non-recurrent traffic conditions. The paper reviews the
mathematical formulation of the car-following and lane-changing modules. The
model is validated using a simulation-based approach. Concluding comments on
the general validation process of the model are provided. The paper finally
presents a sample of the accident patterns replicated by the model together with
their implications for real world validation.
KEY WORDS: Microscopic modeling; traffic simulation; car following;
calibration; incident modeling

Introduction
Traffic Control Centers (TMCs) are becoming essential elements to
monitor, control and manage traffic operations in urban networks.
TMCs operate various types of algorithms for network monitoring,
signal control, incident detection and management, drivers’ informa-
tion systems, variable message sign control, etc. Real-time traffic
simulation is a very effective tool to replicate traffic conditions and

Correspondence Address : Civil and Environmental Engineering Department, College of

Engineering, UAE University, Al-Ain, P.O. Box 17555, UAE. Email: [Link]@[Link]

ISSN 0308-1060 print: ISSN 1029-0354 online # 2007 Taylor & Francis
DOI: 10.1080/03081060701398117
290 Yaser E. Hawas

identify optimal control strategies. Several network simulation models

exist in the literature. Nonetheless, only few of these models offer the
reasonable capabilities of simulating non-recurrent traffic conditions.
The relationships utilized in macroscopic simulation models are
mostly curve fitting relations whose suitability can be investigated from
the method of data collection, frequency and data domain. Ideally, the
data should cover the entire domain the relationship is intended to
cover. Some of the models suffer from serious errors caused primarily
by the mismatch between the data domain and the relationship. The
well-known relation of Greenshield (1935) has been criticized with
regard to its data domain (Hall, 1996). Greenshield is a linear model of
speed and density. Data consisted of a cluster of six points near free-
flow speed, and only a single observation at congested conditions. The
linear relationship comes from connecting the ‘clustered’ points with
the single point. Other relationships in the literature (Greenberg, 1959;
Edie, 1961) are indeed weak in the sense that they are derived from a
very limited data domain.
Macroscopic traffic models are easier to calibrate and less costly to
execute or maintain. They, on the other hand, carry a higher risk that
their representation of the real world is less accurate or perhaps
inadequate. They can be justified if the significant details (such as queue
build up and dissipation for their effect on the delay) are modeled
accurately with more attention given to the parameters that highly
affect the overall performance. Also, they can be justified, if the
application for which the model is employed does not require accurate
description of all the network details, but rather for only some, which
can be modeled microscopically joint with the other macroscopic
details. One limitation of such models is their ability to represent non-
recurrent conditions caused by incidents. Furthermore, they cannot be
used to model vehicle detectors, and, as such, they offer limited
applicability for advanced traffic control systems (such as responsive
detector based signals) and incident detection and management
applications.
Other models use time series techniques to predict speed, traffic
volume or occupancy given existing traffic or historically stored
detector data. The primary advantage of these techniques is the ability
to react to short-term fluctuations and to capture the dynamic flow
pattern effectively (Davis & Nihan, 1984; Tavana & Mahmassani,
2000). The parameters are calculated via some specific off-line
optimization technique (Tavana & Mahmassni, 2000). The results
and parameters are only applicable to the situations from which
detector data were extracted. The model structure does not prevent
errors of the speed estimates at any time t from propagating at further
 A Microscopic Simulation Model for Incident Modeling 291

time intervals. Detector malfunctions, or incorrect data entry to the

systems, might also lead to inaccurate estimates.
Microscopic models depend extensively on car-following models.
The rationale is to utilize observed distributions (e.g. acceleration/
deceleration distributions) to predict the driver’s behavior (selection
of inter-spacing from the leading car, acceleration, deceleration,
etc.) in response to a specific stimulus. Car-following models are
very effective, especially in modeling traffic in a confined space (e.g.
single lane traffic), where driver actions become limited. None-
theless, data available for model calibration and verification are
limited. Car-following models commonly utilize three functions to
describe the perception of the driver to a specific stimulus: the
decision making process in response to stimuli, and the final control
or reaction by the driver. Car-following models differ in their
representation of the control-stimuli function. Some models assume
a linear relationship, where control is expressed as a multiplier
of the stimulus, or a non-linear model, or a more generalized
model.
Several car-following models have been proposed in the literature to
describe driver behavior in a traffic stream. They determine the rate of
acceleration/deceleration of the following vehicle based on the actions
of the lead vehicle(s). Equations of motion are then used to compute the
speed and the position of the following vehicle for any given time
interval.
Four different types of car-following models exist in the literature:
stimulus-response models, safety-distance or behavioral models, psy-
chophysical or action point models, and fuzzy logic-based models
(Brackstone & McDonald, 1999). The limitations of the existing
models can be summarized as follows:
1. Several models assume empirical mathematical functions that may
not adequately explain the actual cause-effect behavior of the driver.
2. Some models assume that following vehicle drivers respond only to
the lead vehicle without observing other vehicles ahead. In actual
traffic situations, drivers account for the traffic conditions ahead,
surrounding and even behind in setting their desired speeds, spacing
and in switching lanes.
3. Several models assume symmetrical driver responses to positive/
negative changes in the traffic stimuli. This may not be realistic. For
instance, the motivation to decelerate as the separation distance
becomes small is based on safety (collision avoidance), and it is more
forceful than the motivation to accelerate as the separation distance
becomes larger than desired.
292 Yaser E. Hawas

4. One limiting assumption in some of the models is the zero

deceleration/acceleration rates of the following vehicle if the lead
vehicles have the same speed regardless of their spacing.
5. The assumption that the following vehicle drivers can accurately
perceive the relative speed of the lead vehicles, absolute speed and/or
acceleration of lead vehicle at any point in time is quite limiting, and
may not be realistic given the different driver reactions with factors
such as aging, impairment, etc.

The recent evolution of advanced traffic management systems stimu-

lated the traffic research community to revisit the development of
the simulation models and assess their applicability vis-à-vis the
functionalities that such advanced systems support. Hawas and
Mahmassani (1997) discussed the data needs for development,
calibration and validation of dynamic simulation models and algo-
rithms for ATIS/ATMS applications. Shin et al. (1999) stressed the
immediate need for a tool to evaluate and validate the recent traffic
ATIS/ATMS developments. Hawas (2002) proposed a sensitivity
analysis, heuristic approach for calibrating macroscopic simulation
models.
This paper reviews a simulation-based integrated system for incident
management ( ) to support the real-time operation of the TMCs. An
integral part of is a dynamic traffic simulator capable of modeling
traffic operation and control in urban network corridors (freeways and
networks). This paper focuses on the simulation model description. The
following section introduces the modules of the envisaged integrated
system for incident modeling ( ). The car-following model is then
presented. The paper also discusses the rationale used in modeling the
lane switching mechanism. The validation of the devised car-following
model is discussed and results are analyzed. Finally, samples of the
accident replicated using the simulation model are presented and
discussed.

Integrated System for Incident Modeling:

Existing simulation models do not offer adequate capabilities of
modeling incidents in urban networks accurately. For instance, the
well-known NETSIM does not allow for the modeling of accidents at
specific locations. Incidents are represented as long-term events which
are generated randomly along the specified approach. The last section
of the paper discusses in detail the limitations of the NETSIM model
from the standpoint of incident modeling. The system comprises three
modules:
 A Microscopic Simulation Model for Incident Modeling 293

1. Traffic Simulation Module ( -s). This module provides strategies

for highway clearance, route diversion and signal control in response
to identified incident alarms. It is capable of modeling the driver’s
behavior in response to route diversion advisory messages dissemi-
nated by the TMCs. The -s simulation model is an integral part
of the overall traffic management module, and is used to model the
network entities, control, driver behavior, etc. This paper will focus
on the -s module.
2. Fuzzy Logic Incident Detection/Verification Module ( -fl). This
module is developed/calibrated using simulation-based detector data
extracted from -s. It uses an artificial neural network approach
together with detector data to calibrate a fuzzy logic approximation
for incident detection (Hawas, 2007).
3. Motorist Information Module ( -mi). This module controls the
disseminated information to motorists in response to the traffic
management decisions made by -s.

This paper focuses on the detail of the car-following and lane-changing

procedures of the simulation model (the first module). -s is a visual
basic microscopic simulation for incident modeling and management. It
offers extended capabilities in modeling incident/accident situations
quite accurately, and, as such, extract incident patterns via the various
approach detectors. The time-dependent accident-related detector data
at intersections can be used to calibrate incident detection logic (as was
the case in developing the fuzzy logic for incident detection, -fl).
-s is an object-based model that allows for the distribution of
various detectors (on each lane) at different locations. It can be used to
model different intersection layouts, traffic control types and timing, as
well as link characteristics (speed, number of lanes, etc.).
-s was initially developed using Visual Basic 6.0 as a program-
ming language. Visual Basic was selected to enable the integration with
the fuzzy logic code for incident detection. Furthermore, Visual Basic
provides a simple programming syntax with the ability to program
using object oriented programming techniques. Currently, the program
is being developed using Java to support internet operation.
-s is a microscopic, object-oriented program. Among the
program objects are streets, lanes, detectors, vehicles, signals and
intersections. Each object is composed of data that represents the
current values of object parameters, and methods (or functions) which
could be applied to the object (e.g. add vehicle or remove vehicle for the
lane object). -s is a hierarchical model in the sense that main or
larger objects contain the sub (or smaller) objects. For example, each
street object composes several lane objects, and each of these lane
objects contains the vehicle objects (running on that lane).
294 Yaser E. Hawas

Car-Following Model
The -s model assumes a string of three vehicles (two lead vehicles
and one following vehicle) traveling along a single lane. The accelera-
tion/deceleration rate, speed and spacing of the following vehicle are
calculated based on the behavior of the downstream lead vehicle(s).
The -s car-following model is coupled with the so-called finite
state model that captures the lane switching behavior of the driver. The
finite state model ‘estimates’ the scores of the lanes that may serve the
vehicle movement, and allocates the vehicle to the lane offering better
score. The lane score is estimated empirically based on the number of
queued vehicles and accident blockage (if any). The newly switched
vehicle among lanes is restricted from further switching for a time
interval (5
10 seconds) to limit the numbers of unnecessary switching.
The model assumes that the acceleration/deceleration rate of the
following vehicle driver depends on the current speed, the control
speed, and the perception-reaction time of the following vehicle driver.
The control speed is defined as the maximum speed at which the
following vehicle driver would travel given the spacing and rate of
change in spacing with respect to the lead vehicle immediately in front
(i.e. second lead vehicle). It assumes also that the following vehicle
adopts a desired spacing between itself and the lead vehicle(s). The
length of the desired spacing is assumed to depend on the current speed
of the following vehicle.
Furthermore, it assumes the level of alertness of a driver affects the
perception/reaction time component of the acceleration/deceleration
rate. The following vehicle driver perception/reaction time depends on
his/her level of alertness. The level of alertness depends on current
spacing, desired spacing between the following vehicle and lead
vehicle(s), and the status of the lead vehicle(s) (e.g. stopping).
The data to calibrate car-following models are quite difficult to
collect, not to mention the extensive recourses needed for even
restricted data sets (Hawas & Mahmassani, 1997). More detail on
the data requirements and procedures for calibration and verification of
dynamic simulation models can be found in Hawas (2003). An
empirical formula was derived to estimate the control speed for steady
conditions (zero speed difference between the following and the lead
t
vehicles) for various values of spacing (SF,2 ) and vehicle speed limits
t
(VSL). For any vehicle speed limit (VSL), and spacing (SF,2 ), the vehicle
t
control speed at the steady conditions (VCSS ) is given by the following
formulae:
SF;2t
45 m VCSSt  VSL
SF;2t
 9 and B 45 m VCSSt  (2:90:54+VSL) ln (SF;2t )(11:91:04 VSL) (1)
SF;2t B 9 m VCSSt  0
 A Microscopic Simulation Model for Incident Modeling 295

The above equations were derived for steady flow conditions (when the
speed difference between the following and the second lead vehicle is
t t
1
zero, and the spacing between them is kept fixed; i.e. SF, 2 SF, 2 ). A
t
parameter a is introduced herein to account for the general flow
conditions (not necessarily steady conditions) as follows:
VCSt  at +VCSSt (2)
where, at is a parameter ranging from zero to one, and VCSt is the
vehicle control speed at time t. at is introduced to capture the variance
in the following driver’s alertness. The driver reduces his/her desired
speed due to increased crash risk, when the spacing between him/her
t t
1 t
and the second lead vehicle is closing (i.e. when SF, 2 BSF, 2 ). a is
assumed inversely proportional to the normalized rate of change in
spacing between the following vehicle and the second lead vehicle,
t t
NRSF, 2. If the spacing is closing, NRSF, 2 will yield only negative values
(ranging from 0 to 1)
at  exp(m NRSF;2t ) if 0
 NRSF;2t
1
(3)
 1:0 Otherwise
NRSF;2t  SF;2t SF;2t
1 =SF;2t
1 (4)
m is a sensitivity parameter introduced to capture the effect of the
normalized rate of change in spacing on at . Ideally, the level of
alertness, at and as such the m value should be dependent on the
NRSFt ,2. However, for simplicity, m is assumed fixed and can be
estimated using calibration sensitivity analysis (Hawas, 2002).
The traffic measures (acceleration, speed, spacing, etc.) are updated
each time interval with length of Dt (the simulation resolution).
The speed of the following vehicle is estimated using the following
equation:
VFt  VFt1 AFt1 Dt (5)
where AFt
1is the acceleration of the following vehicle at time t 1
. AFt
is calculated as follows:
AFt  (VCSFt VFt )=PRT (6)
PRT denotes the perception and reaction time needed to increase/
decrease the speed from the current value to the desired value. The
typical value of PRT varies based on the driver’s perception and
reaction time, as well as attentiveness, which, in turn, varies based on
the spacing with the lead vehicles as compared to the desired spacing.
Furthermore, PRT depends on the status of the lead vehicles; the
following driver’s alertness increases if the status of the lead vehicle is
‘stopping’.
296 Yaser E. Hawas

PRT  NPRT=(ALF;1t +ALF;2t ) (7)

NPRT denotes the normal perception and reaction time to achieve the
desired speed from the current value (ideally should be treated as a time
variant variable, but for simplicity assumed fixed of 2 seconds in -s).
t
ALF,1 represents the level of alertness of the following vehicle due to
the first lead vehicle. It is calculated using Eq. (8) if the status of the first
t
lead vehicle is stopping, and 1 otherwise. ALF, 2 represents the level of
alertness of the following vehicle due to the second lead vehicle. It is
calculated using Eq. (9) if the status of the second lead vehicle is
stopping, and 1 otherwise.
t t
The two parameters ALF, 1, ALF, 2 are introduced to capture the
following vehicle’s driver added alertness due to the actual spacing
compared to the desired spacing between him/her and the lead vehicles.
They are functions of the spacing ratio (the actual spacing to the desired
spacing). When the actual spacing is equal to the desired spacing, the
values are equal to 1 (the driver’s least attentive condition). As the actual
spacing becomes lesser than the desired spacing, the attentiveness level
increases, with the value equals 2 (with such case valid only theoreti-
cally) as the actual spacing becomes 0. Figure 1 illustrates the assumed
relationship (Eqs. 8 and 9) between the alertness level and the spacing
ratio. The parameters of this relationship were calibrated using field
data observations (Mehmood et al., 2003). The desired spacing between
t
the following and the second lead vehicle (DSF, 2) is also assumed
to follow an exponential form of the speed of the following vehicle
(Eq. 10); the higher the speed, the higher the desired spacing. Figure 2
illustrates the calibrated exponential relationship between the desired
spacing and the following vehicle speed. The desired spacing between
t
the following and the first lead vehicle (DSF, 1) is set approximately equal
to double the desired spacing between the following and the second lead

2.5

2
Alertness level

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Spacing ratio (Actual to desired spacings)

Figure 1. Empirical relationship between driver alertness level and spacing ratio
 A Microscopic Simulation Model for Incident Modeling 297

60

Desired spacing
50
40
30
20
10
0
1 21 41 61 81 101
Following vehicle speed

Figure 2. Empirical relationship between desired spacing and following vehicle speed
t
vehicle (DSF, 2) plus the length of the second lead vehicle (L2).

ALF;1t  2+exp(0:7+SRF;1t ); SRF;1t  SF;1t =DSF;1t (8)

ALF;2t  2+exp(0:7+SRF;2t ); SRF;2t  SF;2t =DSF;2t (9)
DSF;2t  7:00+exp(0:016+VFt ) (10)
DSF;1t  2DSF;2t L2 (11)

Lane-changing Model
The lane car-following model does not account for the possibility of
vehicles switching lanes when current speeds become less than the
desired speed, and, as such, vehicles do not actually follow the lead
vehicles as explained earlier. The lane switching, on the other hand,
follows a fairly complicated decision process, where the driver
examines the availability of ‘free’ lanes to accommodate his/her
movement. -s augments the car-following model with the lane-
changing model as explained below.
Since -s is an object-oriented model, the ‘vehicle’ object has the
so-called ‘status’ property that defines the physical status of each
vehicle. Five status properties are defined for each vehicle: moving,
stopping (if the vehicle is decelerating), maneuvering (if the vehicle is
to switch lanes), exiting (if vehicle is to exit current link) or accident
(if vehicle is stopped to simulate an incident). The vehicle status is
determined based on the vehicle actual speed as compared to the
desired speed, leading vehicle(s) speed, spacing, etc.
-s follows an empirical method known as ‘lane scoring’ in
specifying the candidate lane(s) for the vehicle willing to switch lanes. If
the vehicle status is set to ‘maneuvering’, indicating the willingness of
the vehicle to switch lanes, the lane scoring class module is called to
estimate the candidate lane scores. The candidate lanes are these
adjacent lanes (to the current one) that could accommodate the vehicle
intended movement (left, through, or right) at the downstream node of
298 Yaser E. Hawas

its current link. The lane score is heuristically set to be a function of the
current lane’s vehicle count, and the number of active incidents on the
lane (if any). An empirical formula is used for the lane score as follows:

Lane score u1 +Active IncidentsLane Vehicle Count (12)

where u1 is a penalty term for the incidents. Higher values of u1 result

in very low score for the lane, and, as such, the lane would be entirely
avoided. On the other hand, low values will not ‘flag’ the lane as an
incident lane, and, as such, excessive delays might be encountered if
such lane is penalized properly. The ‘best’ penalty value depends on the
average lane volume. Very high volume would necessitate higher
penalty terms to avoid the excessive delays, whereas for low volume
lanes, a u1 value of one might be most appropriate. Sensitivity analysis
was used to find the most suitable value of u1. A value of three was
found to give very reasonable results for a general network; that is, each
incident is approximately equivalent to about three vehicles.
The candidate lane with the least score is flagged as best candidate
for potential switching. However, the maneuvering vehicle does
not actually switch to the lane until a suitable gap is found. The
vehicle switches only if the potential lane provides enough space
for maneuvering (the adequate gap is a function of the speed of
the following vehicle in the potential lane, the current speed of the
maneuvering vehicle, and the spacing between the maneuvering vehicle
and the lead vehicle on the potential lane). Once an adequate gap is
allotted, the maneuvering vehicle switches to the potential lane.

Comparative Analysis of NETSIM and -s

Validating simulation models in general, and microscopic models in
particular, is a challenging task due to the limited availability of
validation data (Hawas, 2002), the complexity of the car-following and
the lane-changing process, and the interaction among the car following/
lane-changing parameters noted earlier.
The approach adopted in this paper to validate -s is simulation-
based. The well-known NETSIM simulation model was utilized to
generate several validation scenarios for comparative analysis vis-à-vis
-s. This paper includes selected studied cases. NETSIM operates
with a car-following model, and has the ability to model detectors and
offers good flexibility in specifying the detector’s locations and the
frequency of readings. This feature was utilized together with the
counter feature in -s to compare the spatial and temporal
distribution of vehicles along the links with different traffic intensities.
 A Microscopic Simulation Model for Incident Modeling 299

The NETSIM model, although capable of simulating incident (as

long term events) in urban street networks, has several limitations,
among which:
1. NETSIM cannot model an incident at a specific point on the
network. Incidents are modeled as long term events on selected links.
The specific locations (on the links) of these events are selected
randomly by the model. Furthermore, the incident can cause a
blockage to one lane only; simultaneous lane blockage cannot be
modeled.
2. NETSIM does not account for simultaneous incidents/accidents on
the network. One incident has to finish before another one starts.
3. The car-following model of NETSIM does not account for the
changes in driving behavior as a result of an accident. For example,
the slowing down of the vehicles as they pass by the accident
location to have a look.

A three-stage approach is suggested to validate -s as follows:

1. Comparative analysis of the -s and NETSIM detectors’ readings
under recurrent traffic conditions (as briefly presented in this paper).
Comparative analysis of the two models’ detectors in conjunction
with ‘limited’ field data for recurrent conditions.
2. Comparative analysis of the two models’ detectors under non-
recurrent traffic conditions. However, due to the limitations of the
NETSIM model in modeling incidents, the following procedure is
devised:
a. use NETSIM to model long term events (incidents) and extract
the randomly selected incident location, and associated detector
readings for such cases;
b. use -s to replicate the same incidents of NETSIM and
extract the corresponding detector readings;
c. conduct comparative analysis of the detector readings in (a)
and (b).

In this paper, some results of the first stage of validation are presented.
More work on the validation of the -s model is underway and will
be published elsewhere.
A simple four-leg intersection with three-lane approaches is utilized
for validation. An approach of 600 m in length is then selected, and two
detectors are installed on each lane; one close to the downstream node
(10 m away from the stop line to verify the queue discharge modeling),
and the other at the middle of the lane (300 m from the down stream
node). The approach is represented using the two models, and the
passage detector readings are recorded every 10 seconds. The detector
300 Yaser E. Hawas

readings were then accumulated over a period of 15 minutes. The

approach was simulated for one hour and the cumulative detector
readings were reported at 15 minutes intervals (15, 30, 45, and
60 minutes). The cumulative detector reading is shown in Table 1 for
the two models. The results shown in Figure 3 clearly demonstrate the
very close spatial and temporal vehicle distribution between the two
models. The maximum discrepancy among the two models is about
8.5%. The figure indicates that higher discrepancy is depicted between
the downstream detectors (closer to the intersection). This could be
attributed to the difference in modeling the queuing build/dissipation,
headway setting between the two models.

Accident Replication Using -s

-s was used to develop various accident scenarios. Several
accident replicates were generated using the -s of a simple
network composed of a single three-lane approach operated with a
pre-timed signal. Accidents were introduced to the middle lane of the
approach at various locations as specified below. The detector counts
were recorded every 10 seconds. The total simulation time was set to
10 minutes (warming up time) plus the accident duration plus
10 minutes after the clearance of the accident. For an accident
duration of 5 minutes, the simulation time was 25 minutes, with a
total of 150 data records, with each data record representing the
detector cumulative hits at 10-second intervals. The variables
considered in generating the various accident scenarios are as
follows:
1. Link length: three various link lengths were considered (150, 300,
and 500 m).
2. Downstream green time: two values were considered for the down-
stream node green time (15 and 30 seconds).
3. Middle detector location: in all the accident scenarios, three
detectors were considered for each lane. This arrangement was
found to be the most economical yet consistently reliable and
accurate arrangement. The first detector was located at the begin-
ning of the approach (10 m away from the downstream node). The
third detector was located at the end of the approach (at the
upstream node). Only the second (middle) detector was placed at
different locations (to study its location effect on the accuracy of
data). Three positions were considered for the middle detector
location: 1/5 of the street length, 1/4 of the street length, and 1/3 of
the detector length (measured from the downstream node).
 Table 1. Comparative detector readings of NETSIM and -s for validation
Upstream detectors Downstream detectors

A Microscopic Simulation Model for Incident Modeling 301

NETSIM -s Diff% NETSIM -s Diff%

Case I: Highway speed limit 60, downstream signal green time 30 seconds, cycle156 seconds, approach volume1350
vehicle/hour
Cumulative counts (15 minutes)
Lane 1 113 112 0.89 87 89 2.25
Lane 2 109 111 1.8 86 82 4.88
Lane 3 114 112 1.79 90 88 2.27

Cumulative counts (30 minutes)

Lane 1 224 222 0.9 188 182 3.3
Lane 2 224 223 0.45 186 178 4.49
Lane 3 222 223 0.45 193 194 0.52

Cumulative counts (45 minutes)

Lane 1 331 334 0.9 289 276 4.71
Lane 2 322 319 0.94 279 284 1.76
Lane 3 326 334 2.4 285 299 4.68

Cumulative counts (60 minutes)

Lane 1 410 418 1.91 390 380 2.63
Lane 2 407 399 2.01 379 376 0.8
Lane 3 423 418 1.2 390 406 3.94
 302 Yaser E. Hawas
Table 1 (Continued )

Upstream detectors Downstream detectors

NETSIM -s Diff% NETSIM -s Diff%

Case II: Highway speed limit 80, downstream signal green time30 seconds, cycle 156 seconds, approach volume 1000
vehicle/hour
Cumulative counts (15 minutes)
Lane 1 83 83 0 76 77 1.3
Lane 2 84 84 0 77 71 8.45
Lane 3 82 80 2.5 78 78 0

Cumulative counts (30 minutes)

Lane 1 167 166 0.6 161 162 0.62
Lane 2 169 167 1.2 166 154 7.79
Lane 3 162 166 2.41 159 171 7.02

Cumulative counts (45 minutes)

Lane 1 252 249 1.2 251 255 1.57
Lane 2 251 251 0 245 242 1.24
Lane 3 246 249 1.2 256 250 2.4

Cumulative counts (60 minutes)

Lane 1 306 308 0.65 313 301 3.99
Lane 2 306 300 2 298 309 3.56
Lane 3 304 309 1.62 308 312 1.28
Table 1 (Continued )

Upstream detectors Downstream detectors

A Microscopic Simulation Model for Incident Modeling 303

NETSIM -s Diff% NETSIM -s Diff%

Case III: Highway speed limit100, downstream signal green time30 seconds, cycle 156 seconds, approach volume 1170
vehicle/hour
Cumulative counts (15 minutes)
Lane 1 96 96 0 86 88 2.27
Lane 2 99 97 2.06 82 85 3.53
Lane 3 97 98 1.02 89 84 5.95

Cumulative counts (30 minutes)

Lane 1 192 194 1.03 189 185 2.16
Lane 2 201 194 3.61 185 179 3.35
Lane 3 191 195 2.05 186 186 0

Cumulative counts (45 minutes)

Lane 1 286 291 1.72 282 283 0.35
Lane 2 301 292 3.08 284 273 4.03
Lane 3 289 291 0.69 285 279 2.15

Cumulative counts (60 minutes)

Lane 1 384 387 0.78 385 387 0.52
Lane 2 400 390 2.56 380 374 1.6
Lane 3 383 390 1.79 383 380 0.79
304 Yaser E. Hawas

Figure 3. Percentile difference between NETSIM and -s cumulative detector

readings

4. Accident location: two locations were considered: 1/2 of the street

length and 1/6 of the street length (measured from the downstream
node).
 A Microscopic Simulation Model for Incident Modeling 305

5. Accident duration: two durations were considered: 5 and 10 minutes

(300 and 600 seconds).
6. Traffic volume: four levels of traffic volumes were considered: 200,
500, 1000 and 1500 vehicles/hour/approach. These levels represent
light to congested traffic conditions.

Other variables were kept fixed as follows:

1. Vehicle type percentages for cars, buses, vans, small and big trucks
were 25, 30, 25, 20 and 0%, respectively.
2. Percentages for left, through and right turns were fixed at 25, 55 and
20%, respectively.
3. Other variables, such as the various vehicle type lengths and the
maximum allowed speed for each vehicle type, were also kept fixed.

Figure 4 illustrates a sample of the generated accident patterns. The

figure indicates the cumulative detector readings corresponding to
various accident scenarios as indicated on the figures. Herein, three
detectors were considered to adequately capture the accident pattern:
upstream detector, D1; middle link detector, D2; and downstream
detector, D3. The upstream detector was located at the link iterance,
and D3 was located 10 m from the link’s stop line. The middle
detector’s location was varied (as indicated above) to identify the most
suitable location.
It is clear from the figures that, as the volume increases, the spatial
separation among the detectors’ patterns increases. At very low traffic
volumes, the three detector patterns become almost identical. The
visual inspection of these patterns indicates the big variation of the
accidents patterns for the different considered variables. Some patterns
can be easily depicted by the flat slope of the accumulative detector’s
reading curve. Others, however, do not exhibit the same curve shape.
This necessitates the need for a tool to ‘learn’ and ‘automate’ the
process of learning these patterns; a tool that can associate the various
pattern shapes with the various input variables. The generated detector
data records were utilized to develop the fuzzy logic module for
incident detection at intersections, -fl (Hawas, 2007).

Concluding Remarks
This paper has provided an overview of a simulation-based integrated
system for incident management ( ) to support the real-time
operation of the TMCs, with emphasis on the microscopic simulator
-s. The paper introduced the modules of the envisaged integrated
system for incident modeling ( ) and the mathematical formulation
 306 Yaser E. Hawas
Accident Pattern (Flow = 1500, Approach Length = 500 m, Detector 2 Location = 343m, Accident Pattern (Flow = 1000, Approach Length = 500 m, Detector 2 Location = 334m,
Accident Location = 250 m, Accident Duration = 600 sec, Green Period = 15 sec) Accident Location = 416 m, Accident Duration = 300 sec, Green Period = 15 sec)
120 100
Accident Duration
90
100 Accident Duration
80
D1_Count 70 D1_Count
Detector Counts

Detector Counts
80 D2_Count D2_Count
D3_Count 60 D3_Count
ACC_Status ACC_Status
60 50

40
40
30

20
20
10

0 0
1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 113 120 127 134 141 148 1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 113 120 127 134 141 148
Time Interval (each of 10 seconds) Time Interval (each of 10 seconds)

Accident Pattern (Flow = 200, Approach Length = 500 m, Detector 2 Location = 400m, Accident Pattern (Flow = 200, Approach Length = 500 m, Detector 2 Location = 400m,
Accident Location = 416 m, Accident Duration = 600 sec, Green Period = 30 sec) Accident Location = 250 m, Accident Duration = 600 sec, Green Period = 30 sec)
30 35

25 30
Accident Duration Accident Duration
D1_Count 25 D1_Count
Detector Counts

Detector Counts
20 D2_Count D2_Count
D3_Count 20 D3_Count

15 ACC_Status ACC_Status

15

10
10

5
5

0 0
1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 113 120 127 134 141 148 1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 113 120 127 134 141 148
Time Interval (each of 10 seconds) Time Interval (each of 10 seconds)

Figure 4. Samples of -s accident patterns

 A Microscopic Simulation Model for Incident Modeling 307

and algorithmic procedure of the car-following module were presented.

The paper also discussed the rationale and procedures used in modeling
the lane switching mechanism, signal control (pre-timed and respon-
sive), and the shortest path finding algorithm.
The car-following logic can be used to model the spatial/temporal
distribution of traffic flows along the network links under recurrent
and recurrent traffic conditions. As such, it can be used to develop
spatial/temporal flow patterns that could be deployed to detect
incidents at congested intersections.
One of the valuable contributions of the devised simulation model is
its ability to model incidents accurately at pre-specified spatial and
temporal values. With reference to the incident patterns shown in
Figure 4, the incidents can be clearly defined and captured by the
detectors along the approaches. It is important, however, to stress the
fact that the incident patterns are quite varied, and, as such, would
require a smart mechanism to detect these incidents effectively. More
discussion on this issue can be found in Hawas (2007).
This paper introduced a novel approach to the validation process of
microscopic simulation models. Most of the validation efforts of
microscopic simulation models were to achieve agreement between
‘aggregate’ observed values and model outputs on variables such as
volume served, average travel time, average speed, density and average
and maximum vehicle queue length. The validation based on averages
of traffic variables, however, has several shortcomings. For instance,
one could demonstrate that for the same aggregate measure (e.g.
average link travel time), the spatial and temporal distribution might
vary significantly. It has been a real challenge for researchers and
practitioners validating simulation models as the data needed are
mostly unavailable and quite expensive to obtain. It is recommended
that a two-stage approach for the validation of the newly developed
simulation models in general should be adopted, and the one presented
herein in particular.
The first (upper) stage would validate a benchmark simulator (e.g.
NETSIM) using field data. The validation should be undertaken using
both aggregate measures (or counts) from the field as well as real
detector data. This benchmark simulator can then be used to validate
any special purpose simulators, such as the one presented here. Usually,
this validation stage is most expensive and cumbersome as it would
require extensive field data collection, manpower, cost, etc. Further-
more, the validation would only be applicable to the investigated sites
or traffic conditions. Due to the huge resources that would be needed to
validate the benchmark simulator, it is advisable that this task is
maintained by an authority or well-established traffic agency over
308 Yaser E. Hawas

extended time periods, and then made accessible to other simulation

developers.
It is also understood that the real-world validation of such a
benchmark simulator might extend for prolonged periods to collect
the data that might be necessary for validation. For instance, validating
the ‘rare’ incident situations with real data would require long periods
of observations and extensive detector installations in the network,
which might not be available or affordable for researchers. This
supports the recommendation of devising systematic methods of data
collection to be authorized by some traffic agency.
The benchmark simulator (and its library of data) can then be used to
validate the developed simulation models, by comparing the detailed
microscopic traffic patterns extracted from the devised simulator
(herein, -s) vis-à-vis those extracted by the benchmark simulator
(NETSIM). This is particularly done by comparing the spatial-temporal
distribution of the vehicles along links and accounting for various
factors, such as traffic volume, link speeds, signal timing, etc. This
could be undertaken extensively and thoroughly, to cover a wide range
of traffic conditions and scenarios.

Acknowledgements
The author expresses his sincere appreciation to the Research Affairs at the United
Arab Emirates University for the financial support of this project under fund grants No.
01-01-7-11/03, 02-01-7-11/04, and 06-01-07-11/05. The author also acknowledges the
efforts of Alaa Al-Alawi and Moumena Cheqfeh for their assistance in programming.

References
Brackstone, M. & McDonald, M. (1999) Car-following: A historical review, Transportation
Research Part F , 2, pp. 181
196.
Davis, G. L. & Nihan, N. L. (1984) Using time-series designs to estimate changes in freeway level
of service, despite missing data, Transportation Research , 18A(5), pp. 431
438.
Edie, L. C. (1961) Car following and steady state theory for non-congested traffic, Operations
Research , 9, pp. 66
76.
Greenberg, H. (1959) An analysis of traffic flow, Operations Research , 7, pp. 78
85.
Greenshield, B. D. (1935) A study of traffic capacity, Highway Research Board Proceedings , 14,
pp. 448
447.
Hall, F. L. (1996) Traffic stream characteristics, in: Gartner, Messer & Rathi (Eds) Chapter 2 of
the Traffic Flow Theory Monograph (Washington, DC Federal Highway Administration).
Hawas, Y. E. (2002) Calibrating simulation models for ATIS/ATMS applications, Journal of
Transportation Engineering , 128(1), pp. 80
88.
Hawas, Y. E. (2007) A fuzzy-based system for incident detection in urban street networks,
Transportation Research Part C: Emerging Technology doi: 10.1016/[Link].2007.02.001.
 A Microscopic Simulation Model for Incident Modeling 309
Hawas, Y. E. & Mahmassani, H. S. (1997) Data requirements for development, calibration of
dynamic traffic models and algorithms for ATMS/ATIS, in: Proceedings of 78th Annual
Meeting of the Transportation Research Board , Washington, DC.
Mehmood, A., Hellinga, B. & Saccomanno, F. (2003) Modeling car following using system
dynamics, in: Proceedings of the 82nd Transportation Research Board Annual Meeting ,
Washington, DC.
Shin, M. S., Ran, B., Oh, R. S. & Choi, K. (1999) Developing an evaluation tool for a real-time
DTA system, in: Proceedings of the 78th Annual Meeting of the Transportation Research
Board , Washington, DC.
Tavana, H. & Mahmassani, S. H. (2000) Estimation and application of dynamic speed-density
relations using transfer function models, in: Proceedings of the 79th Transportation Research
Board Annual Meeting , Washington, DC.
TSIS (1998) User Guide; release 4.2. Federal Highway Administration (Colorado Spring, CO: ITT
Systems & Sciences Corporation).