A Fuzzy Logic Model for Real-time Incident Detection in Urban Road

Network

Faisal Ahmed and Yaser E. Hawas

Department of Civil and Environmental Engineering, COE, United Arab Emirates (UAE) University, Al Ain, U.A.E.

Keywords: Fuzzy Logic and Systems, Intelligent Transport System, Urban Incident Detection, Neuro-Fuzzy, Detector

Count, Average Speed, Detection Rate, False Alarm Rate.

Abstract: Incident detection systems for the urban traffic network are still lacking efficient algorithms or models for

better performance. This paper presents a new urban incident detection system based on the application of

Fuzzy Logic modeling. Offline urban incident and corresponding non-incident scenarios are generated using

a microscopic simulation model assuming varying traffic link flows, phase timing, cycle times, and link

lengths. The traffic measures are extracted from three detectors on each link. Statistical significance analysis

was utilized to identify the significant input variables to be used in developing the Neuro-fuzzy model. A set

of data was generated and used for training of the proposed Neuro-fuzzy model, while another set was used

for validation. The performance of the proposed model is assessed using the success and the false alarm

rates of detecting an incident at a specific cycle time.

1 INTRODUCTION

The loop detector-based freeway incident detection

algorithms in literature could be generally

categorized into adopted analytical and heuristic-

based techniques (Parkany, 2005). Notable roadway

detector-based recent urban incident detection

models are mostly based on statistical regression

(Ahmed and Hawas, 2012), Bayesian network

(Zhang and Taylor, 2006) and fuzzy logic modeling

(Hawas, 2007) techniques. Non-parametric

optimization technique (Liu et al., 2007) and

discriminant analysis (Sermons and Koppelman,

1996) was used for the probe-vehicle based urban

incident detection system. Neural network models

were also developed (Dia and Thomas, 2011) using

both loop detector and probe-vehicle data.

Typically, the focus of these algorithms was

primarily on estimating the performance measures

using the percentage of the total number of incidents

detected or falsely identified incidents for the

simulated duration where the whole incident as a

single unit. These algorithms do not particularly

account for the true start or the terminating times of

individual incidents as a criterion of evaluation.

Moreover, these do not consider the effects of the

link lengths of the approaches, the hourly traffic

volumes, the signal settings and the cycle times of

the intersections. This study strives to fill in some of

these research gaps of urban incident detection areas

for more efficient detection model.

This study assumes that the duration of an

incident is divided into smaller time steps and the

algorithm is operated repeatedly each (shorter time

resolution) step to detect incidents. The proposed

fuzzy-model is capable of identifying whether there

is an incident or not during each time step. The

simulation period may be divided to hundreds of

such shorter time steps. With this approach the

actual incident start and clearance time could be

identified to a great extent.

Therefore, this paper comes up with a new form

of urban incident detection model using fuzzy-logic.

The model detects the incident status each time step,

under various signal cycle times, link lengths and

traffic volumes combinations.

2 METHODOLOGY

The conceptual assumption is that the average

detectors’ readings in the case of incident may

significantly vary from the counter readings in the

case of no incident. A micro-simulation based

methodology is adopted. A typical pre-timed urban

intersection network that consists of four links of

465

Ahmed F. and E. Hawas Y..

A Fuzzy Logic Model for Real-time Incident Detection in Urban Road Network.

DOI: 10.5220/0004239904650472

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 465-472

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

similar geometry and traffic conditions (Figure 1)

was selected as it represents the simplest case of a

signalized urban network. The overall methodology

that this study followed could be summarized as

with the following steps:

Step 1: Preparing a specific simulation test-bed (with

upstream, midblock and downstream detectors) for

the base inputs of a specific Cycle Time (sec) of

downstream signal, associated link length (m) and

hourly traffic volume (veh/hr) combination.

Step 2: Run this specific simulation model without

incident and extract the raw detectors count and

speed data (at every approach split time of the signal

cycle).

Step 3: Run this specific simulation model test-bed

again with an incident generated at a specific time

with specific incident duration, and hence extract the

raw detectors count and speed data.

Step 4: Estimate detectors count and speed data for

every analysis time step [analysis time step = cycle

time] for the both incident-free and incident-induced

runs.

Step 5: Estimate the traffic measures of interest (i.e.

independent variables) for every analysis time step.

Step 6: Repeat steps 1 to 5 for a different base input

specifics (cycle time, link length and hourly traffic

volume) and collect all the traffic measures of

interests for all analysis time-steps from all input

specific combinations.

Step 7: Develop some statistical significance tests

(ANOVA) for extracting the most significant

independent variables to be used in the proposed

incident detection model.

Step 8: Develop a Neuro-Fuzzy Model [A fuzzy

model for each specific base model that is trained

with neural net for calibrations.

Step 9: Validate this Neuro-Fuzzy Model with

validation data set. This data set is developed

following step 1 to step 5 for slightly altered input

specifics from the base cases.

Step 10: Comparison of the measures of

performance estimated by the calibrated models

using the validation data.

2.1 Experimental Set up of the Incident

Modeling

An incident is modeled here as a “lane-blocking”

event that persists at least for 6 minutes on a typical

three-lane urban arterials in the simulation models.

It is quite rational that longer time incidents,

reported in previous studies, could be detected easily

as these might have some significant impacts on the

traffic parameters. The true challenge (and that is

one of the premises of this work) is to detect the

incidents of relatively shorter times. Herein, we

focus on a single-lane blocking incidents of 6, 8 and

10 minutes incident clearance intervals. Such shorter

events will be harder to detect.

Figure 1: A simple signalized (pre-timed) urban

intersection: four approaches, detector placements and a

randomly generated incident on the Lane 1 (i.e. rightmost

lane in the direction of traffic flow) of the West bound

approach.

2.2 Incident Data Development

In the absence of detailed data of real-field detector-

based traffic measures, it is a common practice to

use well-validated simulation data to generate

incident scenarios. Previous studies ((Khan and

Ritchie, 1998); (Yuan and Cheu, 2003) and (Zhang

and Taylor, 2006)) also used simulation models to

generate the incident scenarios. However, these

studies used calibrated simulation models from the

field data. Similarly, this study also adopted

NETSIM to generate incident data. NETSIM places

the incidents randomly on the designated lane with

specific time duration; however, it cannot be

actually used to model an incident at a very specific

designated place (Yuan and Cheu, 2003).

2.3 Incident Data Analysis and Fuzzy

Models Development

The detector data were extracted for both incident

and non-incident simulation cases for various

operating configurations models. The term

‘operating configuration’ refers the combination of a

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

466

specific cycle time, link length and traffic volume.

Specific traffic measures that are likely to vary

between incident and no incident cases were chosen

to develop the fuzzy-logic models. Incident

detection and false alarm rates were chosen as the

measures of effectiveness (MOEs) of the calibrated

fuzzy-logic models.

3 INCIDENT MODELING

For practicality issues herein, we assume that each

detector covers all the approaching lanes for

capturing the traffic data. Each detector was placed

perpendicularly to the direction of traffic flow. The

same logic could be easily adapted in case the

detectors are placed on individual lanes. When a

vehicle hits a detector, the corresponding detector’s

count is increased by one. The detector also captures

the vehicle’s speed.

Only for the simplicity and convenience of the

data extraction from the detectors, it is assumed that

incidents starting time is the start of the green phase

of the incident approach. The incident then lasts for

multiples of cycle times (based on the incident

duration). The incident terminates concurrently by

the end of a cycle time. However, this assumption

might have some impact on the time to detect of the

incidents.

The detector placements are kept fixed; near the

stop-line (downstream detector), at mid-block

position (mid-detector) and at end of the link

(upstream detector). The vehicle composition is kept

also fixed; private-cars 90% and heavy-vehicles

10%. The percentages for left, through and right

turns at each approach were fixed as 25%, 50%, and

25%, respectively. The operating speed limit was

fixed at 60 km/hr. The pre-timed signal operates on

split phase sequencing for the 4 approach legs.

The simulation test beds were varied to reflect

various signal cycle time (60, 80 or 100 seconds),

approach link length (300, 500 or 1000m) and

hourly traffic volumes (100, 500, 1000 or 1500

veh/hr). As the combination of link length of 300

and traffic volume of 1500 veh/hr resulted in link

spill back in the no-incident scenarios, and as such it

was excluded. We have 11 basic link and volume

(LV) combinations for each signal cycle to develop

simulation test beds and to extract the detector data

needed for model calibration. Thus, the LV

combinations, denoted by (Link length: Veh/hr), are

(300, 100), (300, 500), (300, 1000), (500, 100),

(500, 500), (500, 1000), (500, 1500), (1000, 100),

(1000, 500), (1000, 1000) and (1000, 1500). These

11 basic LV models for each cycle time also serve as

the base incident-free models. Then, incidents were

generated on these base test-beds with different

start-times for each incident model. The incident

models were run with the same random seed number

and initial warm-up period as of the corresponding

base incident-free models. Finally, we have 66

incident models for the 60-second cycle time cases,

55 incident models for the 80-second cycle time

cases, and 66 incident models for the 100-second

cycle time cases.

Each simulated incident model (also,

corresponding non-incident base model) was run for

the time-period of around ½ hour (i.e. 30 time steps,

23 time-steps and 18 time-steps for the 60, 80 and

100 sec signal cycle times, respectively, where a

time step is equal to a cycle time). The exact

incident specifics with the 60-second cycle time are

denoted here by the [run no: incident start time,

incident duration]. The exact runs are [R1: 2, 6],

[R2: 6, 6], [R3: 11, 6], [R4: 16, 6], [R5: 21, 6] and

[R6: 26, 5]. The 80-second runs are [R1: 2, 6], [R2:

6, 6], [R3: 11, 6], [R4: 16, 6] and [R5: 21, 3]. The

100-second runs are [R1: 2, 6], [R2: 6, 6], [R3: 11,

6], [R4: 16, 6], [R5: 21, 3] and [R6:17, 2].

4 DATA ANALYSIS

The approach used for the data analysis is based on

the assumption that it is likely that the traffic

measures (extracted from detectors) of the incident-

induced cycle-time will vary from the counter traffic

average values measured in no incident case. The

proposed model operates with a time step (cycle

time) resolution; to detect the incident status at every

cycle time.

The considered traffic measures are the

‘accumulated detector counts’ and the ‘average

detector speeds’ for all the three detectors. The data

extraction period is equal to the green split time of

that cycle. That is, for every cycle time, there are

four data extraction periods.

For the upstream detector and mid-lane

detectors, the traffic measures are estimated for each

cycle time including 4 split phases. For the

downstream detector, only the traffic measures

during the green phase are used. During the red

phases, it is expected that detectors will indicate

fixed counts and zero speed. Except for the front

leading vehicles near the STOP line (near the

downstream detector), no other vehicles would hit

the downstream detector during the red phases.

AFuzzyLogicModelforReal-timeIncidentDetectioninUrbanRoadNetwork

467

5 DEVELOPMENT OF

FUZZY-BASED INCIDENT

DETECTION MODEL

A neuro-fuzzy approach was adopted to develop the

incident status prediction models presented in this

paper. In developing the fuzzy model, the

independent variables (as indicated above in Table

1) are the traffic measures extracted from the

simulation detectors. The parameters UC, US, MC,

MS, DC and DS of Table 1 are the recorded detector

data for each of the simulation time-step of the

incident models or the operating time-step in reality.

The parameters C

, C

, S

and S

are also the

recorded detector data for each time step with no-

incident scenarios out of the corresponding base LV

model of each operating configuration. In reality,

based on the detector readings of (say) the previous

3 to 5 time steps, the model could identify the

closest base scenario for the retrieval of the

parameters.

After some comprehensive statistical

significance tests with Factor Analysis in Minitab, it

was observed that Y

, X

, Y

and Y

in Table 1 are

the most significant independent variables in

predicting the incident status by some general linear

regression (GLM) models. So, the same four

independent variables were considered as the input

membership functions for the fuzzy logic models.

The dependent variable of the fuzzy model is

either an incident status (yes) or a normal recurrent

traffic condition (no incident) of a single time-step.

Because of the unavailability of the neuro-fuzzy

training for the discrete binary dependent variables,

the dependent variable of the proposed fuzzy-logic

was considered as continuous variable while using

the program FuzzyTECH 5.5 (INFORM, 2001).

The ‘true’ range of the dependent variable of an

incident status is allocated the central value of 1 for

an incident, and the ‘false’ range of the dependent

variable of a non-incident status is allocated the

central value of 0.

In applying the fuzzy-logic model to predict the

incident status, a threshold value is utilized. If the

estimated dependent variable is higher that the

threshold value (say 0.500) an incident is indicated,

otherwise it's a normal condition. The threshold

value is chosen to maximize the incident detection

rate and minimize false alarms. Such threshold was

determined through comprehensive iterative

analyses. Initially, the value of 0.500 was set as the

intuitive separating point between incident and non-

incident status. Then, a brute-force search was

adopted with 0.001 units of increase/decrease for

next iteration until the improvement in the incident

detection rate is noticed, while keeping the false

alarm rate within some acceptable limits.

The adopted measures of effectiveness of this

model are as follows:

Incident Detection Rate: The percentage of time

steps that the FLM predicts the incident time steps

correctly. The true detection of incident status of a

time step is defined as the prediction of an incident

status by the model while this time step was truly an

incident-induced simulated time step.

False Alarm Rate: The percentage of time steps

that the FLM predicts an incident status out of all

normal incident-free time steps. The false detection

of a time step is defined as the prediction of an

incident status by the FLM while this time step was

truly incident free.

It is to be noted that the average time to detect the

incident is the duration of one time step (which is

equal to one cycle time) as this model detects

whether an individual time step is incident-induced

or incident-free.

6 FUZZY SYSTEM STRUCTURE

The simple FLM structure shown in Figure 2 was

adopted. The connecting lines symbolize the data

flow.

Four input variables with the associated

linguistic terms were identified for the logic as

shown in Table 2. The output ‘Incident_Status’ is an

index for incident possibility with two linguistic

terms (False and True); the higher the index the

higher the possibility of an incident.

6.1 Fuzzification

The linear (L-shaped) membership function (MBF)

was adopted for all variables. The membership

functions are initially set equally distributed over the

range of all possible values. Each variable’s term is

defined by that single value that corresponds to a

term membership value (µ) of 1 as shown in Figure

The Neuro-Fuzzy training algorithm is used to

optimize these confidence levels and the

membership functions via data training as will be

explained later.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

468

Table 1: Traffic measures used in the incident detection models.

Detector

Traffic measures of the incident

scenarios

[for each analysis

time-step (cycle time)]

Traffic measures of the no-

incident

scenarios

[for each analysis

time-step (cycle time)]

Parameters to be used in the models (for

each cycle time)

n: total number of simulated cycle times

for a specific combination

[n= 30, 23 and 18 for the cycles of 60 ,

80 and 100 seconds, respectively]

Vehicle count

measures

Speed measures

Vehicle count

measures

Speed measures

Vehicle count

measures

Speed measures

Upstream

detector

[data at each

cycle]

Total vehicle

count

(UC)

Average speed

(US)

Total vehicle

count

)

Average speed

)

deviation of

upstream detector

count:

UCX





deviation of

upstream detector

speed:

USY





Midblock

detector

[data at each

cycle]

Total vehicle

count

(MC)

Average speed

(MS)

Total vehicle

count

)

Average speed

)

deviation of

midblock detector

count:

MCX





deviation of

midblock detector

speed:

MSY





Downstrea

detector

[data at each

cycle]

Total vehicle

count [during

green phase]

(DC)

Average speed

[during green

phase]

(DS)

Total vehicle

count [during

green phase]

)

Average speed

[during green

phase]

)

deviation of

downstream

detector count:

DCX





deviation of

downstream

detector speed:

DSY





Input

(Fuzzification)

Inference Engine

(Inference Rule

Base)

Output

(Defuzzi-

fication)

Figure 2: The FLM structure of the urban incident

detection model.

6.2 Fuzzy Inference Process

The fuzzy inference consists of three computational

steps: Aggregation, Composition, and Result

Aggregation (INFORM, 2001). The rules (IF-THEN

logics) were generated to describe the logical

relationship between the input variables (IF part) and

the output variable (THEN part). Initially, all the

possible combinations of rules (3*3*3*3*2=162)

were set initially with equal degree of support (DoS)

of 0.5 as shown in Table 3. The initial value of the

DoS for each rule is adjusted by neuro-fuzzy

training.

Table 2: The FLM input and output variables, numerical

ranges, and linguistic terms.

Variable

category

Variable name

(Denoted in FLM)

Numerical

ranges

Linguistic

terms

Min Max

Input

variables

deviation of

upstream detector

speed (Y

)

-17.96 24.11

Low,

Medium,

High

deviation of

midblock detector

count (X

)

-14.56 42.28

Low,

Medium,

High

deviation of

midblock detector

speed (Y

)

-31.51 21.74

Low,

Medium,

High

deviation of

downstream

detector speed

)

-28.75 43.53

Low,

Medium,

High

Output

variable

Incident Status

(Incident_Status)

-1 +2

False,

True

Among the several available operators (such as

AND, OR, Min–Max, Min–Avg, Gamma), the

‘Min–Max’ operator resulted in minimum training

error.

AFuzzyLogicModelforReal-timeIncidentDetectioninUrbanRoadNetwork

469

(a)

(b)

Figure 3: (a) the initial and (b) finally calibrated MBF of

input Y

(60-second cycle, 50-m link length and 1000

veh/hr scenario).

Table 3: Examples of sample IF-THEN rules.

IF THEN

DoS

(initial:

final)

Incident_

Status

low low low low (0.50: 0.97) False

low low low low (0.50: 0.98) True

low low high low (0.50: 0.49) False

low low high low (0.50: 0.50) True

low low high high (0.50: 0.45) False

low low high high (0.50: 0.50) True

6.3 Defuzzification

Among the several defuzzification methods, the

adopted MoM (Mean-of-Maximum) method delivers

the "most plausible" result that is mostly used in

pattern recognition problems. The MoM method

generates the mean value (z

) of all output units,

whose membership functions (from Figure 4) reach

the maximum as follows:





Where

: output unit whose membership functions reach

the maximum

k: number of such output units.

6.4 Neuro-Fuzzy Data Training

The neuro-fuzzy system can be viewed as a three-

layer feed forward neural network similar to the

above traditional fuzzy system (Figure 2) with a

layer of hidden neurons used to perform each

process. The first layer represents the input variables

fuzzification process, the middle hidden layer

represents the fuzzy rule inference process and the

third layer represents the output variable

defuzzification process.

Figure 4: The final values of output (Incident_Status) after

nuero-fuzzy training of the MBF (60-second cycle, 500-

meter link length and 1000 veh/hr scenario).

The ‘calibration’ here refers to finding the

‘optimal’ fuzzy membership shape and the Degree

of Support (DoS) for the IF–THEN rules. In the first

step, all MBFs and rules were selected for the neuro-

fuzzy training to find the optimised FLM. Then, the

parameters (step width for DoS and terms) were

selected for the training. The whole neuro-fuzzy

training was carried out for five cycles with each

cycle for 1000 iterations.

The step width for the DoS values has been set to

0.1 for each cycle. The step width for the terms has

been set to 5% in the first cycle, which was

increased by 5% in later cycles. The maximum and

average deviations were observed after completion

of each cycle. The cycle, for which the deviation

values are less, was selected as the final FLM. After

the training phase, the MBFs and the DoS values

were determined as shown in Table 3 and Figure

3(b).

7 RESULTS

For the 60-second signal time models, the

performance measures, denoted by (FLM threshold,

incident detection rate, false alarm rate), for the

mentioned 11 LV combinations are (0.500, 26%,

10%), (0.500, 69%, 0%), (0.200, 77%, 24%), (0.200,

46%, 39%), (0.200, 46%, 15%), (0.500, 77%, 9%),

(0.500, 31%, 1%), (0.500, 40%, 19%), (0.500, 66%,

9%), (0.500, 34%, 9%) and (0.500, 43%, 9%),

respectively. With the 80-second signal time models,

the performance measures are (0.500, 41%, 8%),

(0.500, 96%, 0%), (0.500, 78%, 0%), (0.500, 22%,

5%), (0.500, 67%, 3%), (0.500, 85%, 6%), (0.500,

70%, 0%), (0.500, 56%, 51%), (0.500, 0%, 0%),

(0.500, 0%, 0%) and (0.500, 56%, 1%), respectively.

With the 100-second signal time models, the

performance measures are (0.500, 0%, 0%), (0.500,

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

470

90%, 0%), (0.500, 90%, 0%), (0.500, 61%, 21%),

(0.500, 90%, 23%), (0.500, 77%, 17%), (0.500,

81%, 1%), (0.500, 48%, 9%), (0.500, 65%, 5%),

(0.500, 87%, 25%) and (0.500, 68%, 10%),

respectively. Thus, the incident detection rates range

from 0% to 96%, while the false alarm rates range

from 0% to 51%. Except for few operating

conditions, the average detection rate is mostly

above 55%.

The worst performance of the FLM (low

detection rate and/or very high false alarm rate) is

evident with low traffic volumes. At low traffic

volumes, incidents do not significantly impact the

detector readings or the adopted traffic measures

(especially for the cases of partial blockage). Even

with long incident durations, vehicles could easily

bypass the blocked lane through other free lanes.

This limitation (low detection rates at low traffic

volumes) is quite similar to that of the freeway

incident detection models. At such low traffic

volumes one may argue that traffic control centre

does not necessarily have to respond by control

adjustments as the incident does not impact the

traffic flow significantly. Also, for the case of 80-

second cycle time, the FLM for the case of long link

length (i.e. 1000 m) seems performing worse with

relatively low detection rates. This may be attributed

to the delay in detecting incidents caused by the

longer travel times on links.

By excluding the scenarios of low traffic volume

(100 veh/hr), the average detection rate of the

proposed FLM is 64.3% (55%, 57% and 81% for the

signal cycles of 60, 80 and 100 seconds,

respectively), and the average false alarm rate is 7%.

This FLM seems performing better with lower false

alarm rate (7%) as compared to the GLM based

regression models (11.7%) developed by Ahmed and

Hawas (2012). The average detection rate of the

GLM models (64.6%) is close to that of the FLM

(64.3%).

8 VALIDATION TESTS

Another set of different incident scenarios was

modeled with NETSIM for validation test. This

would also test the robustness of the devised FLMs.

Here, all scenarios were modeled with 8 time steps

incident duration (480, 640 and 800 seconds for the

cycle times of 60, 80 and 100 seconds, respectively),

where the incidents starting and ending time steps

were set to 9 and 16, respectively. The calibration

data set was from the lane 1 incidents only, but the

validation data set were generated from the incidents

of both lane 1 and lane 2. Thus, it reflects significant

changes to incident occurrence specifics as

compared to the data used for calibration. Some of

the lane-2 incidents were generated with hourly

traffic volumes of 500 and 1000 veh/hr with various

link lengths and cycle time of 60 seconds. Others

were generated with hourly traffic volumes of 100

and 1500 veh/hr, various link lengths and cycle time

of 80 seconds. Also, some incidents were generated

with hourly traffic volumes of 500 and 1500 veh/hr

with various link lengths and cycle time of 100

seconds.

The developed FLM of each specific

combination (i.e. from the 33 operating

configurations) was used to predict the incident

status using the data of the validation scenarios.

Lane-2 incident validation scenarios resulted in

average detection rate of 32% (standard deviation

20%), and average false alarm rate of 14% (standard

deviation 17%). Lane-1 validation scenarios resulted

in 19% average detection rate (standard deviation

21%) and 8% average false alarm rate (standard

deviation 12%).

It is to be noted that the lower detection rates of

the validation scenarios (as compared to the results

reported in the calibration of the FLM) might be

attributed to the insufficient data as for each specific

combination of cycle time, link length and volume;

only one incident scenario for each combination

whereas there were at least 5 incident scenarios in

calibration. Furthermore, the calibration data was

somehow limited in the sense that it did not consider

overall random variations in incident durations, start

and end times.

9 CONCLUSIONS

This paper presented an FLM approach that

combines simple fuzzy logic models and threshold

values for each specific combination of cycle time,

link length and hourly traffic volume. Except the

relatively lower hourly traffic volumes, the incident

detection and the false alarm rates were satisfactory

for all the cases.

There is still a significant room for improving the

presented FLM to obtain more efficient and robust

models. Also, further challenges remain in

predicting the incident status with significantly wide

variations of the input attributes accounting for other

aspects such as the malfunctioning of the detectors

and the variations with detector placements. Further

research is intended to focus on improving the FLM

by considering the impact of random incident

AFuzzyLogicModelforReal-timeIncidentDetectioninUrbanRoadNetwork

471

locations, durations, and different detector

placements. Other improvements could include the

generalization of the FLM to account for over-

saturated traffic conditions when link spill back

occurs. Further research would also consider

different settings of the FLM structure in terms of

reduced number of input parameters. Also, a general

FLM (for various operational conditions) will be

strongly needed.

REFERENCES

Ahmed, F., Hawas, Y. E., 2012. A threshold-based real-

time incident detection system for urban traffic

network. 4

International Conference: Transport

Research Arena (TRA) Europe, Athens, Greece, April

23-26, 2012.

Dia, H., Thomas, K., 2011, Development and evaluation

of arterial incident detection models using fusion of

simulated probe vehicle and loop detector data.

Information Fusion, 12, 2011, 20–27.

Hawas, Y. E. , 2007. A fuzzy-based system for incident

detection in urban street networks. Transportation

Research Part C, 15, 2007, 69–95.

INFORM, 2001, FuzzyTech 5.5: User’s Manual.

GmbH/Inform Software Corporation, Germany.

Khan, S. I., Ritchie, S. G. 1998. Statistical and neural

classifiers to detect traffic operational problems on

urban arterials. Transportation Research C, Vol 6,

1998, pp. 291–314.

Liu, Y., Lei, Y., Yi, Q., Jianquan, W., Huimin, W., 2007,

Traffic incident detection algorithm for urban

expressways based on probe vehicle data. Journal of

Transportation Systems Engineering & Information

Technology, Vol. 8(4), 2007, 36-41.

Parkany, E. 2005. A complete review of incident detection

algorithms & their deployment: what works and what

doesn’t. The New England Transportation

Consortium. NETCR37.

Sermons, M. W., Koppelman, F. S., 1996. Use of Vehicle

Positioning Data for Arterial Incident Detection.

Transportation Research Part C, 4(2), 1996, 87-96.

Yuan, F., Cheu. R. L. 2003. Incident detection using

support vector machines. Transportation Research C,

Vol 11, 2003, pp. 309–328.

Zhang, K., Taylor, M. A. P., 2006. Effective arterial road

incident detection: a Bayesian network based

algorithm. Transportation Research Part C, 14, 2006,

403–417.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

472