Proceedings of the 12th International IEEE Conference MoBT4.

3
on Intelligent Transportation Systems, St. Louis, MO,
USA, October 3-7, 2009

Prediction of Traffic Density for Congestion Analysis under Indian

Traffic Conditions
Ameena Padiath, Lelitha Vanajakshi, Shankar C. Subramanian, and Harishreddy Manda


Abstract— Traffic congestion is a serious problem which traffic congestion. These include traffic incidents such as vehicle
engineers all over the world are trying to solve. Congestion crashes and breakdowns, work zones, bad weather, and special
increases the uncertainty in travel times leading to human stress events [1].
and unsafe traffic situations. Better management of traffic Traffic congestion is becoming a serious problem in most
through Intelligent Transportation Systems (ITS) applications, Indian cities too. The rapid pace of economic growth and the
especially by predicting the congestion on various roads and resulting growth in vehicular population is one of the main
informing the travelers regarding the same is one possible
reasons for this problem. There is a limit on how much
solution. Accurate and quick prediction is one of the important
factors on which the reliability of such a system depends. If one is additional roadway infrastructure can be added to meet the
able to predict congestion on a roadway, then the travelers can be growing vehicle population. City planners are doing the best
warned of the same either pre-trip or enroute so that they can they can to cope with the increasing demand for transportation
take well informed travel decisions. The number of vehicles in a infrastructure. For instance, flyovers have been, and continue
given stretch of a roadway (usually referred to as “traffic to be, constructed, and existing roads are being widened. But
density”) is one of the most commonly used congestion indicator. creating more infrastructure is not a solution by itself since its
Also, the travelers in general will be more interested to know increase will not be able to match up to the increase in traffic
what they can expect when they make the trip in future rather demand. Hence, a more holistic and integrated approach to the
than the present scenario. This makes the short term prediction to
problem is needed for better management of the existing
future time intervals important. In this study, some of the
reported techniques for density prediction under homogeneous facilities and thus reduce traffic congestion on the roads.
traffic conditions are attempted under heterogeneous traffic For this, in recent years, transportation engineers and
conditions in order to determine their feasibility under the Indian planners have increasingly embraced strategies that deal with
traffic scenario. better management of existing roadways, including Intelligent
Transportation Systems (ITS). ITS includes the application of
Keywords: Intelligent Transportation Systems, Traffic advanced technologies using real-time information about
Congestion, Traffic Density. roadway conditions to implement appropriate control
strategies. Real-time control of roadway operations through a
I. INTRODUCTION Transportation Management Center (TMC) has become a

T RAFFIC congestion is an increasingly serious problem

faced by travelers through out the world. Congestion can
major activity undertaken by transportation agencies. ITS
control strategies take many forms such as metering flow onto
roadways, dynamically retiming traffic signals, managing
be defined in simple terms as the traffic demand exceeding the
roadway capacity. Congestion increases the uncertainty in traffic incidents, and providing travelers with information
travel times leading to human stress and unsafe traffic about travel conditions, alternative routes and alternate modes.
situations. Congestion results from the interaction of many The present study falls in the latter category, where the general
different factors or sources. The first and foremost reason is objective is to provide travelers with information about traffic
that the available infrastructure is not enough to meet the conditions to help them in taking better travel decisions.
demand placed on it. In addition to this physical capacity The information that can be provided to the travelers
constraint, external events can have a major effect on traffic includes expected travel times, traffic conditions such as
congestion, availability of alternate routes, etc. The overall
objective of the present study is informing the travelers
Manuscript received on 15th May 2009. regarding the congestion levels of roadways. This requires real
This work was supported in part by the Department of Science and
time traffic data collection, analysis and efficient and accurate
Technology, Government of India under the grant SR/FTP/ETA-55/2007.
Ameena Padiath is a graduate student in the Department of Civil prediction methodologies. Congestion can be indicated by a
Engineering, Indian Institute of Technology Madras, Chennai 600 036, INDIA. combination of various factors such as speed, volume, density,
Lelitha Vanajakshi (corresponding author) is with the Department of Civil travel time, etc. The number of vehicles within a given section
Engineering, Indian Institute of Technology Madras, Chennai 600 036, INDIA. of roadway (usually referred to as “traffic density”) is one of
E-mail: [email protected], Ph: 91 44 2257 4291, Fax: 91 44 2257 4252.
Shankar C. Subramanian is with the Department of Engineering Design, the main traffic parameters which can directly be used to
Indian Institute of Technology Madras, Chennai 600 036, INDIA. E-mail: determine the level of congestion. Also, the travelers in general
[email protected] will be more interested to know what they can expect when
Harishreddy Manda was an undergraduate student in the Department of they make the trip in future rather than the present scenario.
Civil Engineering, Indian Institute of Technology Madras, Chennai 600 036,
INDIA. This makes the short term prediction to future time intervals
important. There are different studies in the past which

78
978-1-4244-5521-8/09/$26.00 ©2009 IEEE
explored various techniques for the short term prediction of where k is the traffic density, Lv is the average vehicle length,
traffic density including historic and real time method, Ld is the detection zone length and %occ is the percentage
statistical methods, machine learning techniques and model occupancy time. As occupancy increases, the density and in
based techniques [2-12]. turn congestion increase. The present study uses this method
However, all these studies were reported for the for the automated measurement of density.
homogeneous and lane disciplined traffic. The traffic situation For validation of the density calculated from occupancy
in many countries, including India, is different with its values, actual density measurement was carried out manually
heterogeneity and lack of lane discipline. It is characterized by using input output analysis. The input-output technique [13], is
a wide variety of vehicles with varying characteristics using a rather straightforward approach. Here, an initial count is
the same road space, without following any lane discipline and made of the number of vehicles existing in the roadway
occupying the entire road space. Under such conditions many between two count stations and over time the number of
of the reported methods for short term prediction of traffic vehicles entering the section is continuously added and the
parameters may not be suitable. The present study is an number of vehicles leaving the section is continuously
attempt to explore and find out the feasibility of these subtracted from the initial count. This method became more
techniques for the prediction of traffic density under Indian attractive with the introduction of loop detectors for automatic
traffic conditions. Traffic data were collected using traffic data collection. This would be an efficient method to
videographic technique for this study. A comparison of the find out the true density in a section, if the counts are accurate.
performance of the following methods will be carried out: The problem is that section density is calculated on the basis of
Historic method, Artificial Neural Networks (ANN) and a the difference between two large numbers (input and output
model based approach using Kalman filtering technique. counts), and count errors even at the low level of 1 percent can
The following section discusses traffic density in general not be tolerated. The input-output count technique therefore is
and some of the existing literature on the short term prediction practical only if the counts are very reliable and accurate. The
of traffic density and congestion. The data collection process is other problem of using this technique is correctly estimating
described in section III and the various prediction techniques the initial number of vehicles present. A small error in
are described in sections IV. In section V, a comparison of calculating the initial value can lead to large errors in
these various techniques is presented. Concluding remarks are calculating density.
presented in section VI. Some of the reported studies which discussed congestion
prediction, using various traffic parameters as indicators of
II. LITERATURE REVIEW
congestion are given below. Ben-Akiva et al. [6] presented
short descriptions of a model system that was being developed
A brief discussion about traffic density followed by a
to predict traffic congestion as a first step towards providing
discussion of some of the important studies dealing with the
real time traffic guidance in Netherlands. Travel time and flow
short term prediction of traffic congestion and traffic density is
were used as measures of traffic congestion and they discussed
given in this section.
different methods for short term prediction including the
Traffic density is the number of vehicles existing over unit
Kalman filtering technique. Kaysi et al. [7] discussed the use
length of a roadway. This is a parameter which directly
of historical data profiles as part of the development of a
indicates the quality of traffic and influences the ease and
congestion prediction algorithm for a real time driver guidance
comfort with which one can drive. Because of this, density is
system in West Berlin. Their approach involved the creation of
being used as a measure of effectiveness of many traffic
a historical data profile to make the prediction to the future
systems. One of the drawbacks associated with this parameter
time steps. Dougherty et al. [8] reported a congestion
is the difficulty in measuring it in field using existing data
prediction study using neural network method and concluded
collection techniques. Direct measurement of density involves
that the ANN method as a potentially considerable value
elevated photography which will cover the entire section under
addition to the traffic engineering field. Huisken and Coffa [9]
study, and this in most of the cases will be impractical and
reported a comparative study on the use of ANN and Auto
expensive. Hence, various methodologies for the estimation of
Regressive Moving Average (ARMA) time series analysis for
traffic density from other easily measurable traffic parameters
the prediction of traffic congestion. The study results indicated
such as flow, speed, occupancy, etc., have been developed.
a better performance by ANN during non-incident cases. Chen
Out of these, occupancy is a variable which is directly related
et al. [10] proposed a congestion prediction algorithm based on
to traffic density and is used as an indicator of traffic density.
auto regressive fuzzy logic. The reported results indicated that
Occupancy is the fraction of time the measurement location
the fuzzy-AR model based on clustering algorithm gave good
was occupied by vehicles out of the total observation period.
prediction, and provided satisfactory traffic congestion control.
Density estimates were calculated from the occupancy values
A report from the Texas Transportation Institute [11] discussed
using a standard relation between density and occupancy as
various aspects of short term congestion prediction and
[13]
presented a modified input output analysis. It presented a non
%occ continuum approach for the prediction of traffic parameters
k , (1)
( Lv  Ld ) 100 and evaluated their performance using field data. Huisken [12]

79
reported a comparative study of the performance of different For the validation using input-output analysis, the initial
data driven prediction techniques such as regression, ARMA number of vehicles existing inside the section at the start of
time series analysis, four different ANN models and fuzzy data collection was needed. This was carried out by taking a
logic, for the short term prediction of traffic congestion. The still picture of the section from the pedestrian over bridges at
results were also compared with a ‘do nothing’ scheme to get a the start of the data collection and manually counting the
better grasp of the impact of the prediction results. number of vehicles present in the section at that time.
Thus, it can be seen that most of the reported studies on
congestion prediction were developed for the homogeneous IV. PREDICTION TECHNIQUES
traffic such as that existing in countries such as USA. The
present study tries to analyze the suitability of some of these Once the density of traffic is estimated from the occupancy
techniques for heterogeneous conditions. A comparison is values, prediction into future time steps was carried out. The
carried out between a historic technique, an ANN based present study compares three different techniques namely, a
technique and a model based technique. historic technique, an ANN based technique and a model based
technique, which are detailed below.
III. DATA COLLECTION
Historic technique
The present study calculates density from the automated
occupancy measurements. Also, for validation purpose, an The historic approach is based on the assumption that the
input output analysis was carried out. Thus, the data historic profile can represent the traffic characteristics for a
requirement included the occupancy values, initial number of given time of the day. Thus, a historical average value will be
vehicles inside the section and other parameters such as used for predicting future values. This method can be valuable
number of vehicles entering and leaving the section and speed in the development of prediction techniques since they explain
of vehicles. a substantial amount of the variation in traffic over many days.
Data for the present study were collected using videographic However, for the same reason, the reliability of the prediction
technique from a 1 km stretch on the Rajiv Gandhi road, is limited because of its implicit assumption that the projection
Chennai, India. The selected roadway is a three lane road in ratio remains constant. Though, a simple method, it was tried
each direction. Video data were collected at the entry and exit out since this is still one of the most popularly adopted
points of the selected stretch from pedestrian overbridges. The methods for short term prediction of traffic parameters in the
section has one side entry ramp in between and the vehicles field. The average of the density data from four days data was
entering and exiting through the same is counted manually. A taken as the predicted density for the fifth day and a
schematic representation of the study site is given in Figure 1. comparison with the actual density on the fifth day is carried
out.

ANN Technique

An Artificial Neural Network (ANN), in the most general

sense, is an information processing structure whose design is
motivated by the design and functioning of human brains and
components thereof. Thus, ANNs are computing techniques,
which can be trained to learn a complex relationship in a data
set. Basically it is a parallel computing system composed of
Fig. 1. Schematic representation of the study site interconnected simple processing nodes, which are non-
algorithmic, non-parametric and intensely parallel [15].
Data were collected in the afternoon peak hours of five In this study, for ANN analysis, the previous time steps
week days, each for a duration of approximately 60 minutes. values were used as inputs and the future time step values were
Data extraction was carried out using a video image processing obtained as outputs. This follows the idea of time series
software namely Trazer (Traffic Analyzer and Enumerator) analysis where historical data are analyzed in an attempt to
[14]. Trazer is an offline video processing system developed identify a pattern, and then assuming that the pattern will
for mixed traffic conditions such as the one existing in India. continue in the future, it is extrapolated in order to produce
The software can be used for extracting traffic data on speed, forecasts.
volume and occupancy at specified intervals. In the present A multi layer network with back propagation algorithm was
study data were extracted for every one minute interval using selected for the present study. The network selected had three
this software. The volume data were manually extracted from layers, with 5 neurons in the hidden layer. These parameters
the video since the accuracy of the input output analysis were selected by trial and error for the best result. Four days’
depends mainly on the count data as explained under the data were used for training the network and 5th day’s data were
literature review section. used for testing the results. During training, the network takes

80
the first 5 one minute interval density values for computing the model presented above is non-linear, the Extended Kalman
next one, i.e., the 6th minute density value. Then the network Filter (EKF) is used for prediction.
compares the actual density value with the predicted value. The Kalman filter [17] is a popular tool for recursive
Using the back-propagation algorithm the network tries to estimation of variables that characterize a system (these
minimize the error between actual and the predicted values. variables are usually referred to as ‘state variables’). The
This way the network was trained and then was tested using Kalman filter is a model-based estimation scheme that takes
the testing set. into account the stochastic properties of the process
disturbance and the measurement noise. The process
Model Based Approach disturbance and the measurement noise are assumed to be
independent of one another, white and normally distributed
In this approach, a set of equations is developed to relate the with zero mean. The Kalman filter works like a predictor-
x corrector algorithm, i.e., it first predicts an ‘a priori’ estimate
various traffic parameters in a section of a roadway. Let N en of the state variables using the system model and the state
x estimate from the previous time interval, and then corrects the
and N ex denote the time rate at which vehicles enter and exit same using measurements to obtain an ‘a posteriori’ state
x estimate. The Kalman filter has been widely used in many
the section respectively. Let N ramp denote the net time rate at disciplines including the field of transportation [18] – [25].
The Kalman filter is used for estimation and prediction
which vehicles enter the section from the ramp. Let N s denote
when the governing equations of the system are linear. When
the number of vehicles in the section at any instant of time. the governing equations are nonlinear, an ‘Extended Kalman
Then, from the balance of vehicles in the section, we have filter’ (EKF) is commonly used. The EKF linearizes the
x x x x governing equations at each time step about the estimate
Ns N en  N ex  N ramp , (2) obtained from the previous time step. Let us assume that the
where the dot indicates the derivative with respect to time. governing equations of the system can be written as
Then, the time rate of vehicles exiting the section was xk 1 f xk , uk  H k wk , (6)
modeled using [16]
x vN s yk h xk  vk , (7)
N ex , (3)
Ls where xk , uk and yk denote respectively the state variables,
where v denotes the aggregate speed of the vehicles in the input and output at the kth instant of time. Here, wk and vk
section at any instant of time and Ls denotes the length of the denote respectively the process disturbance and measurement
section under study. noise. The process disturbance covariance is denoted by Qk
Discretizing the above equations using a time step h, the
following equations are obtained and the measurement noise covariance is denoted by Rk . Let
§ x x
·x
xˆk and xˆk denote respectively the a priori estimate and the a
N s (k  1) N s (k )  h ¨ N en  N ex  N ramp ¸ , (4)
© ¹ posteriori estimate of the state variables at the kth instant of
and § wf · § wh ·
time. Let us define Ak : ¨ ¸ and Gk : ¨ ¸ .
§ vN · © wx ¹ x © wx ¹ x
N cum
ex (k  1) N cum
ex (k )  h ¨ s ¸, (5) xˆk xˆk

© Ls ¹ Let Pk and Pk denote respectively the a priori and the a
cum
where N ex ( k ) denotes the cumulative number of vehicles posteriori error covariance at the kth instant of time. Then the
that have exited the section till the kth interval of time. Now, if following recursive algorithm is used to obtain the estimate of
one wants to find the number of vehicles present in a given the state variables:
section during the (k+1)th time interval, the number of vehicles 1. The a priori estimate in the (k+1)th interval of time is
present in the section during the kth time interval should be obtained through xˆk1 f xˆk , uk .
known in addition to the rate of change of the number of
vehicles between the two time intervals. Thus, if the number of 2. The a priori error covariance in the (k+1)th interval of
vehicles in a section of a roadway during a desired interval of time is obtained through
time is to be found, the main issue is that the initial number of Pk1
Ak Pk AkT  H k Qk H kT .
vehicles present in the section is unknown (and this number is
not readily available in any of the existing automated data 3. The Kalman gain K k 1 is calculated through
collection procedures). Hence, a model-based scheme that 1

predicts the number of vehicles in the section using the

K k 1 Pk1GkT1 Gk 1 Pk1GkT1  Rk 1 .
available measurements has been developed here. Since the

81
 4. Then, the a posteriori state estimate is calculated process of collecting more video data which may improve the
   prediction accuracy of data driven techniques such as the
through xˆ
k 1 xˆ
k 1  K k 1 yk 1  h xˆ
k 1 .
historic technique and the ANN based technique.
5. Finally, the a posteriori error covariance is obtained
80

through Pk 1 I  K k 1Gk 1 Pk1 . Prediction
Measurement
The above five steps are repeated for each time interval to 70

Number of Vehicles in the Section

obtain the estimate of the state variables.
The number of vehicles in the section N s and the 60

cumulative number of vehicles exiting the section N excum are 50

cum
taken as the state variables of interest. The variable N ex is
40
taken as the output variable and the number of vehicles
entering the section and the net number of vehicles entering
30
through the ramp during each time interval are provided as
inputs to the prediction scheme. The initial values of the state
variables were assumed and the prediction scheme was 20
0 10 20 30 40 50 60
implemented. Time (min)

Fig. 2. Comparison of the predicted and actual number of vehicles in the

V. COMPARISON OF THE TECHNIQUES section – using model based technique.

In this section, the performance of the above techniques is 80

compared with actual density values obtained using input- Prediction
Measurement
output analysis. The density calculated using input output 70
analysis was used for measuring the accuracy of prediction in
Number of Vehicles in the Section

each case. The performance of the models was quantified 60

using the Mean Absolute Percentage Error (MAPE) and is
obtained using
50
§ 1 N N pred (k )  N meas (k ) ·
MAPE ¨ ¦ ¸ *100%, (8)
¨N k 1 N meas (k ) ¸ 40

© ¹
where N pred ( k ) and N meas (k ) are the predicted and the 30

measured number of vehicles in the section during the kth

20
interval of time with N being the total number of time 0 10 20 30 40 50 60
Time (min)
intervals.
As mentioned before, the data from the first four days were Fig. 3. Comparison of the predicted and actual number of vehicles in the
section – using ANN based technique.
used to get the predicted values for the fifth day in the case of
the historic technique and the ANN based technique. The
model based approach does not require the data from the first 90
Prediction
four days and this is an advantage of the same since it is not Measurement
dependent on past days’ data. The comparison of these three 80

techniques has been done for the fifth day and the results are
Number of Vehicles in the Section

70
presented in Figs. 2-4. The MAPE for the historic technique,
the ANN based technique and the model based technique was
60
obtained to be 38.57 %, 32.47 % and 40.83 % respectively.
It should be pointed out here that traffic data collection in 50
India is not yet automated. For the current study, the videos
had to be recorded by manually setting the cameras at the test 40
locations. Also, the time of data collection was limited to sixty
minutes due to other constraints. Since the accuracy of the 30
Kalman filtering technique depends on the length of time for
which the analysis is carried out, we feel that the accuracy of 20
0 10 20 30 40 50 60
the model based approach may improve once we are able to Time (min)

collect video data for longer time periods. Also, we are in the Fig. 4. Comparison of the predicted and actual number of vehicles in the
section – using historic technique.

82
 [7] I. Kaysi, M. Ben-Akiva, and H. Koutsopoulos, “An integrated
approach to vehicle routing and congestion prediction for real-time
VII. CONCLUSIONS driver guidance,” Transp. Res. Rec. 1408, Transportation Research
Board, Washington, D.C., 1993, pp. 66-74.
[8] M. S. Dougherty, H. R. Kirby, and R. D. Boyle, “The use of neural
Congestion prediction and informing the travelers on the networks to recognize and predict traffic congestion,” Traffic
congestion status of various alternate routes are important Engineering and Control, 34(6), 1993, pp. 311 – 314.
facets of ITS. These have been already implemented in many [9] G. Huisken, and A. Coffa, “Short term congestion prediction:
developed nations. However, implementing such an comparing time series with neural networks,” IEE conference on
application under a heterogeneous, less lane disciplined traffic road transport information and control, 2000, pp. 66 -69.
[10] B. Chen, S. Peng, K. Wang, “Traffic Modeling, Prediction, and
condition, such as the one on Indian roads, is more
Congestion Control for High-Speed Networks: A Fuzzy AR
challenging. There is a need to check whether the existing Approach,” IEEE Transactions On Fuzzy Systems, VOL. 8, NO. 5,
techniques and models which have been demonstrated to work 2000, pp. 491 – 508.
satisfactorily under the near-homogeneous, lane disciplined [11] Dynamic Traffic Flow Modeling for Incident Detection and Short-
traffic conditions will work for the traffic conditions existing Term Congestion Prediction, Report No. FHWA/TX-06/0-4946-1,
in India. The present study is one attempt in this direction. Texas Transportation Institute, URL:
http://tti.tamu.edu/documents/0-4946-1.pdf, 2005, Accessed July 17,
Congestion can be indicated by various parameters or a 2008.
combination of parameters such as speed, volume, density, [12] G. Huisken, Inter-Urban Short Term Traffic congestion Prediction,
travel time, etc. However, traffic density, which is defined as Trail Thesis Series, The Netherlands TRAIL Research School, 2006.
the number of vehicles existing within a given length of a [13] A. D. May, Traffic flow fundamentals, Prentice-Hall, Inc., New
roadway, is a good indicator of congestion. But, density is a Jersey, 1990.
parameter which is difficult to measure and hence is usually [14] http://www.kritikalsolutions.com/trazer.asp.
[15] S. S. Haykin, Neural networks: A comprehensive foundation,
inferred from other traffic parameters which can be readily Prentice-Hall, Inc., New Jersey, 1990.
measured from field. The present study explored the use of a [16] V. Tyagi, S. Darbha and K. R. Rajgopal, “A dynamical systems
historic technique, an ANN based technique and a model based approach based on averaging to model the macroscopic flow of
technique for the prediction of traffic density in Indian traffic freeway traffic,” Nonlinear Analysis: Hybrid Systems, Vol 2, Issue 2,
conditions. The results from these various techniques are 2008, pp 590-612.
[17] R. E. Kalman, “New Approach to Linear Filtering and Prediction
corroborated using data collected from the roads in Chennai,
Problems,” Transactions of the ASME, Journal of Basic Engineering
India, using videographic technique. The results obtained were Series D, Vol. 82, 1960, pp. 35 – 45.
compared with actual numbers obtained through input output [18] S. Chien, X. Liu, and K. Ozbay, “Predicting Travel Times for the
analysis and the results were found to be promising. This is a South Jersey Real-Time Motorist Information System,” Presented at
preliminary study which will be ultimately used for predicting the 82nd Annu. Meeting (CDROM), Transportation Research Board,
congestion on Indian roads. The next step of this research National Research Council, Washington, D.C., 2003.
[19] C. Nanthawichit, T. Nakatsuji, and H. Suzuki, “Application of Probe
work will involve modifying this model-based scheme to Vehicle Data for Real Time Traffic State Estimation and Short Term
include the various categories of vehicles so that the effects of Travel Time Prediction on a Freeway,” Presented at the 82nd Annu.
heterogeneity can be investigated, which may improve the Meeting (CDROM), Transportation Research Board, National
accuracy of the prediction scheme. Research Council, Washington, D.C., 2003.
[20] A. Shalaby, and A. Farhan, “Bus Travel Time Prediction Model for
REFERENCES Dynamic Operations Control and Passenger Information Systems,”
Presented at the 82nd Annu. Meeting (CDROM), Transportation
[1] Congestion report. Traffic Congestion and Reliability: Linking Research Board, National Research Council, Washington, D.C.,
Solutions to Problems, Available: 2003.
http://ops.fhwa.dot.gov/congestion_report_04/congestion_report.pdf, [21] H. Liu, H. van Lint, H. van Zuylen, and K. Zhang, “Two Distinct
2004, Accessed July 30, 2008. Ways of Using Kalman Filters to Predict Urban Arterial Travel
[2] D.C Gazis and H. Knapp, “On-line estimation of traffic densities Time,” IEEE Conference on Intelligent Transportation Systems,
from time-series of flow and speed data,” Transportation Science, 2006, pp. 845 – 850.
Vol 5, No.3, 1971, pp 283-301. [22] J. S. Yang, “Travel Time Prediction Using the GPS Test Vehicle and
[3] L. Muiioz, X. Sun, R. Horowitz', and L. Alvarez, “Traffic Density Kalman Filtering Techniques,” American Control Conference,
Estimation with the Cell Transmission Model,” Proceedings of the Portland, OR, USA, 2005, pp. 2128- 2133.
American Control Conference, 2003. [23] I. Okutani, and Y. J. Stephanedes, “Dynamic Prediction of Traffic
[4] A. L. Icaza, L. Munoz, X. Sun, and R. Horowitz, “Adaptive observer Volume through Kalman Filter Theory,” Transportation Research,
for traffic density estimation,” Proceeding of the American Control Part B, vol. 18B, 1984, pp. 1-11.
Conference, 2004. [24] W. Guan, and X. Hua, “A Combination Forecasting Model of Urban
[5] A. Kurkjian, S. B. Gershwin, P. K. Houpt, and A.S. Willsky. Ring Road Traffic Flow,” IEEE Intelligent Transportation Systems
“Estimation of roadway traffic density on freeways using Conference, 2006.
presence detector data,” in proceedings of 9th IEEE Intelligent [25] Y. Wang, M. Papageorgiou, and A. Messmer, “Real –time freeway
Transportation Systems Conference, 2006. traffic state estimation based on extended Kalman filter,”
[6] M. Ben-Akiva, G. Cantarella, E. Cascetta, J. Ruiter, J. Whittaker, Transportation Science, 41(2), 2007, pp. 167-181.
and E. Kroes, “Real time prediction of traffic congestion,” IEEE
Vehicle Navigation & Information Systems, The 3rd International
Conference On Vehicle Navigation 8 Information Systems, 1992,
pp. 557 – 562.

83