2010 13th International IEEE MA8.

2
Annual Conference on Intelligent Transportation Systems
Madeira Island, Portugal, September 19-22, 2010

Traffic and Mobility Data Collection for Real-

Time Applications
J. Lopes, J. Bento E. Huang, C. Antoniou, M. Ben-Akiva
Instituto Superior Técnico, Intelligent Transportation System Laboratory
Brisa Auto-Estradas de Portugal MIT, Cambridge MA, USA
jlopes, jbento@brisa.pt enyang, costas, mba@mit.edu
Abstract - Successful development of effective real-time traffic traffic data. Figure 1 presents a schematic overview of the
management and information systems requires high quality traffic process chain from data acquisition until transfer for third
information in real-time. This paper presents the state-of-the-art of party applications and services.
traffic and general mobility sensory technology and a suite of methods
for data pre-processing and cleaning for real-time applications. We
propose a suite of methods and techniques to be applied from traffic Data Mining and
Preprocessing Fusion
Collection Transfer
data acquisition, preprocessing, transformation and integration until
data advanced processing and transfer. Next, we detail some techniques • Data • Data • Data • Knowledge
for data preprocessing and integration, or fusion, phases. Even though acquisition measuring reduction discovery
• Data • Data • Patterns
the comprehensive use of historical traffic data and assignment models cleaning fusion recognition
to support the most part of online services and operations, real-time • Data • Data
data is extremely important to promote models’ accuracy and, Clustering transfer
therefore, the reliability of information and outputs derived from data
fusion and processing. Together with techniques and theoretical
Figure 1: Schematic workflow of traffic data collection and processing
formulas we present a case study applied to the Portuguese Brisa’s A5
motorway, a 25 km inter-urban highway between Lisbon and Cascais.
Since all data available is accurately collected from the
Traffic on this motorway heading to Lisbon in the morning rush hours
transportation system, or it is related with it, the ability to
typically experiences high levels of congestion. Brisa, the motorway
use all datasets combined in the most appropriate way is a
operator company, has equipped A5 with a variety of traffic sensors to
be used in a real-time multi-purpose way, either for traffic
key issue for the information systems solution providers and
management and control or for traveler information and third-part
a challenge for the data processing systems and methods. On
applications. top of those pressing solutions, traffic and congestion
estimation and forecasting methods aims to maximize the
Keywords: Traffic data acquisition, Data preprocessing, Data use of the pool of versatile data, through the development of
transformation and fusion algorithms and techniques to fuse those seemingly unrelated
data into high level information which is readily usable by
I. INTRODUCTION predicting systems.
Continuing deployment of a large number of sensors,
This paper is concerned with the traffic data collection,
telemetry and telematics devices, and other on-line traffic
preprocessing and fusion chain to support real-time
and mobility data collection tools, pushed by road traffic
applications, whereas data completeness, consistence,
engineering, such as traffic monitoring, signal control,
performance and reliability promote an equilibrium equation
automatic incident detection and recovery, traffic forecasting
for effective implementations.
and traveler information, has increased hugely the amount of
time-series traffic data [Antoniou et al., 2008]. The range of The paper is organized as follows: State-of-the-art sensory
new data generated provides remarkable opportunities for technology for traffic and generic mobility data collection is
enhancing decision-making in areas such as transport presented in Section II. Then some descriptions and figures
systems management and control, traveler information about the case study area in Section III. This is followed by
services, and urban and transportation planning. However to the introduction of methods and techniques for data
manage and make efficient use of such amounts of preprocessing and cleaning, along with some experimental
heterogeneous data sets in a comprehensive way, it’s also a results in Section IV. Finally, in the Section V we discuss
big challenge, and even bigger for real-time use cases and results and introduce some works and guidelines for future
applications. research.
The foremost data-collection task is therefore an essential
component on the process chain, and definitively does not II. TRAFFIC AND MOBILITY SENSORY TECHNOLOGY
end on the data acquisition. Together with data For many years, under growing pressure for improving
preprocessing and fusion, implements a silent but crucial traffic management and control, traditional on-road sensors,
mission to provide the essential raw material for upcoming (e.g. inductive loop detectors), were massively installed and
applications and systems: stable, coherent and efficient collecting methods have been evolving to obtain, compute

978-1-4244-7659-6/10/$26.00 ©2010 IEEE 216

and transfer traffic data. [Klein, 2006] Along the roadway,
such in-situ technologies, either based on intrusive or non-
non
intrusive techniques, revealed necessary but not sufficient
because of their limited coverage and expensive costs of
implementation and maintenance. In the last years, several
alternatives technologies and datata sources emerged, driven
by innovative methods and models. This section presents a
comprehensive overview of sensor technology applied to
transport engineering.
Traffic data collection is primarily categorized
orized in three major
methods: site data, floating car data and wide-area
wide data. Figure 2:: Probe vehicles systems for vehicle
vehicle-vehicle and vehicle-
infrastructure cooperation
Each of these methods has different technical characteristics
and principles of operation, including structure of data Wide-area:
area: under research method and technology aims to
collected, accuracy of the measurements and network carry out wide-area
area traffic flow monitoring capabilities
coverage. [Antoniou et al., 2008] An alternative based on multi-sensor tracking options such as
classification establishes such sensor types based on their photogrammetric processing, video analysis, sound
functionality as point, point-to-point,
point, and wide area. The recording, and space-based radar.
third and under research methodology, wide-area
wide survey
Table I classifies the various traffic data collection methods
aims to get automated traffic condition snapshots from the
and technologies by their network and vehicle coverage, and
road network.
traffic data collected. As previously defined, based on the
Site data:
a: refers to traffic data measured by the means of network coverage, collection methods can be limited to a
sensors located along the roadside with diverse technologies particular site location; can be stretched to fixed road
and application techniques. Some have been employed for segments or trips defined by identifying sensors; or can be
many years such as inductive magnetic loops, pneumatic extended to a wide-area
area network thr
through autonomous on-
road tubes, piezoelectric loops arrays
rrays and microwave radars. board equipped vehicles and airborne sensors.
With the recent technology developments, new sensors for
roadside sites came out powered by flexibility, multi-multi
purpose and cost effectiveness. Examples of those new TABLE I
sensors are ultrasonic and acoustic sensor systems,
TRAFFIC AND MOBILITY DATA COLLECTION METHODS
magnetometerr vehicle detectors, infrared systems, LIDAR –
AND TECHNOLOGIES
light detection and ranging, and video image processing and
detection.
Traffic and Mobility Data
Floating car data (FCD) also defined as vehicle-based
detection: refers to mobility data collection by locating and
Speed,

Network Collection

Sub-path flows
recognizing vehicles att multiple points in a network, where Coverage Method
Classification

Travel Times
specific detectors are real or virtually installed. [Huang,

OD Flows
Volume,

2010] Some of those point-to-point point sensors provide

Density

complete transversal of the travel path, providing excellent

and confident information for route choice analysis
ana and OD
estimations. Examples of FCD sensors are: license plate Inductive
X X
recognition (LPR), automatic vehicle identification (AVI) loops
transponders including probe vehicles and electronic toll Road tubes X X
tags. Varying from previous methods, global positioning
Piezoelectric X X
system (GPS) combined
mbined with wireless communication
service GSM/GPRS, provides mobility data from dynamic Microwave
X X
Site- Radar
segments where equipped vehicles, in terms of delay and based
congestion level. LIDAR/
Infrared/ X X
More recently, new technologies arisen both from road Acoustic
infrastructure and vehicle side, in thee first instance Video
established for vehicle-to-infrastructure
infrastructure and vehicle-to-
vehicle X X(1)
processing
vehicle cooperation. Those identifiable vehicles, equipped
Toll plazas X X
with wireless communications devices, are able share
mobility data each other and with roadside devices, as LPR X X
shown in the Figure 2. Finally wireless cell
cell-phone tracking Road Transponders X X
also provides important mobility data about moving mobile-
mobile segments
phones intensity and times along the road network.
netwo Wireless
X X X
devices

217
 GPS X X X toll collection service, based on dedicated short-range
Cell-phone communications (DSCR) microwave communications, A5
Wide- X X X toll plazas, located between kilometer 11 and 19, are
tracking
area
Airborne
equipped with the electronic toll collection (ETC) devices.
X X X Due to the ETC use rate, and additional DSRC detectors
sensors
were installed to collect travel time information between
(1) Classification of the vehicles into two classes: regular cars and lorries,
from vehicles volume and estimated length; not compatible with axels
location points. Automatic video processing cameras and
classification methods. microwave radars are now under deployment to dense the
sensory infrastructure, as shown in the Figure 4.
First conclusion from methods and technologies comparison
is the highest complementarity between solutions, in terms Behind roadway telematics infrastructures and systems,
of functionalities and data gathered from transportation Brisa motorways, including A5, are fully equipped with a
system. private high-speed fiber-optic cabling and wireless solutions
to enable real-time remote monitoring of the network
III. CASE STUDY AREA conditions, but also, traffic management, tolling operations
and enforcement.
For the experimental analysis of this research study, we use
the A5 motorway, a 25 km inter-urban highway connecting TABLE II
Lisbon and Cascais in the West-cost of Portugal, presented
TELEMATICS SYSTEMS INSTALLED IN THE CASE STUDY
in Figure 3. The first stretch of this motorway, linking
AREA
Lisbon to the National Stadium in Oeiras, with 8 km length,
opened to the traffic in 1944 and became the first motorway Operating functions
in Portugal and one of the firsts in the world.
Number of

Management
A5 Main Figures

Traffic data
Telematics system

Monitoring
units

collection

& control
Geometry (km) 25

Tolling
Nodes 14
Ramp Toll gates 40 X X X
64
connections
ETC toll gates 20 X X X
Toll plazas 6
PTZ Video cameras 47 X X
Annual Average
daily traffic 67,200 VMS Variable
9 X
(AADT) message Signs

AADT near Loop detectors 3 X X

135,400
Lisbon
Microwave radars 2 X X
ETC rate 71%
Video processing
4 X X
Light vehicles cameras
93%
rate
Point-to-point DSCR 12 X X
Occurrences
average/day Roadway telematic infrastructures are then connected to
Brisa’s traffic control center (TCC) where, a state-of-the-art
Incidents 28
suite of systems and applications corporate an advanced
Accidents 4 traffic management and information system (ATMIS),
Obstructions/
providing 24-hour automated electronic tolling, traffic
6 monitoring, traveler information and decision support
Lane closures
services for traffic control and management. At ATMIS
level, all data is managed together and, even though the
Figure 3: The Portuguese A5 inter-urban tolled motorway main figures
specific use for specific functions, such as toll tax charging,
A5 motorway is widely equipped with telematics systems it is available for multiple purpose applications, from traffic
for tolling and traffic management and control. It includes a management to road maintenance and planning.
variety of sensors for traffic monitoring, for the most part This work is part of a research project to generate real-time
used in a multi-purpose way, according to the Table II. anticipatory short-term traffic conditions to be integrated
Primarily telematics installations on the roadway regarded with traffic management and traveler information services,
toll collection systems for open tolling service, where tolls both for recurrent and non-recurrent, unpredictable, incident-
fees are levied at certain points on the highway, once on the based conditions. Although the primarily A5 motorway
main carriageway and other at interchanges. Since 1991, the application, success results will be applied to the complete
high acceptance of Via Verde – the national-wide electronic network.

218
 Sintra
Amadora

Cascais Lisboa

Oeiras

Loop detectors Toll plaza Microwave radars Video processing Point-to-Point DSRC

Figure 4: A5 motorway sensory infrastructure

5 shows the correlation relationship between FCD average

IV. DATA PREPROCESSING AND CLEANING
speed and flow rate in the A5 corridor.
Real-world raw sensor data, technology independently, is
highly susceptible to noise, missing values, and inconsistent 160
data due to sensor failures, measurement errors, and data 140

Average speed (km/h)

link errors or simply because of their huge size. Low-quality 120
100
data will lead to low-quality data processing and outcomes,
80
whichever downstream traffic data application is to run. 60
Furthermore, due to measurement errors derived data, such 40
as average speeds or vehicles counts, may have values that 20
are physically impossible such as negative volume counts or 0
vehicles speed, unless wrong-way driving incidents. 0 500 1000 1500 2000 2500 3000

However, this straightforward approach to improve data Traffic flow rate (veh/h)
quality becomes ineffective in case of road incidents. For
this, extended algorithms for pattern analysis focus precisely Figure 5: A scatter plot for FCD average speed versus Point flow rate in
on the identification of unexpected significant variations so the A5 corridor
called incident-affected data or outliers, either in data
measured or predicted. Consistent values leads to identify Both central tendency and dispersion properties are obtained
outlying events, correspondent to roadway incidents. This as distributive functions, calculated by partitioning the
automatic process for incident detection is extremely useful dataset into smaller subsets and computing measures for
for network operations, to proceed with automatic responses each subset. The global and unique measure is obtained by
and control. However, new incidents conduct to unexpected merging results.
traffic patterns commonly not found in the historical The end result of this process step is a package of discrete
database. Therefore, data preprocessing can be used as a characteristics about traffic data, either for a short sliding
parallel process, valid and useful until an unpredicted window reaching real-time timeline or for the entire dataset
scenario event occurs and disrupt traffic conditions. for analysis, such as, the complete day. The spatial domain
In a short definition, data preprocessing is a combination of those characteristics is also stretchy and can be applied to
chain of techniques to be applied to improve data quality a single measure site, to a set of sites defined individually or
through the completeness, consistency and simplification of within a corridor, or to the network as a whole.
datasets. However, for some following pre-processing steps and for
A. Data properties summarization the some applications further on, it is essential to know how
current traffic conditions compare with past periods in the
To be successful and effective with real-time data historical database. Once again the time and space scale to
preprocessing it is essential to have a comprehensive, overall find out the reference period, depends on the application
picture of existing datasets. It can be based on summarized goal.
representative data properties, including highlights of data
values to be treated as noise or outliers. [Han, 2006] The B. Data cleaning – Missing values
technique is to understand the distribution of the data based Most real-time applications, such as monitoring, simulation,
on descriptive statistics, regarding both central tendency and require complete datasets without missing values. Data
dispersion of data. It includes mean, median, mode and cleaning techniques attempt to fill in missing values, smooth
midrange as measures of central tendency and quartiles, out noise while identifying outliers, and correct other
inter-quartile range and variance for data dispersion. Figure inconsistencies in the data. Several methods endeavor to

219
complete missing gaps dynamically, implementing A basic assumption is to consider that the evolution of traffic
estimation strategies balanced between data accuracy and patterns on a given day of the week is the same as the
computational complexity. Beyond this technical approach evolution of the traffic pattern on the corresponding day in a
to estimate wanting data, there is an essential variable prior reference week, constructed as a moving average over
to the strategy definition: the gap size. several weeks in the past. [Bellemans, 2000] This
construction requires dealing with scenarios such as
Taking into account the typical interval for roadside data
“special” days or official holidays on a weekday and days
acquisition and aggregation, varying from one to 5 minutes,
with major events where traffic patterns may differ
a missing gap up to 15 minutes, corresponding to a one-step

regression analysis, involving a response traffic value,
, and For instance, to estimate a missing value
for a time step 
significantly from regular days.
iteration bootstrapping, can be estimated with a straight-line

a single predictor time-based period, . That is, in the current day , we use the following equation.

=  + 
  − 1

  = ×
 

  − 1
(1)
(5)

Where
  is the corresponding data value at the time 

where b and w are regression coefficients, thought as

in the reference day . Such matching value is then scaled

weights to be solved by the method of least squares. So that
we can equivalently write,

=  +   (2) to the traffic value intensity that precedes the missing value,

  − 1
using the factor (6).

values  ,
 ,  ,
 , … || ,
|| ,

Let be the set of existing time-based observation

  − 1
the regression (6)
coefficients can be estimated using the method with the
following equations,
∑||  − ̅
 −
̅
The application of this method is shown with the chart in the

 = 
Figure 7, where missing values are estimated through the

∑  − ̅ 
||
(3) interpolation process (5) using values from the reference day
22.Jan - the regular weekday from previous week and the
correction factor.

 =
̅ −  ̅ (4) 120

where ̅ and
̅ are the respective mean values of existing

100
Average speed (km/h)

dataset . Figure 6 shows an implementation example of 80

missing values estimation using linear regression method, as
60
described in the equations (1) to (4).
40
120
20
Average speed (km/h)

100

80 0
7:30

7:40

7:50

8:00

8:10

8:20

8:30

8:40

8:50

9:00

9:10

9:20

9:30
60

40 January 29th 2010 - Carcavelos/Oeiras Service Area

20
22.Jan 29.Jan with missing gap
0
29.Jan estimated values 29.Jan complete data
7:30

7:40

7:50

8:00

8:10

8:20

8:30

8:40

8:50

9:00

9:10

9:20

9:30

January 29th 2010 - Carcavelos/Oeiras Service Area Figure 7: Missing data estimation for a 30 minutes missing gap using
reference data interpolation
29.Jan with missing gap 29.Jan estimated values 29.Jan complete data
The big advantage of this method is the simplicity that leads
Figure 6: Missing data estimation using linear regression to ease implementation and computation, useful for real-time
applications. However it fails thoroughly when traffic
For larger missing data gaps, (e.g. 45 to 120 minutes), but patterns suddenly changes because of any roadway incident
reasonable to be a punctual system or communications or demand fluctuations. To handle that situation, keeping the
failure, data completion can be estimated using historical low complexity, we propose a statistical-based analysis to
data in combination with the most recent observations. For establish dynamically an effective connection supported by
the present research work, gap sizes over 120 minutes are evolutionary traffic patterns.
considered input data failure, not able to be estimated from
Space-based similarity search
online and historical data sets.
This new process is based in time-series data analysis and
aims to find the most similar traffic pattern to the real-time

220
sequence pattern, or trail, to be used as reference data to From the real-time traffic trail, this function aims to discover
complete missing values gaps. Because of the heterogeneity in the historical database, the most similar trail to be defined
and complexity of the network of sensors and traffic as the reference trail. The similarity between two trail
measure database, we implement an experience-based structures can be measured by the normalized root mean

0 = {
& ,
&' ,
&' , … ,
&'( } where
 are the discrete

heuristic to optimize patterns recognition and matching. The square deviation (NRMSD). Let the real-time traffic trail be

values for the 9-th element, and let the historical set of traffic
strategy starts with a space-time based correlation to identify

trails for a reference element obtained in (9), be 0< where

space-time connections among network elements. The

each 0<= is >-th trail index and the
<= is the counterpart 9-
objective of this first step is to elect a representative pair of
elements, to be used on the next step.
th element.
To summarize such linear connections between elements

between the real-time trail 0 and historical trail set 0< ,

numerically is used the mean and standard deviation of each Then we can find the minimum value of the NRMSD

Pearson correlation coefficient. Let  and be a pair of !

variable-element separately plus a measure known as the
applying the following formulas,

interval ",  − !#,  = { %& , %&' , %&' , … , %&'( } and =

1
(
observations from two independent data sources, at time

{ *& , *&' , *&' , … , *&'( } , the correlation coefficient + is +?@ 0, 0<= = A B‖
 −
<= ‖
!
(10)

∑( % − %̅ * − *,
given by the formula

+ , = over all selected historical trails, where ‖ . ‖ denotes the

-∑( % − %̅  ∑( * − *, 
(7)
norm value. To normalize the deviation or error found in

from the covariance ./01.3 normalization, by division it by computed values
DEF −
D( ,
In other words, the coefficient of correlation is obtained terms of percentage, RMSD is then divided by the range of

the standard deviation 41 × 43 of each variable element.

From the correlation coefficient + we obtain the coefficient +?@ 0, 0<=
of determination + , by squaring +. The magnitude of the G+?@ 0, 0<= =

DEF −
D(
(11)
coefficient of determination + indicated the proportion of

Hence, the most similar historical trail 0<HI is obtained as

variance in one variable-element, explained from the

+ =+
knowledge of the second variable-element.

trail 0 and the historical trail 0<= ,
(8) the minimum value of the NRMSD between the real-time

0<HI = min G+?@ 0, 0<=

Finally we calculate the amount of variation that the two (12)

From here 0<HI will be used as the reference data set to

variable-elements have in common - the proportion or

coefficient for free – the coefficient of alienation 1 − + , complete missing values in 0.

percentage of shared variance. There is also the opposite

which defines the measure of non-association or variation

120

5ℎ78
797!:8 = + × 100

between such two variables.
100
(9)
Average speed (km/h)

80
Therefore, shared variance is the variance accounted for in
one element by another element. For each sensor or data 60
source of the network, this space-based similarity process
40
establishes a reference element, to be used in the following
processing steps. 20

Time-based similarity search 0

7:30

7:40

7:50

8:00

8:10

8:20

8:30

8:40

8:50

9:00

9:10

9:20

9:30

Traffic conditions can be stated as an evolution chain of

values, sequence or trail, established from plain measures, January 29th 2010 - Carcavelos/Oeiras Service Area
summarized indicators or from a complex, highly elaborated

for the discrete time  . The set of values, behavior as time

gauge. Each value is the representative of the traffic network 20.Jan 29.Jan with missing gap
29.jan estimated values 29.Jan complete data

the network conditions for the time ( , where ! the near

series sliding window, can be defined as the sequence trail of Figure 8: Missing data estimation based on space-time similarity search

real-time index. The trail structure is topologically linear, The application of this method combination is shown in the
which can be represented by a vector of measures in the Figure 8. For the 29th of January, missing values in the
common timeline. sensor point-to-point Carcavelos/Oeiras Service Area, in the
time period 8:20 to 8:50 AM, are estimated using data from

221
20th of January, found in the space-time similarity search. the middle of the distribution in either direction”. They may

data is 0<HI .
Ultimate values are applying factor (6) where the reference be due to sensor noise, acquisition process instability,
equipment degradation, computer or communication system
TABLE III fails, or human-related errors. However, with some
applications, abrupt changes in the acquisition field may
RESULTS COMPARISON FOR 40 MINUTES MISSING GAP occur and cause fluctuations in upcoming observations from
Method RMSD the bulk of values. In case of traffic data measurement, such
Linear regression 13.05 sudden changes are usually related with traffic congestion
Interpolation with reference data 18.25 caused by accidents, broken vehicles or any other type of
Space-time similarity search 12.40 incident. For real-time applications, based on online data
collection and processing, it is crucial to assure data quality
As presented in the Table III, for short missing values through the identification and isolation of outliers.
periods, both linear regression and space-time similarity In this research work, automated detection of outliers and
search are satisfactory to estimate missing values. This way
removal were developed and integrated with automatic
promotes the usage of the linear regression, due the low
incident detection, in order to preserve all data, including
computational complexity and easy implementation.
such apparently unreliable data. Therefore, this process is
Next we process and present estimation values for a large made up of two distinct functions: i) identify, isolate and
missing gap, for the same reference day. replace outliers; and ii) automatic data-driven incident
120
detection. In this paper and section we focus on the first
function, in order to support real-time data applications.
Average speed (km/h)

100
80 The generation of outliers can be described by the time-
series process analysis additive outlier model [Martin, 1986].
M= = N= + O=
60
40
(13)

Where M= are the observations data sequence, N= is the

20

expected data sequence and O= represents a sequence of

0

contaminating outliers. The expected sequence N= is based

7:30

7:40

7:50

8:00

8:10

8:20

8:30

8:40

8:50

9:00

9:10

9:20

9:30

strictly on the current and past data observations M='P

January 29th 2010 - Carcavelos/Oeiras Service Area

for > ≥ R ≥ 0, and the median value MS. For a data point %,
the distance T is
29.Jan with missing gap
Linear regression

T = U MT − MST U
Interpolation with reference data
(14)

Is the distance T exceeds some specific threshold VT > 0,

Interpolation with space-time similarity

and then we declare MT to be an outlier and replace it with an

29.Jan complete data

Figure 9: Missing values estimation methods comparison

Using the same package of techniques and methods, we estimated value obtained from the process to complete
produced a 90 minutes missing values gap for the 29th of missing values, defined previously.
January in the same sensor. Figure 9 presents the estimation
100 30
values results graphical comparison, using precisely the
28
same reference data sets. 90
26
Average speed (km/h)

TABLE IV 80 24

Distance dp
70 22
RESULTS COMPARISON FOR 90 MINUTES MISSING GAP
20
60 18
Method RMSD
Linear regression 21.03 50 16
Interpolation with reference data 29.81 14
40
Space-time similarity search 9.16 12
30 10
0:20
0:40
1:00
1:20
1:40
2:00
2:20
2:40
3:00
3:20
3:40
4:00
4:20
4:40
5:00
5:20
5:40
6:00
6:20
6:40
7:00

As shown in the Table IV, for large missing values periods,

space-time similarity search with NRMSD is definitively the
most appropriated method to complete dataset. measured values distance dp OUTLIERS

Figure 10: Automated outliers detection and replacement process

C. Data cleaning – Outliers
Outlying observations are measures that numerically deviate The method application is shown in the Figure 10 with the
radically from the rest of the data. [Mendenhall, 1993] variable average speed in the sensor point-to-point
defines the term “outliers” to values “that lie very far from Carcavelos/Oeiras Service Area for the 29th of January.

222
D. Data reduction will lead to low-quality results processing. For real-time
For conventional surveillance systems, raw data is gathered traffic data applications this postulation is even more
from a multi-variety of sensing methods either installed on significant, since low-quality information for decision
the network operators’ infrastructure or on fleeting vehicles, support will lead to incompetent control and management.
as previously described. This increasing array of data Data preprocessing and cleaning defines a set of techniques
sources leads to an increasing difficulty to accomplish and methods to analyze databases, identify errors and
desired results by making use of the whole data. [Huang et inconsistencies and proceed with dataset correction,
al., 2009] Some of the main reasons for that relies on the completion and simplification.
intrinsic characteristics of data sources and types: i) Are This paper presents a suite of combination methods to
deployed with uneven density over the network; 2) Are analyze and summarize real-time traffic data sets, to estimate
missing values and to identify and correct outliers in
heterogeneous in type; 3) Provide highly correlated data; 4)
datasets. Either used separately in the simplest way for short
Report at non-uniformed resolution; 5) Report at different
missing gaps or in a complex way, combining several
frequencies. Data reduction techniques can be applied to methods and time-space based historical datasets, data
work out some of those difficulties by harmonizing data preprocessing techniques aims to improve data quality
references and dimensions, and bringing down the size and through the harmonization and completeness.
complexity of datasets, maintaining the integrity of the
original data. Our contribution is the designing and implementation of a
In the present work we design a two-tier approach for data systematic methodology to measure, check and repair traffic
reduction: i) at the data acquisition level; and ii) at data data in order to enable following steps in the processing
fusion level. For the first tier, data acquisition process chain till decision making and transfer. With this approach,
computes raw or elementary data and, along with events facing data problems in early stage, including data
detection, proceeds with data aggregation and simplification, reduction and integration, we promote the
summarization per regular periods, varying in space, time multi-purpose usage of traffic data. However, the key
and measure type. For the second tier, heterogeneous data advantage of our process definition goes for real-time data
sets are merged together do obtain a single data platform, applications, used to manage primitive data with noise and
and is defined as data integration and fusion. errors inside. Future works in this research program will take
Data fusion is the process of merging together information advantage of this work, and will focus data fusion and
gathered from various heterogeneous sensors, into a single decision making processes.
data platform. In space-time domain, such as traffic field,
data fusion is synonymous with data integration and aims to ACKNOWLEDGMENT
combine diverse data sets into a unified, or fused, data set This research work is supported by Brisa Auto-Estradas de
which includes all of the data points and time frames from Portugal, a leading world toll motorway operator and
the input data sets. The resulted data set differs from a transportation infrastructures manager. Brisa is also member
simple merged superset in that data tuples contains attributes of MIT-Portugal research program and, with an active
and metadata which might not have been included for these collaboration of Brisa Innovation research group and the
points in the original data set. MIT ITS lab, to develop this work.
TABLE V REFERENCES
[Antoniou et al., 2008] Antoniou, C., Balakrishna, R., and Koutsopoulos,
TRAFFIC DATA, INFORMATION AND KNOWLEDGE H. N. (2008). Emerging data collection technologies and their impact on
Data traffic management applications. Proceedings of the 10th International
Knowledge
acquisition Conference on Application of Advanced Technologies in Transportation,
Data Fusion and decision
and Athens, Greece.
making
preprocessing [Bellemans, 2000] Bellemans, T., Schutter, B., Moor, B. (2000). Data
Elementary Situation acquisition, interfacing and pre-processing of highway traffic data.
Object data Response
data information Proceedings of Telematics Automotive 2000, Birmingham, UK, vol. 1, pp
- Sensor data - Point speed - Road point - Traffic 4/1-4/7, Apr. 2000
- Tolling data - Point flow conditions control [Han, 2006] Han, J., Kamber, M. (2006). Data Mining Concepts and
- Road - OD travel - Link conditions - Driver Techniques (2nd Edition). Elsevier, Morgan Kaufmann Publishers.
segment data times - OD conditions warning [Huang et al., 2009] Huang, E., Antoniou, C., Ben-Akiva, M., Lopes, J.,
- GPS data - OD flows - Driver-choice - Congestion Bento, J. (2009). Real-time multi-sensor multi-source network data fusion
- Cell-phone - Class-relative options pricing using dynamic traffic assignment models. In proceedings to the 12th
data density - Incidents - Maintenance International IEEE Conference on Intelligent Transportation Systems.
- (…) - (…) - (…) - (…) [Huang, 2010] Huang, E. (2010). Algorithmic and Implementation Aspects
Table V presents a data architecture overview, concerning of Online Calibration of Dynamic Traffic Assignment. Master's thesis,
Massachusetts Institute of Technology.
the evaluation chain from elementary data to knowledge for [Liu, 2004] Liu, H., Shah, S., Jiang, W. (2004). On-line outlier detection
decision making support. and data cleaning. Journal of Computers and chemical engineering. Pages
1635-1647. Elsevier
[Klein, 2006] Klein, L., Mills, M., Gibson, D. (2006). Traffic Detector
Handbook: Third Edition—Volume I. Report No. FHWA-HRT-06-108.
V. DISCUSSION AND FUTURE WORKS Federal Highway Administration, USA
Real-world traffic databases are highly susceptible to noise, [Martin, 1986] Martin R., Yohai, V. (1986). Influence Functionals for Time
redundancy and inconsistent data due to their typically huge Series. Ann. Statist., Volume 14, Number 3
size and their likely origin from multiple, heterogeneous [Mendenhall, 1993] Mendenhall, W., Reinmuth, J., Beaver, R. (1993).
sources and sensory technologies. Low-quality traffic data Statistics for Management and Economics. Belmont, Duxbury Press

223