Content uploaded by Vesal Ahsani
Author content
All content in this area was uploaded by Vesal Ahsani on Dec 10, 2018
Content may be subject to copyright.
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=gits20
Journal of Intelligent Transportation Systems
Technology, Planning, and Operations
ISSN: 1547-2450 (Print) 1547-2442 (Online) Journal homepage: http://www.tandfonline.com/loi/gits20
Quantitative analysis of probe data
characteristics: Coverage, speed bias and
congestion detection precision
Vesal Ahsani, Mostafa Amin-Naseri, Skylar Knickerbocker & Anuj Sharma
To cite this article: Vesal Ahsani, Mostafa Amin-Naseri, Skylar Knickerbocker & Anuj
Sharma (2018): Quantitative analysis of probe data characteristics: Coverage, speed bias
and congestion detection precision, Journal of Intelligent Transportation Systems, DOI:
10.1080/15472450.2018.1502667
To link to this article: https://doi.org/10.1080/15472450.2018.1502667
Published online: 17 Oct 2018.
Submit your article to this journal
Article views: 68
View Crossmark data
Quantitative analysis of probe data characteristics: Coverage, speed
bias and congestion detection precision
Vesal Ahsani
a
, Mostafa Amin-Naseri
b
, Skylar Knickerbocker
c
, and Anuj Sharma
a
a
Department of Civil, Construction and Environmental Engineering, Iowa State University, Ames, Iowa, USA;
b
Department of Industrial
and Manufacturing Systems Engineering, Iowa State University, Ames, Iowa, USA;
c
Institute for Transportation, Iowa State University,
Ames, Iowa, USA
ABSTRACT
In recent years, there has been a growing desire for the use of probe vehicle technology
for congestion detection and general infrastructure performance assessment. Unlike costly
traditional data collection by loop detectors, wide area detection using probe-based traffic
data is significantly different in terms of the nature of data collection, measurement tech-
nique, coverage, pricing, and so on. Although many researches have studied probe-based
data, there remains critical questions such as data coverage and penetration over time, or
the influential factors in the accuracy of probe data. This research studied probe-sourced
data from INRIX, to profoundly explore some of these questions. First, to explore coverage
and penetration, INRIX real-time data was illustrated temporally over the entire state of
Iowa, demonstrating the growth in real-time data over a 4-year timespan. Furthermore, the
availability of INRIX real-time and historical data based on type of road and time of day,
were explored. Second, a comparison was made with Wavetronix smart sensors, commonly
used sensors in traffic management, to explore INRIX’s speed data quality. A statistical ana-
lysis on the behavior of INRIX speed bias, identified some of the influential factors in defin-
ing the magnitude of speed bias. Finally, the accuracy and reliability of INRIX for congestion
detection purposes was investigated based on the road segment characteristics and the
congestion type. Overall, this work sheds light onto some of the less explored aspects of
INRIX probe-based data to help traffic managers and decision makers in better understand-
ing this source of data and any resultant analyses.
ARTICLE HISTORY
Received 30 November 2017
Revised 16 July 2018
Accepted 17 July 2018
KEYWORDS
congestion detection;
coverage; probe data;
sensor data; speed
bias analysis
Introduction
Many transportation agencies and state Departments
of Transportation (DOT) utilize fixed, infrastructure –
mounted sensors for collecting relatively accurate and
real-time traffic information such as lane by lane traf-
fic speed, volume, occupancy, and so on. Compared
to alternatives provided by most nontraditional
data streaming sources, the cost of deploying and
maintaining these sensors could be high. Another
limitation of sensors is their geographical scalability;
they need to be installed in a large number to deter-
mine the traffic situation in an area (Young, 2007).
Accordingly, most Traffic Management Centres
(TMC) install them on major freeways and critical
urbanized areas rather than throughout the highways
and arterials. The lack of sufficient coverage on high-
ways and arterials spurs the interest to augment
infrastructure mounted sensors with new data stream-
ing sources.
With the rapid rise of telecommunication and
wireless technologies over the past few years, traffic
data collection, processing, analyses and utilization
have changed significantly. Wide area probe techno-
logy is an example which collects traffic information
from millions of mobile devices, connected cars,
trucks, delivery vans, and other GPS-enabled fleet
vehicles. Probe-based methods of measuring travel
time and speed data can easily scale across large net-
works without the need for deploying any additional
infrastructure (Young, 2007). This makes several
agencies to use a single, uniform source of data
as a cost-effective way for monitoring traffic across
most roadways in a State (FHWA, 2013). Some of the
third-party probe data providers are INRIX, HERE,
TomTom, and so on.
CONTACT Vesal Ahsani ahsaniv@iastate.edu Department of Civil, Construction and Environmental Engineering, Iowa State University, Ames,
IA 50011.
Color versions of one or more of the figures in the article can be found online at www.tandfonline.com/gits.
ß2018 Taylor & Francis Group, LLC
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS
https://doi.org/10.1080/15472450.2018.1502667
Several studies have been conducted to compare
the accuracy and reliability of probe sourced data
against local sensor data such as radar sensor data,
loop detector data, and so on (Adu-Gyamfi, Sharma,
Knickerbocker, Hawkins, & Jackson, 2017; Coifman,
2002; FDOT, 2012; Feng, Bigazzi, Kothuri, & Bertini,
2010; Haghani, Hamedi, & Sadabadi, 2009; Kim &
Coifman, 2014; Lindveld, Thijs, Bovy, & der Zijpp,
2000). Many of them evaluated performance of probe
data by travel time reliability measures, such as the
90th or 95th percentile of travel time, the standard
deviation, the coefficient of variation, the percentage
of variation, the buffer index, the planning time
index, the travel time index, congestion hour, and so
on. (Aliari & Haghani, 2012; Araghi, Hammershøj
Olesen, Krishnan, Tørholm Christensen, & Lahrmann,
2015; Chakraborty et al., 2018; Cookson & Pishue,
2016; Day et al., 2015; FHWA, 2017; Gong & Fan,
2017; Higatani et al., 2009; Hu et al., 2016; Lomax,
Schrank, Turner, & Margiotta, 2003; Miwa, Ishiguro,
Yamamoto, & Morikawa, 2015; MoDOT, 2017; Pu,
2011; Rakha, El-Shawarby, & Arafeh, 2010; Remias
et al., 2013; Sanaullah, Quddus, & Enoch, 2016;
Schrank, Eisele, Lomax, & Bak, 2015; Schrank, Eisele,
& Lomax, 2012; Sekuła, Markovi
c, Laan, & Sadabadi,
2017; Sharifi et al., 2017; Turner, 2013; Uno,
Kurauchi, Tamura, & Iida, 2009; Venkatanarayana,
2017; WSDOT, 2013,2014; Zheng, Li, van Zuylen,
Liu, & Yang, 2018). An overview of these studies and
the performance measures used to evaluate travel time
reliability of probe-source data is provided in Table 1.
In addition to studies on INRIX travel time reliabil-
ity, more recent studies have been conducted using
INRIX data to evaluate other aspects of INRIX data.
For instance, Eshragh and colleagues estimate the
accuracy of probe speed data on arterial corridors
using roadway geometric attributes (Eshragh, Young,
Sharifi, Hamedi, & Sadabadi, 2017). It was also shown
that INRIX and benchmarked results were most simi-
lar in external-external trips (Hard et al., 2017). Also,
Lu and Dong compared INRIX with radar sensor data
for travel time estimation and showed its reliability
(Lu & Dong, 2018). Moreover, models were developed
for detecting abnormal traffic patterns and traffic
speed prediction using INRIX and Wavetronix data
sets, and obtained satisfactory results (Barajas, Wang,
Kaiser, & Zhu, 2017). Day and colleagues made use
of INRIX XD and connected vehicle data to optimize
traffic signal offsets (Day et al., 2017). Additionally,
(Elhenawy, Chen, & Rakha, 2014) examined the qual-
ity of INRIX data and showed its good quality for
travel time prediction.
Overall, the recent studies have reaffirmed the validity
and value of INRIX data while pointing out improve-
ment in its quality over years. The quality improvement
is shown in Figure 2 of the paper. The figure shows a
significant increase in the real-time data availability.
Moreover, based on the report performed by the
University of Maryland and published by INRIX, INRIX
was never penalized for data quality during the life of
vehicle probe project (VPP). This report mentioned 57%
and 46% improvement of INRIX speed error results in
heavy and moderate congestion from 2008–2009 to
2012–2013, respectively. Moreover, 87% overall improve-
ment was observed in INRIX speed bias results from
2008 to 2013 (INRIX, 2015).
Inversely, very few research has been conducted on
probe data coverage, probe data penetration over
time, speed bias and congestion detection perform-
ance with respect to segment’s characteristics. This is
while probe data, unlike sensor data, comes from an
ever changing source, thus making it critical to study
the patterns and trends in coverage and penetration.
To the best of our knowledge, no other research has
been looked into the coverage of probe-sourced data
temporally over a 4-year timespan.
In terms of INRIX speed quality, several works have
estimated the speed bias to be 6 mph on freeways relative
to ground truth (Lattimer & Glotzbach, 2012)andmore
generally, the overall average speed errors were estimated
to be within 10 mph throughout various levels of conges-
tion (Kim, Motiani, Spasovic, Dimitrijevic, & Chien,
2014). Adu-Gyamfi (2017) also noted that high speeds
(>60 mph) generally have less error, whereas low speeds
(<60 mph) show higher speed error. Table 2 below
includes positive and negative results these mentioned
papers have been concluded. Despite the invaluable
information that these works provide, yet a quantitative
analysis that studies the significant factors influencing
speed bias is not in place. Similarly, a quantitative study
on the factors (e.g., segment length or congestion type)
that influence the congestion detection quality using
probe-based data, have not been performed. This work
studied INRIX data to learn more about this source
of data from these less considered perspectives.
Data
The different sources of data utilized in this work, are
explained in this section.
Probe-sourced data
With the help of today’s technologies including
connected vehicles and smartphones, INRIX leverages
2 V. AHSANI ET AL.
the great amount of historical and real-time data
which can be analyzed and investigated to improve
transportation networks performance. This study
utilized the historical and real-time traffic data col-
lected through the INRIX TMC monitoring platform.
For each of the TMC segments, the speed, as well
as the corresponding date and time of traverse were
provided for every 1 minute.
Infrastructure-mounted sensors
The benchmark dataset used in this paper was
obtained from Wavetronix sensors which utilizes
radar technologies for data collection. Although we
acknowledge that sensors might have some inherent
errors, yet Wavetronix Smart Sensors have been
commonly utilized as ground truth for comparison pur-
poses (Chakraborty, Adu-Gyamfi, Poddar, & Ahsani,
2018;Luetal.,2014; Poddar, Ozcan, Chakraborty, &
Ahsani, 2018;Sharifi,etal.,2011). Each Wavetronix
sensor unit is built up of a Doppler radar, a wireless
modem, solar panel and on-board processors for real-
time processing of traffic data such as speed, volume,
etc. High-resolution (20 second) traffic speed data were
provided by Wavetronix sensors.
Roadway asset management system (RAMS)
Information on the roadway geometry and characteris-
tics can play an important role in studying performance
Table 1. Overview of the studies and the performance measures used to evaluate travel time reliability
of probe-source data.
Study Source of probe data used Performance measures
Pu (2011) Not mentioned 95th percentile travel time, standard deviation,
coefficient of variation, percent variation, skew
statistic buffer index (w.r.t. average), buffer
index (w.r.t. median), planning time index, fre-
quency of congestion, failure rate (w.r.t. aver-
age), failure rate (w.r.t. median), travel
time index
Lomax et al. (2003) Not mentioned Travel time window, percent variation, variability
index, displaying variation, buffer time, buffer
time index, planning time index, travel rate
envelope, on-time arrival, misery index
Turner (2013) INRIX Annual hours of delay per mile, hours of target
delay per mile, Travel Time Index, Planning
Time Index, top N congested segments
Uno et al. (2009) Not mentioned Average travel time, covariance of travel time, level
of service (LOS)
Rakha et al. (2010) Not mentioned Travel time coefficient of variation
Day et al. (2015); Remias et al. (2013) INRIX Congestion hours, distance-weighted congestion
hours, congestion index, speed profile, speed
deficit, travel time deficit, congestion cost, top
N bottlenecks
MoDOT (2017) Not mentioned Average travel time per 10 miles, additional travel
time needed for on-time arrival (80% of time),
annual congestion costs
FHWA (2017) NPMRDS Congested hours, planning time index, travel
time index
Schrank et al. (2012;2015) INRIX Travel speed, travel delay, annual person delay,
annual delay per auto commuter, total peak
period travel time, travel time index, planning
time index, number of rush hours, percent of
daily and peak travel in congested conditions,
percent of congested travel
WSDOT (2013,2014) Not mentioned Lane-miles congested, total and cost of delay,
travel time index
Sharma, Ahsani, and Rawat (2017) INRIX Congestion detection latency, count of congestion,
congestion durations, buffer time index, reliabil-
ity curve
Hu et al. (2016) INRIX Delay saving, buffer index, 95th percentile
travel time
Cookson and Pishue (2016) INRIX INRIX travel time index, wasted time in congestion
Aliari and Haghani (2012) INRIX Travel time, average speed
Gong and Fan (2017) INRIX Travel time reliability, planning time index, fre-
quency of congestion
Sekuła et al. (2017) INRIX Hourly traffic volume
Venkatanarayana (2017) INRIX, NPMRDS Traffic delay, planning time index, travel time
index, AASHTO reliability indexes (RI80, for all
days and weekdays), congested hours, and con-
gested miles
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 3
of transportation networks. Iowa DOT’s Roadway
Asset Management System (RAMS) provides an
inventory of roadway geometry and characteristics for
the entire state. Information included in the database
include the number of lanes, speed limit, AADT and
surface type, as well as other information that may be
useful when building a model. The RAMS database
uses the DOT’s linear referencing system which can
be used when requesting information for a specific
location. The coordinates for each TMC event were
passed through the linear referencing systems REST
services to provide the corresponding route and mile-
age values. The route and mileage could again be used
with the REST services to retrieve the roadway
Table 2. Summary of the findings on INRIX speed bias analysis.
Study Source of probe data used Performance measures Pros Cons
FDOT (2012) NAVTEQ, TrafficCast,
INRIX
Absolute average speed error,
average speed bias, abso-
lute average travel time
error, travel time bias
1. All probe data sources are
generally consistent with
the ground truth data.
2. INRIX data in some cases
appeared to be more
accurate compared to other
probe datasets
1. TMC segments in urban
areas with traffic signals
experienced a larger vari-
ability in the results.
Haghani et al. (2009) INRIX Speed error, speed
error bias
1. Speed data provided by
INRIX is generally of
good quality
1. Segments with length less
than one mile are in-accur-
ate.
2. Different confidence scores
30, 20, and 10 are not sig-
nificant indicator of INRIX
data quality.
3. For speeds below 45 mph,
INRIX overestimates the
speeds and for speeds over
60 mph, it underestimates
the actual speed
Lattimer and Glotzbach (2012) INRIX, NAVTEQ,
TrafficCast
Travel time, Speed bias –1. INRIX speed has a 6 mph
bias relative to ground
truth on an uncon-
gested freeway
Kim et al., (2014) INRIX, NAVTEQ,
TrafficCast
Travel speed, Speed error –1. Overall average speed
errors tend to be within
10 mph throughout various
levels of congestion.
2. Data providers missed a
major incident lasting more
than 4 hours.
Adu-Gyamfi et al. (2017) INRIX Speed bias, latency,
similarity index
1. Probe data is reliable for
monitoring the perform-
ance of transportation
infrastructure over time.
2. Latency on freeways is less
than non-freeways
1. Various levels of amplitude
bias between INRIX and
benchmarked data
Table 3. Summary of the INRIX evaluation procedure steps.
Step Name Research motivation
Data
Time Location
1 Coverage Temporal distribution of INRIX real-time data (score 30) Whole 4 years of 2013, …, 2016 Entire state of Iowa
Availability of INRIX real-time and historical data
(scores 10, 20, 30)
Road type
Time of day
April 2016 to April 2017 Entire state of Iowa
2 Speed bias Characteristics of INRIX speed bias
Speed value
Time of day
truck-AADT
Number of lanes
Type of TMC Segment
Segment length
April 2016 to April 2017 Des Moines Area (Iowa)
3 Congestion detection Characteristics of congestion detection by INRIX
Type of congestion
Type of TMC segment
Segment length
April 2016 to April 2017 Des Moines Area (Iowa)
4 V. AHSANI ET AL.
geometry and characteristics. This system allows for
the system to be deployed in real-time in the future to
quickly obtain the roadway characteristics. The data
requested for the model using this service are: (1)
AADT, (2) federal functional class, (3) median type
and width, (4) number of lanes, (5) right and left
shoulders type and width, (6) speed limit, (7) surface
type and width, and (8) terrain.
Data stream and pre-processing
Most of the time in real-world scenarios, raw traffic
data are incomplete, highly susceptible to noise, and
inconsistent for many reasons, such as sensor failures,
measurement technique errors, huge data size, etc.
Data pre-processing can be used to try to detect and
correct corrupt and erroneous traffic data. However,
the storage and analysis of massive amounts of INRIX
and Wavetronix requires proper infrastructure and
computational power to handle masses of data. A
high performing cluster was used for data processing.
Hadoop Distributed File System (HDFS) (“Apache
Hadoop,”2017) was used for storage of the data and
map-reduce was used for processing. Pig Latin
(“Apache Pig,”2017) was used as the language to
implement map-reduce algorithms.
Evaluation procedure
Incorporating a probe data stream into traffic opera-
tions, planning, and management activities requires
several key evaluations in the reliability and accuracy
of the probe-sourced data. For this purpose, this study
utilized real-time and historical traffic data which
were collected through two different data sources;
INRIX and Wavetronix. The INRIX probe data stream
is compared to a benchmarked Wavetronix sensor
data source in order to explain some of the challenges
and opportunities associated with using wide area
probe data. In the following, INRIX performance will
be thoroughly evaluated in three major criteria
(Table 3):
1. Coverage and penetration
2. Speed bias
3. Congestion detection
1. Coverage
The most critical consideration in evaluating probe
data is the geographic coverage provided by the
vendor. The quality of probe data is heavily dependent
on the number of probes on the road network. The
more probes on the network, the better the coverage.
In addition to real-time data, INRIX provides histor-
ical data whenever real-time data are not available.
The higher the device penetration (i.e., more probes),
the better the data are. INRIX reports two measures
of confidence, score and value. Based on Interface
Guide for Public Sector Applications from INRIX, for
each speed measurement, INRIX reports a measure of
confidence, reported as one of three possible values
(IowaDOT, 2017):
Score 30: speed estimate for that segment based
completely on real-time data (the highest confi-
dence score),
Score 20: speed estimate based on real-time data
across multiple segments and/or based on a com-
bination of expected and real-time data, and
Score 10: speed estimate based primarily on his-
toric data (the lowest confidence score).
Additionally, INRIX reports a second measure of
confidence, which it is called the confidence value.
Based on INRIX Interface Guide, the confidence value
is based on a comparison against historical trends. It
should be taken into consideration that the confidence
value only applies when the confidence score is 30.
In Figure 1, the 2016 yearly coverage of INRIX
real-time data (score 30) for interstate and non-inter-
state roadways in the state of Iowa is shown by a
range of colors. Red represents minimum possible
availability of real-time data on roads through green
representing the maximum.
However, the coverage and quality of probe-based
data, due to its nature (being provided by probes), is
not guaranteed to stay constant over time. Thus, it is
critical to monitor the trends in coverage accuracy
over time, a point which has been less considered in
the literature. Figure 2(a) and (b) depict the empirical
cumulative distribution function (CDF) of real-time
INRIX data availability for years 2013–2016 on inter-
states and non-interstates respectively. The INRIX
data are reported every minute for each TMC seg-
ment. It is visually apparent that the percentage of
real-time INRIX data availability on both interstates
and non-interstates for year 2016 is higher than the
prior 3 years. For instance, red arrow in Figure 2(a)
indicates for the 60th percentile of road segments on
the interstates, the availability of real-time INRIX data
was increased from nearly 70% for years of 2013–2015
to almost 90% for year 2016. Moreover, the number
of roads that had no coverage in 2013 to 2015 had
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 5
decreased in 2016, as the number of probes increased.
INRIX has not shared the reason for this significant
increase in the availability of real-time data but our
hypothesis is that additional sources of data were pro-
cured which increased the penetration in Iowa.
Therefore, the further analyses was conducted on
2016 data to capture the most recent characteristics of
INRIX data.
The daily availability of INRIX traffic data is shown
in Figure 3(a) and (b), reflecting how traffic speed
data from interstates and non-interstates are spread
over a span of a full day based on confidence scores
10, 20, and 30. In Figure 3, the INRIX time interval is
considered again as 1-min for the analysis. The hori-
zontal axis shows 1440 minutes of a day. The vertical
axis is the probability of having confidence score 10,
20 or 30 of INRIX data in each minute of a day with
three colors of blue, red, and yellow, respectively. The
probability of having each of the scores in each
minute of a day is computed by considering that spe-
cific minute for all days in 1 year, examine how many
score 10, 20, and 30 were observed in that specific
Figure 2. Temporal empirical CDF of INRIX real-time (score 30) data on (a) Interstates and (b) Non-interstates in the entire state of
Iowa over 4 years of 2013–2016.
Figure 1. Geographical INRIX real-time data availability for the state of Iowa in year 2016.
6 V. AHSANI ET AL.
Figure 3. Daily score-wise availability of INRIX traffic speed data on (a) Interstates and (b) Non-Interstates for whole year of 2016.
Yellow, red, and blue lines represent scores 30, 20, and 10 data, respectively. (a) Interstates, (b) Non-interstates.
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 7
minute over the span of a year. In the analysis, each
specific minute of a day is considered with all corre-
sponding confidence scores (which can only be one of
the values 10, 20, or 30), and calculate number of
times score 10, 20, and 30 were seen in that specific
minute of day over 365 days in a year. In other words,
the summation of probabilities of scores 10, 20, and
30 in each minute always equals to 1. One point
which should be noted here is that this figure does
not show that it is probable to have multiple confi-
dence scores for each minute. According to our ana-
lysis on INRIX data in the year 2016, for example in
Figure 3(a) at the time 4:48 am, it is 84% probable to
have confidence score 30, 0% probable to have score
20, and 16% probable to have score 10. As expected,
INRIX was able to provide real-time speed data (score
30) most of the day on the interstates (Figure 3(a)),
whereas on non-interstates, real-time data were pro-
vided mostly from around 6 am to 6 pm (Figure 3(b)).
Thus, INRIX provides a higher percentage of real-
time data on interstates compared to non-interstates
and the data are more reliable during the day than
the night.
For the further analysis, we focus on performance
on Interstates as the quality of data for Iowa
Interstates was significantly superior to rest of the
network. For this purpose, a specific location includ-
ing a total of 163 TMC segments and Wavetronix
sensors for approximately 164 miles of Iowa primary
network along I-35, I-80, and I-235 near Des Moines
area is selected, as shown in Figure 4. The criterion
for selecting a sensor–segment pair was based on two
main association rules. First, the Wavetronix sensor
should be located in its corresponding INRIX seg-
ment, and, second, the bearing of the road on which
the sensor and segment are located should be the
same. There are several locations in Iowa which one
TMC segment corresponds to multiple sensors.
2. Speed bias
Speed bias is defined as the difference of speed
between the two traffic speed data providers. There is
almost always a speed bias between data streaming
from probes and traditional infrastructure-mounted
sensors. Although part of this difference is inevitable
Figure 4. Location of segments and sensors used.
8 V. AHSANI ET AL.
due to the differences in the two data collection
methods, yet a model that provides insight about the
underlying factors that influence speed bias, would
further the community’s understanding of this probe-
sourced data. Different factors, such as INRIX speed
value, time of the day, the number of probes on road,
road segment type, number of lanes, and so on, can
be influential in the magnitude of probe data speed
bias. A statistical model was used to investigate the
role of these factors.
In this study, speed bias was calculated by subtract-
ing INRIX speed from Wavetronix speed (Equation 1).
Speed bias ¼Wavetronix speedINRIX speed (1)
Probe technology calculates speed as the average
speed of vehicles over a length of road which is called
space mean speed (SMS). Time mean speed (TMS)
which is arithmetic mean of vehicles’speed passing a
point is the calculated speed for Wavetronix sensors.
There is always a difference between SMS and TMS
due to measurement technique. The relationship
between TMS and SMS is shown in Equation 2 below
(Knoop, Hoogendoorn, & Van Zuylen, 2009):
vt¼rM2
vs
þvs(2)
where
v
s
¼SMS,
v
t
¼TMS, and
r
M
¼variance of SMS.
Ideally, TMS to SMS conversions (or vice versa)
should be performed before the two data sets are
compared. However, in this paper speed data obtained
from Wavetronix and INRIX were already aggregated.
For that reason, the speed variance (r
M
) could not be
calculated within the 20-s and 1-min period. However,
a previous study showed that the most probable range
of error introduced by the measurement technique
is between 0 and 2 mph (Adu-Gyamfi et al., 2017).
To further explore the characteristics of speed bias
in INRIX data, a statistical analysis was performed
to quantitatively explore the significant contributors
to INRIX speed bias.
First, INRIX speed values were explored by dividing
observations into five groups (0–25, 25–45, 45–55,
55–65, and greater than 65 mph) and the box plot of
speed bias for each group is depicted in Figure 5(a).
The notable observation in this figure is that as INRIX
speed value increases, two attributes decrease: (1)
the interquartile range (variation), and (2) median of
speed bias. Smaller interquartile ranges with higher
INRIX speeds, indicated there is less variation in speed
bias when INRIX speed is greater than 45mph.
To determine the effect of INRIX speed in the
speed bias level, a one way analysis of variance
(ANOVA) was performed on five predefined groups
of INRIX speed. Based on the ANOVA, the differen-
ces of the mean INRIX speed bias among all groups
of INRIX speed were statistically significant
Fð4;1948120Þ¼12323;p<:0001. Tukey’s post-hoc test
showed that all five groups had statistically significant
differences in the mean speed bias. Figure 5(b) shows
the location of all segment-sensor pairs and their cor-
responding speed biases for each speed group. Range
of speed bias is from –20 (green color) to þ20 (red
color). Yellow and orange colors represent low magni-
tude of speed bias. It reaffirms the fact that speed bias
decreases for most of the sensor-segment pairs as
INRIX speed values increase. It should be noted that
negative speed bias means that INRIX speed value is
more than Wavetronix sensor speed value.
In terms of INRIX speed quality, Haghani et al.
(2009) mentioned that INRIX overestimates speeds
below 45 mph and for speeds over 60 mph, it underes-
timates the actual speed. Our observation in Figure 5
does not comply with their finding. Among the 163
segments and sensors which were used in this paper,
although there were cases where the general under-
standing of speed underestimation and overestimation
were observed, in many cases it was contradicted (i.e.,
INRIX overestimated for speeds greater than 60 mph
and underestimated for less than 45 mph. Moreover,
in cases the INRIX speed was almost equal to the
actual speed). The authors are further investigating
the contributing factors, such as segment location,
segment length, time of day, etc., to this inconsistent
behavior. Furthermore, it should be noted that as
mentioned in the literature, the quality of INRIX data
has been improved over time which could be a con-
tributor to this different observation. However, this
topic is beyond the scope of this article and will be
presented in separate forthcoming work.
As observed in Figure 5, five ranges of speed were
chosen. According to Highway Capacity Manual ver-
sion 6 (Transportation Research Board, 2016), 45 mph
is considered as the breakdown speed. However, to
delve deeper into the characteristics of lower speeds
and their variations, the DOT has conventionally
studied speeds less and greater than 25 mph to calcu-
late congested hour. Thus, to align with these efforts
and make the findings comparable, we determined the
bins as presented.
Based on the exploratory analysis, the magnitude
and variation of speed bias for INRIX speeds below
45 mph was found to be greater than others. This is
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 9
Figure 5. (a) Boxplots of speed bias for five different ranges of INRIX speed. (b) Location of all segment-sensor pairs and their
corresponding speed biases for each group. Range of speed bias is from –20 (green color) to þ20 (red color). Yellow and orange
colors represent low magnitude of speed bias. It illustrates the fact that speed bias decreases for most of sensor-segment pairs by
increase of INRIX speed value. It should be noted that negative speed bias means that INRIX speed value is more than Wavetronix
sensor speed value.
10 V. AHSANI ET AL.
while speed bias for INRIX speeds greater than
45 mph, had less variation within an acceptable range
(mostly less than 3 mph). Therefore, a statistical model
was run on 6331 observations to further dissect the
influential factors in speed bias, within less than
45 mph INRIX speed. A linear regression model was
performed to ascertain the effects of speed value, time
of day, truck-AADT, number of lanes, type of TMC
segment, and segment length on the magnitude of
INRIX speed bias. The model was statistically signifi-
cant Fð28;6331Þ¼95:34;p<:0001 and explained
39.35% (Adjusted R
2
) of the variability in speed bias.
Of the twenty predictor variables, the statistically sig-
nificant ones were noted with details in the Table 4.
The model indicated an inverse relationship
between INRIX speed and speed bias (Less speed bias
in higher speeds), confirming the observation in
Figure 5(a). Moreover, certain timespans of the day,
had a significant impact on determining the speed
bias. In general, day hours (6 a.m.–7 p.m.) have lower
mean speed biases than the rest of the night. More
specifically, during morning and afternoon peak traffic
hours, speed bias is less than off-peak hours.
To examine the impact of traffic volume on speed
bias, AADT was considered. Since INRIX mostly
provides traffic data via trucks in the state of Iowa,
between AADT and truck-AADT, the truck-AADT
was observed to have contributed more significantly
to the model. Road segments with higher truck-
AADT, have less speed bias. This implies, higher
device penetration (number of probes) leads to more
accurate traffic information (speed) from INRIX. This
observation about the impact of volume on INRIX
speed bias, reinforces our interpretation of the less
speed bias in the crowded hours of the day.
Moreover, the above model shows that road segments
with higher number of lanes, lead to more capacity
(volume) on the road, which again leads to lower of
speed bias.
On the other hand, the model indicated that longer
segments have higher speed bias. As mentioned
before, INRIX calculates speed by averaging it over
the length (space mean speed). As the length of the
segment increases, the difference of space mean speed
(INRIX speed) and time mean speed (sensor
speed) increases.
Finally, considering the two types of TMC seg-
ments (internal and external), there was a significant
impact in speed bias. The implication is that internal
segments have higher speed bias than the external
ones. The reason is completely explained in the
next section.
3. Congestion detection
Improving traffic safety and operations have long
been areas of motivation among researchers and traf-
fic engineers. Traffic incidents, particularly traffic
crashes, are of great interest due to the huge delay
and costs that traffic injuries and fatalities impose on
society. According to the United States Department of
Transportation, traffic incidents are the main cause
for more than half of traffic congestions that occur
along US highways (Peniati, 2004). Generally, there
are two types of congestion, recurring and non-
recurring. Recurring congestion is regarded as the
congestion caused by the routine traffic in a normal
environment which is somehow expected, whereas
nonrecurring congestion is unexpected and is most
likely caused by an incident. Nonrecurring congestion
Table 4. Significant influencers in INRIX speed bias.
Fð28;6331Þ¼95:34;p<:0001, Adjusted R
2
¼39.35%, sample size ¼6360
Variable Estimate pValue Interpretation
Speed value –0.337 <.001 (INRIX speed ¼<45 mph) As
INRIX speed increases,
speed bias decreases
(Figure 5(a))
Time of the day 06:00–09:00 (morning
peak hour)
–16.31 <.001 In morning and afternoon
peak hours speed bias
decreases significantly
compared to off-peak
hours (09:00–16:00)
09:00–16:00 –8.47 .005
16:00–19:00 (afternoon
peak hour)
–14.28 <.001
Truck-AADT 0.002 <.001 Increased number of trucks
yields a decrease in
speed bias.
Number of lanes –1.119 <.001 Higher number of lanes yields
lower speed bias.
Segment length (mile) 0.393 <.001 Longer segments have higher
speed bias.
Type of segment 1.240 <.001 Internal TMC segments have
higher speed bias.
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 11
may emerge as a result of a variety of factors like lane
blocking crashes or disabled vehicles, work zone
lane closures, adverse weather conditions, etc. These
incidents may also have other consequences, such as
secondary crashes and delays in emergency medical
services, which can cause further complications and
impose additional costs. Consequently, monitoring the
transportation network and being able to detect and
report anomalies in real time are of great importance
in the realm of traffic management. This section
of the paper evaluated the influence of type of
congestion (recurring vs. non-recurring), type of TMC
segment, and segment length on the performance
of probe-sourced data in detecting congestion. For
this purpose, Wavetronix sensors are considered as
the benchmark.
Modified congestion detection algorithm
After data pre-processing, an adaptive incident
detection algorithm adopted by (Chakraborty, Hess,
Sharma, & Knickerbocker, 2017) was modified to
detect and classify congestion onset throughout the
network for the study period. The algorithm calculates
median and inter-quartile range for each time of day
(15 min period) and day of week from two month his-
tory. A dynamic threshold value is set for each 15 min
period for each weekday at median speed minus twice
the inter-quartile range. Recurring congestion inci-
dents were identified when speed dropped below
45 mph but it remained above the threshold calculated
for that location. Most of the recurring congestions
were also verified by CCTV cameras. Nonrecurring
congestions were identified based on three criteria: (a)
speed data of INRIX segment or the mean of 1-
minute aggregated speed data of Wavetronix sensor
for that location must drop below 45 mph, (b) it also
drops below dynamic threshold calculated based on
2 months of historical data for that specific location
for a significant period of time (15 min and more),
and (c) a matching incident must be reported by Iowa
Advanced Traffic Management System (ATMS). The
Iowa ATMS records all incidents, hazards, and con-
gestion detected by various sensors and cameras or
the reports by the highway helpers or police. The inci-
dents in this dataset are validated by the ATMS opera-
tors, thus serve as a reference for evaluating other
sources of data. However, not all incidents, particu-
larly congestion, are recorded in this dataset. The
detection algorithm for recurring and non-recurring
congestion is illustrated in Figure 6.
When studying the congestion detection perform-
ance and exploring recorded INRIX speeds, it was
observed that some segments have low speeds virtually
all of the times. The congestion events detected on
Figure 6. Visualization of the recurring and non-recurring congestion detection process with the modified dynamic thresh-
old algorithm.
12 V. AHSANI ET AL.
these segments, however, were mainly false alarms.
This negatively impacted the congestion detection
performance. Moreover, in the regression speed bias
model, TMC segment type turned out to have a sig-
nificant impact on the value of speed bias. Thus, the
characteristics of road segments (e.g., segment type,
segment length, etc.) were thoroughly investigated to
further understand their potential impact on conges-
tion detection applications.
Segment length
For different purposes, length of the road segments
for which probe-sourced data are available vary
significantly. INRIX uses either TMC segments or XD
segments as their basis for defining road sections on
which to report traffic data. XD segments with 1.5
miles as a maximum length are more constant than
TMC segments which can vary remarkably to even
more than 15 miles in the state of Iowa. There are
two types of TMC codes in INRIX; internal and exter-
nal. Traffic data, such as speed and travel time, are
provided for both internal and external TMC codes.
An external INRIX TMC code is the road segment
between interchanges, typically from the last merge
ramp of the upstream interchange to the first exit
ramp of the downstream interchange, while an
internal TMC code presents the road segment within
an interchange, typically from the first exit ramp to
the last entrance ramp (Young, Juster, & Kaushik,
2015). Hence, external TMC segments tend to be lon-
ger than internal TMC segments, which are usually
less than 0.4 miles. Figure 7(a) shows samples of
external and internal INRIX TMC segments in this
study. Figure 7(b) provides one day’s speed heat map
for I-35, I-80, and I-235 near Des-Moines area in
November 2016 as an example. Time of day is shown
on the vertical axis and several sample TMC segments
on the horizontal axis. Each cell represents the
reported INRIX speed for that specific time and seg-
ment of the road. The cells are color coded based on
the recorded speed values. Distinct recurring conges-
tion events are observed during the morning and
evening peak hours. There are few vertical light blue
lines around segments 10–15, which represent speeds
less than 45 mph for all minutes of the day. Those
lines correspond to internal TMC segments.
Basically, internal TMC segments, due to their
locations, commonly show low speeds throughout
the day. In these segments cars are either accelerat-
ing to enter or decelerating to exit the freeways,
thus the reported speed is mostly below 45 mph.
This explains the reason why speed bias on internal
segments is usually higher than external ones as
shown in Table 4.
Moreover, segment type affects precision of the
congestion detection algorithm. Thus, two scenarios
were considered for further analysis; I) using all
TMC segments in the study area, and II) removing
internal TMC segments and segments with less than
0.4-mile length, to compare the congestion detec-
tion performance.
Table 5 shows the reliability of INRIX in detecting
congestion events for the two predefined scenarios.
True positive (TP) represents a similar event which
is detected by both Wavetronix sensor and INRIX
segment; false negative (FN) means an event detected
in sensor data sets cannot be found in probe data set;
and false positive (FP) denotes an event which is
detected in probe data set and cannot be found in
sensor data set. Finally, the values in the last column
show the precision of congestion detection by INRIX,
calculated using Equation 3.
Figure 7. (a) INRIX TMC code segmentation, (b) space–time speed contour map.
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 13
Precision ¼TP
TP þFP (3)
The results are summarized in Table 5.As
observed, by removing internal TMC codes and
segments with length less than 0.4 miles and their
corresponding sensors (scenario II), FP which shows
false alarms associated with INRIX data were signifi-
cantly decreased for both non-recurring and recurring
congestion events. Simultaneously, the overall
precision of the congestion detection algorithm was
improved by excluding these segments. Thus, for con-
gestion detection applications using TMC segments, it
is recommended to exclude internal and less than
0.4 mile segments. Finally, it is evident from numbers
that precision of congestion detection by probe-source
streaming data (INRIX) is higher for recurring rather
than non-recurring.
Conclusion and recommendation
This research evaluated probe-sourced streaming data
from INRIX, to study its characteristics as a data
source for the ATMS. For this purpose, Wavetronix,
a commonly used infrastructure sensor data source,
was selected as benchmarked. Accuracy and reliability
of INRIX was evaluated by 3 different measures;
coverage, speed bias, and congestion detection.
In terms of coverage, INRIX covered almost all road
networks in Iowa, however, it mostly provides
real-time data on the interstates. It was also shown that
INRIX virtually always provides real-time data through-
out the day on the interstates. However, it is more reli-
able from 6 am to 6 pm on non-interstates. Moreover,
the real-time availability of INRIX speed was compared
for four consecutive years (2013–2016) and the results
indicated a significant improvement in 2016.
INRIX speed bias analysis, found meaningful
interpretations of influential factors in INRIX speed
bias. These findings further our understanding of this
probe-sourced data. In particular, INRIX speed value,
time of day, truck-AADT, number of lanes, type of
TMC segment, and segment length had significant
effects on the magnitude of speed bias. It should be
mentioned that use of XD data and higher market
penetration rates can potentially reduce the bias.
For the congestion detection analysis, three factors
of type of congestion, type of TMC segment,
and segment length were thoroughly examined. It was
concluded that probe segments with less than 0.4
miles length were observed to have the highest false
calls with regards to congestion. The majority of such
segments are located on interchanges where speeds
are typically lower. Also, it is determined that
precision of INRIX in detecting recurring congestion
is more than non-recurring one.
Finally, a major limitation of the analysis carried
out in this study was using sensor data as the bench-
marked dataset. We acknowledge that sensor data
would have its inherent errors and thus not really the
ground truth. Yet, there is an inevitable error within
the benchmarked sensor data that was unavoidable.
However, we believe that this error does not meaning-
fully impact our findings.
The following recommendations are offered by the
authors for transportation agencies and state DOTs
considering the augmentation of traditional traffic
data with probe-based services for wider coverage
under restricted budgets:
In terms of geographic coverage, INRIX was found
reliable for throughout the day on the interstates.
Moreover, this study showed that INRIX is more
reliable during the day than at night, especially
during peak hours. INRIX also has shown improve-
ment in real-time data coverage over the years.
Travel time estimation and incident detection
applications should be completely based on real-
time data. Substitutions with historical data are not
accurate and therefore not advised. In areas with
limited probe penetration, an agency could augment
probe data with infrastructure-mounted sensors.
The length of segments for which probe data are
available varies greatly, from 0.2 miles to more
than 15 miles. Agencies must examine whether
the space granularity of probe data is sufficient for
the intended application. For incident detection
applications, high space granularity may lead to
false alarms. Segments with shorter lengths (less
than 0.4 miles) are recommended to be excluded.
There will always be a bias between traffic speed
data from probe sources and benchmarked sensors.
Speed bias directly affects incident detection, travel
time estimation, calculating performance measures
Table 5. Reliability of probe data in detecting conges-
tion events.
Congestion type
Test
Congestion
detection (%) Precision (%)
Scenario I II I II
Non-recurring TP 35.0 32.0 51.2 61.3
FN 17.0 16.0
FP 33.3 20.2
Recurring TP 63.0 62.0 81.6 85.3
FN 5.0 5.0
FP 14.2 10.7
14 V. AHSANI ET AL.
(such as congested hour, BTI, etc.), and other
traffic-related measures. It is important to under-
stand the factors that influence these biases and
how to correct for them.
Future work
Knowing the potential value in probe-sourced data
encourages further investigation in the characteristics
of this data source, to build models that assist traffic
managers on the times and locations which they can
trust probe-sourced reports without further validation.
The authors plan to focus on developing models that
could be utilized for automatically correcting speed
measurements from probe data.
Acknowledgments
Our research results are based upon work supported by the
Iowa DOT Office of Traffic Operations Support Grant. Any
opinions, findings, and conclusions or recommendations
expressed in this material are those of the author(s) and do
not necessarily reflect the views of the Iowa DOT.
Funding
Our research results are based upon work supported by the
Iowa DOT Office of Traffic Operations Support Grant. Any
opinions, findings, and conclusions or recommendations
expressed in this material are those of the author(s) and do
not necessarily reflect the views of the Iowa DOT.
References
Adu-Gyamfi, Y. O., Sharma, A., Knickerbocker, S., Hawkins,
N., & Jackson, M. (2017). Framework for evaluating the
reliability of wide-area probe data. Transportation Research
Record: Journal of the Transportation Research Board,
2643,93–104. https://doi.org/10.3141/2643-11
Aliari, Y., & Haghani, A. (2012). Bluetooth sensor data and
ground truth testing of reported travel times. Transportation
Research Record: Journal of the Transportation Research
Board,2308 (1), 167–172. https://doi.org/10.3141/2308-18
Apache Hadoop (2017, November 28). Retrieved from
http://hadoop.apache.org/
Apache Pig (2017, November 28). Retrieved from https://
pig.apache.org/
Araghi, B. N., Hammershøj Olesen, J., Krishnan, R.,
Tørholm Christensen, L., & Lahrmann, H. (2015).
Reliability of bluetooth technology for travel time estima-
tion. Journal of Intelligent Transportation Systems,19 (3),
240–255. https://doi.org/10.1080/15472450.2013.856727
Barajas, V. L., Wang, Z., Kaiser, M., & Zhu, Z. (2017).
Improving estimates of real-time traffic speeds during
weather events for winter maintenance performance
measurement. Retrieved from https://trid.trb.org/view/
1465615
Chakraborty, P., Adu-Gyamfi, Y. O., Poddar, S., Ahsani, V.,
Sharma, A., & Sarkar, S. (2018). Traffic congestion
detection from camera images using deep convolution
neural networks. Transportation Research Record: Journal
of the Transportation Research Board, Retrieved from
https://doi.org/10.1177/0361198118777631
Chakraborty, P., Adu-Gyamfi, Y., Poddar, S., & Ahsani, V.
(2018). Traffic congestion detection from camera images
using deep convolution neural networks. Retrieved from
https://trid.trb.org/view/1496345
Chakraborty, P., Hess, J. R., Sharma, A., & Knickerbocker,
S. (2017). Outlier mining based traffic incident detection
using big data analytics. Retrieved from https://trid.trb.
org/view.aspx?id¼1439336
Coifman, B. (2002). Estimating travel times and vehicle
trajectories on freeways using dual loop detectors.
Transportation Research Part A: Policy and Practice,36 (4),
351–364. Retrieved from https://doi.org/10.1016/S0965-
8564(01)00007-6
Cookson, G., & Pishue, B. (2016). INRIX global traffic
scorecard. Inrix Global Traffic Scorecard, 44. Retrieved
from https://media.bizj.us/view/img/10360454/inrix2016-
trafficscorecarden.pdf
Day, C. M., Li, H., Richardson, L. M., Howard, J., Platte, T.,
Sturdevant, J. R., & Bullock, D. M. (2017). Detector-
free optimization of traffic signal offsets with connected
vehicle data. Transportation Research Record: Journal
of the Transportation Research Board,2620,54–68.
Retrieved from https://doi.org/10.3141/2620-06
Day, C., Remias, S., Li, H., Mekker, M., Mcnamara, M.,
Cox, E., …Wasson, J. (2015). 2013–2014 Indiana
Mobility Report. Retrieved from https://docs.lib.purdue.
edu/cgi/viewcontent.cgi?article ¼1006&context¼imr
Elhenawy, M., Chen, H., & Rakha, H. A. (2014). Dynamic
travel time prediction using data clustering and genetic
programming. Transportation Research Part C: Emerging
Technologies,42,82–98. Retrieved from https://doi.org/
10.1016/J.TRC.2014.02.016
Eshragh,S.,Young,S.E.,Sharifi,E.,Hamedi,M.,&
Sadabadi,K.F.(2017).Indirectvalidationofprobe
speed data on arterial corridors. Transportation
Research Record: Journal of the Transportation Research
Board,2643,105–111. Retrieved from https://doi.org/
10.3141/2643-12
FDOT (2012). Probe Data Analysis Evaluation of NAVTEQ,
TrafficCast, and INRIX V
RTravel Time System Data in the
Tallahassee Region Evaluation of NAVTEQ, TrafficCast,
and INRIX V
RTravel Time System Data. Retrieved from
http://www.fdot.gov/traffic/ITS/Projects_Deploy/2012-03-
26_Probe_Data_Analysis_v2-0.pdf
Feng, W., Bigazzi, A., Kothuri, S., & Bertini, R. (2010).
Freeway sensor spacing and probe vehicle penetration:
Impacts on travel time prediction and estimation
accuracy. Transportation Research Record: Journal of
the Transportation Research Board,2178(1), 67–78.
Retrieved from https://doi.org/10.3141/2178-08
FHWA (2013). Work zone performance measurement using
probe data. Retrieved from https://ops.fhwa.dot.gov/wz/
resources/publications/fhwahop13043/fhwahop13043.pdf
FHWA (2017). 2016 urban congestion trends. Retrieved
from https://ops.fhwa.dot.gov/publications/fhwahop17010/
fhwahop17010.pdf
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 15
Gong, L., & Fan, W. (2017). Applying travel-time reliability
measures in identifying and ranking recurrent freeway
bottlenecks at the network level. Retrieved from https://
doi.org/10.1061/JTEPBS.0000072
Haghani, A., Hamedi, M., & Sadabadi, K. F. (2009). I-95
corridor coalition vehicle probe project: validation of
INRIX data July–September 2008. Retrieved from http://
www.i95coalition.org/wp-content/uploads/2015/02/I-95-
CC-Final-Report-Jan-28-2009.pdf
Hard, E. N., Chigoy, B. T., Songchitruksa, P., Farnsworth,
S. P., Borchardt, D. W., & Green, L. L. (2017).
Comparison of cell, GPS, and bluetooth derived external
O-D data –results from the 2014 Tyler, Texas Study.
Transportation Research Record: Journal of the
Transportation Research Board. Retrieved from https://
trid.trb.org/view/1439253
Higatani, A., Kitazawa, T., Tanabe, J., Suga, Y., Sekhar, R., &
Asakura, Y. (2009). Empirical analysis of travel time reli-
ability measures in Hanshin expressway network. Journal
of Intelligent Transportation Systems,13 (1), 28–38.
Retrieved from https://doi.org/10.1080/15472450802644454
Hu, J., Ph, D., Fontaine, M. D., Ph, D., Park, B. B., Ph, D.,
…Ph, D. (2016). Field evaluations of an adaptive traffic
signal —using private-sector probe data. Journal of
Transportation Engineering,142 (1), 04015033–04015039.
Retrieved from https://doi.org/10.1061/(ASCE)TE.1943-
5436.0000806.
INRIX (2015). INRIX jI-95 corridor coalition: vehicle probe
project data validation summary. Retrieved from http://
inrix.com/wp-content/uploads/2016/11/INRIX-I-95-VPP-
Data-Summary-Validation-1.pdf
IowaDOT (2017). Advanced traveler management system/
advanced traveler information system combination. Retrieved
from https://iowadot.gov/purchasing/20227pro.pdf
Kim, K., Motiani, D., Spasovic, L. N., Dimitrijevic, B., &
Chien, S. (2014). Assessment of speed information based
on probe vehicle data: a case study in New Jersey.
Transportation Research Board, 93rd Annual Meeting,
9(January). Retrieved from http://trid.trb.org/view/2014/
C/1289396%5Cn http://assets.conferencespot.org/file-
server/file/166151/filename/14-4464.pdf
Kim, S., & Coifman, B. (2014). Comparing INRIX speed
data against concurrent loop detector stations over sev-
eral months. Transportation Research Part C: Emerging
Technologies,49,59–72. Retrieved from https://doi.org/
10.1016/j.trc.2014.10.002
Knoop, V., Hoogendoorn, S. P., & Van Zuylen, H. (2009).
Empirical differences between time mean speed and space
mean speed. In Traffic and Granular Flow 2007, 351–356.
Retrieved from https://doi.org/10.1007/978-3-540-77074-
9-36
Lattimer, C., & Glotzbach, G. (2012). Evaluation of third-
party travel time data in Tallahassee, FL. Retrieved from
https://itswc.confex.com/itswc/AM2012/webprogram/
Paper10870.html
Lindveld, C. D. R., Thijs, R., Bovy, P. H. L., & der Zijpp,
N. J. (2000). Evaluation of online travel time estimators
and predictors. Transportation Research Record,1719,
45–53.
Lomax, T., Schrank, D., Turner, S., & Margiotta, R. (2003).
Selecting travel reliability measures. Retrieved from
https://static.tti.tamu.edu/tti.tamu.edu/documents/TTI-
2003-3.pdf
Lu, C., & Dong, J. (2018). Estimating freeway travel time
and its reliability using radar sensor data. Transportmetrica
B: Transport Dynamics,6(2), 97–114. Retrieved from
https://doi.org/10.1080/21680566.2017.1325785
Lu, X., Lee, J., Chen, D., Bared, J., Dailey, D., & Shladover,
S. E. (2014). Freeway micro-simulation calibration: case
study using Aimsun and VISSIM with detailed field data.
Transportation Research Board 93rd Annual Meeting.
January 12–16, Washington, D.C,1–17.
Miwa, T., Ishiguro, Y., Yamamoto, T., & Morikawa, T.
(2015). Allocation planning for probe taxi devices aimed
at minimizing losses to travel time information users.
Journal of Intelligent Transportation Systems,19 (4),
399–410. https://doi.org/10.1080/15472450.2014.995760
MoDOT (2017). Tracker: measures of departmental perform-
ance. Retrieved from http://www.modot.org/about/docu-
ments/Tracker_July17/July2017FinalTracker.pdf
Peniati, J. (2004). Operational solutions to traffic congestion.
FHWA Journal, 68 (3). Retrieved from https://www.fhwa.
dot.gov/publications/publicroads/04nov/01.cfm.
Poddar, S., Ozcan, K., Chakraborty, P., & Ahsani, V. (2018).
Comparison of machine learning algorithms to determine
traffic congestion from camera images. Retrieved from
https://trid.trb.org/view/1496348
Pu, W. (2011). Analytic relationships between travel time
reliability measures. Transportation Research Record:
Journal of the Transportation Research Board,2254(1),
122–130. Retrieved from https://doi.org/10.3141/2254-13
Rakha, H., El-Shawarby, I., & Arafeh, M. (2010). Trip
travel-time reliability: Issues and proposed solutions.
Journal of Intelligent Transportation Systems,14 (4),
232–250. Retrieved from https://doi.org/10.1080/15472450.
2010.517477
Remias, S., Brennan, T., Day, C., Summers, H., Cox, E.,
Horton, D., & Bullock, D. (2013). 2012 Indiana Mobility
Report. Retrieved from https://docs.lib.purdue.edu/cgi/
viewcontent.cgi?article ¼1004&context¼imr
Sanaullah, I., Quddus, M., & Enoch, M. (2016). Developing
travel time estimation methods using sparse GPS data.
Journal of Intelligent Transportation Systems,20 (6),
532–544. https://doi.org/10.1080/15472450.2016.1154764
Schrank, D., Eisele, B., Lomax, T., & Bak, J. (2015). 2015
Urban Mobility Scorecard.Texas A&M Transportation
Institute (Vol. 39). Retrieved from https://doi.org/
DTRT06-G-0044
Schrank, D., Eisele, B., & Lomax, T. (2012). TTI’s 2012
urban mobility report. Texas A&M Transportation
Institute. Retrieved from http://d2dtl5nnlpfr0r.cloudfront.
net/tti.tamu.edu/documents/mobility-report-2012.pdf
Sekuła, P., Markovi
c, N., Laan, Z. V., & Sadabadi, K. F.
(2017). Estimating historical hourly traffic volumes via
machine learning and vehicle probe data: A Maryland
case study. Retrieved from http://arxiv.org/abs/1711.00721
Sharifi, E., Hamedi, M., Haghani, A., & Sadrsadat, H.
(2011). Analysis of vehicle detection rate for bluetooth
traffic sensors: A case study in Maryland and Delaware.
Sharifi, E., Young, S. E., Eshragh, S., Hamedi, M., Juster,
R. M., & Kaushik, K. (2017). Outsourced probe data
effectiveness on signalized arterials. Journal of Intelligent
16 V. AHSANI ET AL.
Transportation Systems,21(6), 478–491. Retrieved from
https://doi.org/10.1080/15472450.2017.1359093
Sharma, A., Ahsani, V., & Rawat, S. (2017). Evaluation of
opportunities and challenges of using INRIX data for
real-time performance monitoring and historical trend
assessment. Reports and White Papers. 24. Retrieved from
https://lib.dr.iastate.edu/ccee_reports/24
Transportation Research Board (2016). Highway capacity
manual, sixth edition: A guide for multimodal mobility
analysis. Transportation Research Board. Retrieved from
http://app.knovel.com/hotlink/toc/id:kpHCMAGMM2/
highway-capacity-manual/highway-capacity-manual
Turner, S. (2013). Developing twin cities arterial mobility
performance measures using GPS speed data. Retrieved
from https://www.lrrb.org/pdf/201314.pdf
Uno, N., Kurauchi, F., Tamura, H., & Iida, Y. (2009). Using
bus probe data for analysis of travel time variability.
Journal of Intelligent Transportation Systems,13 (1), 2–15.
Retrieved from https://doi.org/Pii 908329318\rDoi 10.
1080/15472450802644439
Venkatanarayana, R. (2017). Considerations for calculating
arterial system performance measures in Virginia. Retrieved
from http://www.virginiadot.org/vtrc/main/online_reports/
pdf/17-r2.pdf
WSDOT (2013). The 2013 corridor capacity summary.
Retrieved from http://wsdot.wa.gov/publications/fulltext/
graynotebook/CCS13.pdf
WSDOT (2014). Gray notebook 52-for the quarter ending
December 31, 2013. Retrieved from http://wsdot.wa.gov/
publications/fulltext/graynotebook/Dec13.pdf
Young, S. (2007). Real-time traffic operations data using vehicle
probe technology. 2007 Mid-Continent Transportation
Research Symposium. Retrieved from http://www.ctre.iastate.
edu/pubs/midcon2007/YoungVehicleProbe.pdf
Young, S. E., Juster, R. M., & Kaushik, K. (2015). Traffic
message channel codes: impact and use within the I-95
corridor coalition’s vehicle probe project. Retrieved from
http://i95coalition.org/wp-content/uploads/2015/02/TMC_
White_Paper-Final.pdf?x70560
Zheng, F., Li, J., van Zuylen, H., Liu, X., & Yang, H. (2018).
Urban travel time reliability at different traffic conditions.
Journal of Intelligent Transportation Systems,22 (2),
106–120. Retrieved from https://doi.org/10.1080/15472450.
2017.1412829
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 17
