ABSTRACT
The foundation of traffic operations and management is
the ability to monitor traffic conditions. One
approach to traffic monitoring is to sample conditions by “tracking” a limited
number of probe vehicles as they traverse a network. An emerging technology known as wireless location
technology (WLT) has been developed to allow for the geolocation of mobile wireless
devices (the most common of which are cellular telephones).
Over the past decade, a number of research studies and operational tests
have attempted to develop probe traffic monitoring systems based on WLT.
However, there still exists significant confusion and misperceptions concerning
WLT-based traffic monitoring. To address
this problem, this paper seeks to provide a comprehensive assessment of WLT-based
traffic monitoring. To do so, the specific
purposes of this paper are to: (a) fully describe the concept of WLT-based
traffic monitoring, (b) present a critical assessment of past studies of WLT-based
traffic monitoring, (c) document the evaluation of one of the most recent operational
tests – the 2001 Virginia Department of Transportation (VDOT), Maryland State
Highway Administration (MSHA), and US Wireless Corporation (USWC) effort in the
Washington, D.C. region, and (d) discuss the unique challenges that these systems
pose to the field of traffic engineering.
INTRODUCTION
Since
the intelligent transportation systems (ITS) program began to take shape in the
early 1990s, transportation agencies at the federal, state, and local levels have
focused significant resources on using information technology to improve surface
transportation. While ITS has evolved
and now takes a myriad of forms, the foundation of nearly every ITS initiative
is the ability to measure traffic conditions on the network – generally referred
to as traffic condition monitoring. Without knowledge of traffic conditions, transportation
professionals can do little to manage traffic or provide traveler information.
Furthermore, traffic condition data collected by ITS are now being used
in a wide range of transportation applications, such as planning and infrastructure
management.
Transportation
professionals generally measure traffic conditions using a network of “point”
sensors installed at strategic locations throughout the network. While this approach is functional, it suffers
from practical limitations that have resulted in incomplete and erratic traffic
condition data. The communications, installation,
and maintenance costs of the sensor networks have forced transportation agencies
to install them only on the most critical routes – often with widely spaced sensors.
This leaves many “holes” in the system that lack traffic condition data.
An
alternative approach to traffic condition monitoring is to sample a portion of
vehicles as they traverse the network. This
approach is generally referred to as probe-based monitoring. An emerging information technology, wireless
location technology (WLT), holds great promise to provide the platform for a probe-based
traffic condition monitoring system. This technology allows wireless devices to be geolocated while in
use. WLT supports “location-based services”
(LBSs) in which wireless subscribers (whether using cellular phones, automobile-based
telematic devices, mobile computing products, etc.) are provided with targeted
information content based on their specific location (i.e. latitude/longitude).
Thus, WLT offers an existing and growing infrastructure that can conceptually
be tapped to provide probe-based traffic information, without the need for transportation
agencies to install extensive infrastructures solely devoted to traffic monitoring.
Given
the appeal of this concept, it is no surprise that WLT-based monitoring has captured
the interest of the ITS community over the past decade. In fact, there have been a number of relatively
large-scale operational tests undertaken in an effort to accelerate development
of WLT-based traffic monitoring systems. While these tests have generally been categorized as unsuccessful,
the project participants have gained significant insight into the abilities of
WLT-based traffic monitoring. Unfortunately,
relatively little of this insight has made its way into formal research papers
or reports. This has led to significant
confusion in the transportation community, and a lack of information on which
to base future decisions concerning this technology.
To
address this need, this paper seeks to provide a comprehensive introduction to
the concepts, experiences with, and performance of early-generation WLT-based
traffic monitoring systems. The specific
purposes of this paper are to (a) describe the concept of WLT-based traffic monitoring,
(b) present a critical assessment of past studies of WLT-based traffic monitoring,
(c) document the evaluation of one of the most recent operational tests – the
Virginia Department of Transportation (VDOT), Maryland State Highway Administration
(MSHA), and US Wireless Corporation (USWC) effort in the Washington, D.C. region,
and (d) discuss the challenges that these systems pose to the field of traffic
engineering.
WLT-BASED
TRAFFIC MONITORING
WLT
can be classified into two general groups: handset-based systems and network-based
systems. Handset-based systems rely on
global positioning system (GPS)-enabled wireless phones. The GPS unit in the handset determines the
location of a phone, and this information is relayed from the phone to a central
processing system maintained by the wireless carrier. Network-based systems utilize signal information
from cell phones to derive their location. In some cases, network-based systems require
special equipment to be installed throughout a metropolitan area in order to analyze
signal characteristics of calls. For example,
some network-based systems determine positions by analyzing signal power and angle
of arrival at multiple cellular towers. In other cases, network-based systems derive
location estimates purely from signaling information already available at cellular
towers. Since network-based systems do
not require users to have GPS-enabled phones, they generally provide less spatial
accuracy than handset-based systems.
There
have been several attempts to use WLT data to generate traffic condition information.
In order to produce this type of information, a WLT system has to be able
to perform three basic tasks:
- Location
Determination. The location of the probe must be determined.
These position estimates are usually inaccurate to some degree, and may
not lie directly on the roadway network.
- Map Matching. In map matching, the position estimates
are matched to a specific road. Techniques
for map matching vary from simple geometric methods to more complex statistical
approaches that account for link connectivity and past travel history.
- Determination
of Traffic Information. The probes represent a small sample of the entire
traffic stream. A WLT-based system must
be able to use a set of these samples to estimate the speed or travel time for
all traffic on a link.
Several
evaluations of early generation WLT-based traffic condition monitoring systems
have occurred, but none of these systems adequately performed all three of these
tasks. Past evaluations have taken one
of two forms: simulation studies or field operational tests. Major findings of these studies are summarized
below.
Simulation Studies
Researchers have used simulation studies to explore the potential accuracy
and effectiveness of WLT-based systems. While these studies do not replicate the actual conditions precisely,
they do provide some indication of the potential performance of a WLT-based system.
French
Simulation Study
A
study conducted by the French transportation research organization, INRETS, focused
on developing a discrete event simulation of traffic flow in order to determine
the sample size requirements and accuracy of a hypothetical WLT system (Ygnace,
et. al., 2000). The simulation examined
the impact of varying levels of probe vehicle penetration on the accuracy of travel
time estimates. A location error of 150
meters was assumed, and researchers examined a series of traffic and geometric
conditions. The simulation results showed
that freeway link travel times could be estimated to within 10 percent of their
actual value if there is at least 5 percent penetration of wireless devices in
the traffic stream. These promising results are based on relatively
simple geometric conditions.
Berkeley
Simulation Study
A
recent evaluation by the Berkeley Institute for Transportation Studies examined
factors that could affect the utility of WLT-based traffic monitoring systems
(Cayford and Johnson, 2003). The researchers
examined three variables in their simulation: location accuracy, frequency of locations of
a single wireless device, and the total number of locations that could be determined
per square mile per second. The variation
in the number of roads that could be traversed by at least one vehicle within
a five-minute period was used as the measure of effectiveness to compare different
alternatives. The researchers did not
attempt to address whether the observed sample sizes were sufficient to produce
accurate estimates of speeds or travel times for the entire traffic stream, however.
The major findings of this research effort were:
- Assuming a network-based system accurate to within
100 meters, at least one probe speed sample can be generated on 85 percent of
the roads every 5 minutes. This assumes
that positions are updated every 30 seconds, and a maximum of 40 locations are
determined every second per square mile.
- Assuming a handset-based system accurate to within
50 meters, a probe speed sample can be generated on 90 percent of the roads every
5 minutes. This assumes that positions
are updated every 30 seconds, and a maximum of 40 locations are determined every
second per square mile.
Again,
these results only show whether a tracked probe vehicle traveled on a particular
road at least once during a 5-minute period. The researchers did not state whether this
would be sufficient to determine actual speeds or travel times.
Field
Operational Tests
Several
field tests of WLT-based systems have been performed in the United States. Network-based systems have been used in all
cases since GPS-enabled phones currently account for a relatively small portion
of available probes.
CAPITAL
Field Operational Test
The
first major operational test of a WLT-based system was conducted over a 27-month
period in the mid-1990’s on several interstates and state routes in Virginia (UMD,
1997). This project was named CAPITAL
(Cellular APplied to ITS Tracking And Location), and was the result of a cooperative
agreement between FHWA, VDOT, MSHA, and several private sector firms.
This evaluation produced the following major findings:
- By the end of testing, wireless telephones could
be located within 100 meters of their actual position. The accuracy of the position estimates
improved considerably as the number of cellular towers providing directional information
increased. The evaluators noted that accuracies
on the order of 5 to 25 meters might be needed to perform accurate speed estimation
for a network.
- In order to calculate speed, at least four position
estimates had to be identified for each phone, and this occurred only 20 percent
of the time. As a result, link speed estimates
could not be generated for the network. This was due to the small number of data
points, as well as a lack of well-developed algorithms to match vehicles to links.
While
the CAPITAL test showed that WLT could provide reasonably accurate positional
data, it was unsuccessful in producing traffic information.
U.S.
Wireless Operational Tests
In
addition to the VDOT/MSHA/USWC evaluation documented later in this paper, the
U.S. Wireless Corporation (no longer in business) also participated in an operational
test in Oakland, California (Yim and Cayford, 2001). Researchers at the University of California
- Berkeley obtained 44 hours of wireless location data. The researchers found that the position estimates
generally had a 60-meter accuracy, although 66 percent of all probe vehicle tracks
had at least one data point that deviated from the caller’s actual position by
more than 200 meters.
While
the researchers were generally able to identify the location of a vehicle, they
were not successful in matching vehicles to roads or in generating speed or travel
time information. The researchers noted
that the call lengths were generally very short, with a median call length of
only 30 seconds. This made it impossible
to estimate speeds on links since position estimates were not available for long
distances. As a result, the researchers were not able to match 60 percent of
vehicles to a roadway link.
The
most recently completed large-scale field test of WLT-based traffic monitoring
was the VDOT/MSHA/USWC effort that took place in the Washington D.C. region beginning
in the year 2000. The next sections of
this paper introduce this test, describe the evaluation methodology, and present
a summary of the evaluation results.
DEPLOYMENT DESCRIPTION – VDOT/MSHA/USWC OPERATIONAL TEST
Beginning in 2000, VDOT and MSHA participated in a WLT
demonstration project with USWC in the southern suburban region of Washington
D.C.. This region includes the Capital
Beltway, a heavily traveled 8-lane freeway that experiences significant congestion,
and many major arterials. The purpose
of the demonstration project was to demonstrate the feasibility of
WLT-based traffic condition monitoring.
EVALUATION METHODOLOGY
The University of Virginia’s Center for Transportation Studies served as
the traffic monitoring evaluator in this field test (the University of Maryland
also served as an evaluator, focusing on location estimation of individual vehicles).
This section details the methodology developed for the WLT-based traffic
monitoring evaluation.
Data Collection
A key evaluation challenge was
to collect accurate baseline traffic data to serve as the ground truth against
which WLT-based system estimates of macroscopic traffic parameters could be compared.
It was essential that the baseline data accurately represent the actual
conditions on the facilities – and that this could be verified through manual
means. Two possible approaches to collecting the baseline
data were available to the evaluator: probe
vehicles and point sensors.
Conceptually, the use of probe vehicles for baseline
data collection is desirable. Given that
the WLT-based system collects data on “probes” over the entire length of roadway
links, baseline probes would allow for the same measurement.
However, without tolled facilities in the project region, it was impractical
to collect significant samples of baseline probe data.
The use of a point sensor, on the other hand, allows the speed of every
vehicle passing a point on a link to be collected.
However, it requires the key assumption that conditions are uniform throughout
the link in order to compare with the WLT-based system results. Traffic volumes can also be determined using
point sensors, something that is impossible with a probe-based approach. This is important in that it allows for the
assessment of the temporal adequacy of samples from the WLT-based system by providing
true population data for the macroscopic traffic parameters. Given these advantages, as well as data collection
constraints, it was determined that point sensors would be used as the baseline
measurement approach in this evaluation.
A van-mounted video detection system was used for
baseline data collection. The video detection
system was used to derive point measures (spot speeds and counts) by processing
video from a camera mounted on a 45-foot telescoping mast installed on the van.
This system provided the flexibility to position the location of collection
in the mid-point of WLT-based system defined links (relatively short, 0.4 miles
in length), and allowed for manual verification of baseline data accuracy (using
a hand-held Lidar unit for speed, and manual counts for volume).
Given that the focus of this evaluation was on the
ability of WLT-based systems to support traffic monitoring applications, it was
decided that a relatively short “polling” interval should be used in the effort.
Based on information in the Highway Capacity Manual (TRB, 2000) and
recent research on flow rate stability (Smith and Ulmer, 2003) 10 minutes was
selected as the polling interval used.
Finally,
it was important to clearly define the term “sample” in the data collection effort.
In many cases, a WLT-based system may sample the speed of the same vehicle
multiple times as it traverses a single link.
While it is possible to argue the validity of treating each sample of the
same vehicle as an independent sample of link speed (i.e. assuming a vehicle’s
speed is solely governed by conditions over the entire length of the link),
a conservative approach that most directly corresponds to traditional traffic
monitoring practice is to consider the average speed of the multiple samples from
the same vehicle as the single sample for that vehicle/link pair. This was the approach used in this research.
Comparative
Analysis
The comparative analysis included two key components.
First, the baseline link population data was used to compute confidence
intervals on mean speed estimates to identify minimum required sample sizes for
deriving traffic data of particular levels of “quality.”
These were then compared with the WLT-based system samples to ascertain
if the system is theoretically capable of providing sufficient numbers of samples.
Second, link data from the baseline video detection system and WLT-based system
were directly compared to determine if the WLT-based system produced link mean
speed results that accurately reflect the ground truth.
Sample Size Adequacy Evaluation
Given that a WLT-based traffic
monitoring system will produce individual vehicle speed samples on links, or samples
of the random variable, U, and the system
is attempting to estimate the mean link speed, μ, the Central Limit Theorem
can be used to estimate the number of samples required to estimate μ to within
some level of allowable error at an assumed confidence level.
This approach is widely used in experimental design and has been used,
for example, to determine the number of probes required for speed estimation in
the Houston automatic vehicle identification (AVI) based traffic monitoring system
(Turner and Holdener, 1996).
Based on the Central Limit Theorem,
one can collect n probe samples and
compute a confidence interval about the population mean as follows:
(1)
In other words, there is
an a
probability that the true mean link speed falls in the interval defined above.
Note that this effort’s methodology allows for the direct computation of
the speed population parameters since the van-based video detection system measured
the speed of every vehicle traversing the link over the 10-minute interval.
Working with equation (1), if
the “width” of the confidence interval is defined as 2d miles/hour (i.e., an error of ±d
miles/hour can be accepted in the estimation of the true link mean speed, m) an equation can be derived
to determine the minimum required sample size, n. Note that a confidence
level, α, must be assumed.
(2)
Finally, it must be noted
that this methodology assumes that a WLT-based traffic monitoring system measures
the speed of each vehicle traversing a link without error. Certainly, this will not be the case. As such, the values of n presented in the results represent the absolute minimum number of
samples required. A fielded system will
likely require a somewhat larger number of samples to account for errors in the
system’s ability to measure individual vehicle speeds.
Link Conditions Estimation Evaluation
For each 10-minute interval, mean speeds produced
by the WLT-based system were compared with the baseline data to measure percentage
error in the speed estimates of the WLT-based system. In addition, hypothesis tests were conducted
to rigorously assess the difference in mean speed values. The Wilcoxon Rank-sum Test, the most widely
used nonparametric alternative to the independent samples t-test, was selected
for this purpose. The mean speed of baseline and WLT-based system samples were
compared at the 95 percent confidence level using the following hypothesis test:
Ho:
m = mo
Ha: m ą mo
Where,
m = WLT-based
system mean 10-minute interval link speed
mo=
Baseline mean 10-minute interval link speed
RESULTS
Three major data collection efforts were conducted to
gather the baseline data needed for the evaluation. These took place in the Fall of 2001, and included links that could
be classified as freeway, high-speed major arterial, and low-volume/speed urban
links. Data collection was conducted primarily
during daylight hours, however, data were collected in the evening for several
hours on a freeway link. It should first
be noted that the WLT-based system was unable to reliably collect sufficient samples
to estimate conditions on low-volume/speed urban links – therefore no results
are available for this category of facility. Table 1 provides detail on the links analyzed
in this research.
TABLE
1,
Link
Descriptions
| Link ID | Road Name | Direction | Speed Limit (mph) | AADT | # Lanes | Description |
| 201 | I-495 | East | 55 | 65,000 |
4 | Between
Telegraph Rd & US1 |
| 202 | I-495 | West | 55 | 73,000 |
4 | Between
Telegraph Rd & US1 |
| 103 | US-1 | North | 45 | 58,000 (Bidirectional) |
3 | Immediately
south of I-495 interchange |
| 104 | US-1 | South | 45 | 58,000 (Bidirectional) |
3 | Immediately
south of I-495 interchange |
| 105 | US-1 | North | 30 | 64,000 (Bidirectional) |
3 | Immediately
north of I-495 interchange |
| 121 | George Washington Parkway | North | 40 | 27,000 (Bidirectional) |
1 | Immediately
south of I-495 interchange |
| 122 | George Washington Parkway | South | 40 | 27,000 (Bidirectional) |
1 | Immediately
south of I-495 interchange |
| 242 | Duke Street | West | 40 | 23,000 (Bidirectional) |
3 | Immediately
west of US-1 |
WLT-Based System Sample
Size Adequacy Evaluation
The
WLT-based system was only capable of sampling a relatively small portion of total
vehicles traversing the links analyzed. Table
2 presents mean results for links classified as freeway (daylight), freeway (evening),
and arterial. It should also be noted
that on many occasions, the WLT-based system only sampled 1 or 2 vehicles per
link per 10-minute interval. During four
10-minute intervals, the WLT-based system did not sample a single vehicle on a
link (this occurred twice on link 201, I-495, and twice on link 104, US-1).
TABLE 2.
WLT-Based System Mean Sample Sizes
| Link Classification |
Mean WLT Samples (per 10 minutes) |
Mean Traffic Count (per 10 minutes) |
Percentage of Traffic
Stream Sampled |
| I-495 (daytime) | 7.0 | 636 | 1.1% |
|
I-495 (evening) | 3.7 | 462 | 0.8% |
|
Arterials | 4.1 | 216 | 1.9% |
While the above results seem to paint a rather bleak
picture, it is important to consider needed sample sizes based on the Central
Limit Theorem. The Central Limit Theorem
shows that relatively small sample sizes are required when speed variance is low,
so the small sample sizes observed in the field test may be acceptable.
Table 3 provides the minimum sample sizes required for freeway links given
two confidence intervals. The first confidence
interval was intended to capture the strict requirements of traffic management
and control – with the intervals defined as ±5 mph with a confidence level of
99%. The other interval was designed to
represent the less strict requirements of traveler information systems – with
the intervals defined as ±10 mph with a confidence level of 95%. In addition to these minimum required sample
sizes, the actual numbers of samples collected by the WLT-based system are reported.
TABLE 3.
Freeway Link
Results – Sample Size
| | Time Interval | Total Vehicles | Population Mean Speed m | Population Standard Deviation,
s | Sample Required (99% C.I., Error=±5mph) | Sample Required (95% C.I., Error=±10mph) | USWC Sample Size |
|
Sept I-495 East Link 201 | 9:50-10:00 | 430 | 68.7 | 7.9 | 16.6 | 2.4 | 3 |
| 10:00-10:10 | 466 | 51.6 | 21.6 | 124.2 | 18.0 | 5 |
| 10:10-10:20 | 592 | 22.0 | 6.2 | 10.3 | 1.5 | 4 |
| 10:20-10:30 | 745 | 20.0 | 6.4 | 10.9 | 1.6 | 13 |
| 10:30-10:40 | 824 | 23.7 | 6.5 | 11.2 | 1.6 | 6 |
| 10:40-10:50 | 729 | 51.8 | 21.6 | 123.4 | 17.9 | 3 |
| 10:50-11:00 | 694 | 67.5 | 7.7 | 15.9 | 2.3 | 12 |
| 11:00-11:10 | 697 | 68.2 | 8.8 | 20.4 | 3.0 | 3 |
| 11:10-11:20 | 405 | 68.0 | 8.4 | 18.9 | 2.7 | 7 |
|
Oct I-495 East Link 201 | 11:35-11:45 | 597 | 65.7 | 7.4 | 14.4 | 2.1 | 4 |
| 11:45-11:55 | 613 | 66.2 | 7.0 | 13.0 | 1.9 | 6 |
| 11:55-12:05 | 567 | 66.0 | 7.7 | 15.6 | 2.3 | 4 |
| 12:05-12:15 | 361 | 46.0 | 19.1 | 96.5 | 14.0 | 1 |
| 12:15-12:25 | 574 | 19.1 | 6.0 | 9.6 | 1.4 | 5 |
| 12:25-12:35 | 741 | 22.4 | 7.7 | 15.7 | 2.3 | 8 |
|
Oct I-495 West Link 202 | 13:30-13:40 | 733 | 66.4 | 6.4 | 10.8 | 1.6 | 5 |
| 13:40-13:50 | 732 | 66.1 | 6.0 | 9.6 | 1.4 | 6 |
| 13:50-14:00 | 828 | 64.9 | 6.4 | 11.0 | 1.6 | 2 |
| 14:00-14:10 | 750 | 65.9 | 6.3 | 10.7 | 1.5 | 7 |
| 14:10-14:20 | 705 | 65.3 | 6.6 | 11.4 | 1.7 | 6 |
| Nov I-495 East Link 201 | 17:40-17:50 | 850 | 17.0 | 5.2 | 7.1 | 1.0 | 22 |
| 17:50-18:00 | 736 | 14.2 | 7.4 | 14.5 | 2.1 | 11 |
| 18:00-18:10 | 756 | 15.4 | 6.8 | 12.4 | 1.8 | 6 |
| 18:10-18:20 | 873 | 19.2 | 6.5 | 11.2 | 1.6 | 16 |
| 18:20-18:30 | 667 | 40.9 | 19.1 | 100.4 | 14.5 | 9 |
| 18:30-18:40 | 523 | 64.8 | 7.9 | 16.4 | 2.4 | 4 |
| 18:40-18:50 | 530 | 65.7 | 7.6 | 15.2 | 2.2 | 5 |
| 18:50-19:00 | 498 | 65.3 | 7.5 | 14.9 | 2.2 | 5 |
| 19:00-19:10 | 510 | 66.5 | 7.0 | 13.1 | 1.9 | 0 |
| 19:10-19:20 | 318 | 64.9 | 7.7 | 15.9 | 2.3 | 3 |
| 19:20-19:30 | 439 | 67.2 | 7.7 | 15.8 | 2.3 | 3 |
| 19:30-19:40 | 470 | 65.8 | 7.3 | 14.2 | 2.1 | 0 |
| 19:40-19:50 | 483 | 64.6 | 7.7 | 15.6 | 2.3 | 2 |
| 19:50-20:00 | 443 | 67.3 | 7.2 | 13.6 | 2.0 | 1 |
| 20:00-20:10 | 352 | 66.8 | 6.5 | 11.1 | 1.6 | 3 |
| 20:10-20:20 | 392 | 67.4 | 7.2 | 13.8 | 2.0 | 4 |
| 20:20-20:30 | 394 | 66.2 | 6.9 | 12.8 | 1.9 | 8 |
| 20:30-20:40 | 363 | 65.9 | 7.0 | 13.1 | 1.9 | 4 |
| 20:40-20:50 | 328 | 67.2 | 7.0 | 13.1 | 1.9 | 2 |
The results presented in Table 3 indicate that the feasibility of an early-generation
WLT-based system providing adequate quantities of samples is dependent on accuracy
requirements. For the more stringent traffic
management and control requirements, the WLT-based system only provided sufficient
samples in 3 of the 39 intervals. On the other hand, when the requirements are
relaxed to traveler information standards, the WLT-based system met the sample
size requirements in 30 of the 39 intervals.
As evident in Table 3, speed variance plays a major role in the minimum
sample size requirements. As seen in equation
2, as the standard deviation of U increases,
the sample size requirements increase dramatically. For example, when the speeds dropped significantly
during the 12:05-12:15 interval on link 201 in October, a minimum of 97 samples
are required to produce an average speed estimate within ±5 mile/hour error at
a 99% confidence interval. Thus, it is
clear that when any probe-based system is deployed in an area that experiences
frequent changes in conditions (such as those that experience frequent incidents),
the sample size requirements will increase significantly.
The sample size adequacy analysis was also conducted
for arterial links. A total of 35 10-minute
intervals were analyzed on the major arterials (US 1, Duke Street, George Washington
Parkway). Table 4 presents a summary of
the results.
TABLE 4.
Arterial Link Results – Sample
Size
| | Time Interval | Total Vehicles | Population Mean Speed m | Population Standard Deviation,
s | Sample Required (99% C.I., Error=±5mph) | Sample Required (95% C.I., Error=±10mph) | USWC Sample Size |
|
Oct US-1 North Link 103 | 15:55-16:05 | 240 | 36.0 | 6.0 | 9.4 | 1.4 | 7 |
| 16:05-16:15 | 244 | 35.3 | 5.7 | 8.5 | 1.2 | 3 |
| 16:15-16:25 | 268 | 29.6 | 11.2 | 33.1 | 4.8 | 1 |
| 16:25-16:35 | 261 | 29.6 | 8.5 | 19.2 | 2.8 | 4 |
| 16:35-16:45 | 248 | 32.3 | 7.5 | 15.1 | 2.2 | 4 |
| 16:45-16:55 | 261 | 22.2 | 10.2 | 27.4 | 4.0 | 3 |
|
Oct US-1 South Link 104 | 17:35-17:45 | 116 | 37.3 | 8.6 | 19.6 | 2.8 | 2 |
| 17:45-17:55 | 292 | 33.4 | 8.2 | 17.9 | 2.6 | 3 |
| 17:55-18:05 | 290 | 33.0 | 8.6 | 19.8 | 2.9 | 0 |
| 18:05-18:15 | 298 | 32.8 | 8.9 | 20.9 | 3.0 | 3 |
| 18:15-18:25 | 375 | 26.6 | 9.5 | 24.0 | 3.5 | 4 |
| 18:25-18:35 | 338 | 28.4 | 7.7 | 15.6 | 2.3 | 4 |
|
Oct GW North Link 121 | 8:25-8:35 | 251 | 43.8 | 8.7 | 20.2 | 2.9 | 2 |
| 8:35-8:45 | 329 | 44.5 | 5.4 | 7.8 | 1.1 | 1 |
| 8:45-8:55 | 355 | 44.4 | 5.6 | 8.3 | 1.2 | 2 |
| 8:55-9:05 | 282 | 46.1 | 6.0 | 9.4 | 1.4 | 3 |
| 9:05-9:15 | 232 | 46.7 | 5.8 | 8.9 | 1.3 | 1 |
| 9:15-9:25 | 185 | 46.7 | 6.2 | 10.2 | 1.5 | 4 |
|
Oct US-1 South Link 104 | 17:35-17:45 | 116 | 37.3 | 8.6 | 19.6 | 2.8 | 2 |
| 17:45-17:55 | 292 | 33.4 | 8.2 | 17.9 | 2.6 | 3 |
| 17:55-18:05 | 290 | 33.0 | 8.6 | 19.8 | 2.9 | 0 |
| 18:05-18:15 | 298 | 32.8 | 8.9 | 20.9 | 3.0 | 3 |
| 18:15-18:25 | 375 | 26.6 | 9.5 | 24.0 | 3.5 | 4 |
|
Nov US-1 North Link 105 | 8:15-8:25 | 283 | 15.0 | 5.7 | 8.5 | 1.2 | 2 |
| 8:25-8:35 | 299 | 16.1 | 5.3 | 7.4 | 1.1 | 7 |
| 8:35-8:45 | 282 | 15.9 | 6.0 | 9.5 | 1.4 | 3 |
| 8:45-8:55 | 271 | 17.6 | 5.2 | 7.2 | 1.0 | 5 |
| 8:55-9:05 | 271 | 12.8 | 4.1 | 4.5 | 0.6 | 9 |
| 9:05-9:15 | 243 | 15.1 | 4.8 | 6.1 | 0.9 | 7 |
|
Nov Duke St. Link 242 | 10:30-10:40 | 160 | 27.2 | 8.3 | 18.3 | 2.7 | 4 |
| 10:40-10:50 | 178 | 27.0 | 9.0 | 21.6 | 3.1 | 5 |
| 10:50-11:00 | 180 | 28.1 | 8.0 | 17.1 | 2.5 | 12 |
| 11:00-11:10 | 152 | 28.4 | 8.7 | 20.0 | 2.9 | 8 |
| 11:10-11:20 | 211 | 26.3 | 9.3 | 23.1 | 3.3 | 8 |
| 11:20-11:30 | 178 | 26.3 | 10.2 | 27.5 | 4.0 | 4 |
The results presented in Table 4 are consistent with the results for freeway
links presented in Table 3. For the more
stringent requirements, the WLT-based system only provided sufficient samples
in 2 of the 35 intervals. On the other
hand, when the requirements are relaxed to ±10 mile/hour error with a 95% confidence
interval, the WLT-based system met the sample size requirements in 26 of the 35
intervals.
In general, the sample size requirements are greater for arterial links
than for freeway links when controlling for the level of accuracy. This is to be expected given the larger variations
in speeds on arterials due to traffic control devices and greater number of access
points. Furthermore, the arterials chosen
for this case study are major arterials, and likely to experience less variation
than many arterial facilities. Thus, the results presented in Table 4 should
be considered as a “best case” scenario.
WLT-based System Link Condition Estimation Evaluation
Before summarizing the results of the link speed evaluation,
it is important to note that the WLT-based system was unable to produce a speed
estimate due to a lack of any samples in 4 out of 74 ten-minute intervals considered
in the evaluation. This illustrates that early-generation WLT-based systems struggle
with the ability to consistently produce condition data. Table 5 presents the
average 10-minute interval mean error for the link classifications considered
in this evaluation. Every interval was
considered, regardless if the sample size adequacy analysis conducted in the previous
section indicated that the interval did not have sufficient WLT samples.
TABLE 5.
Average
10-minute Interval Mean Absolute Error – WLT-based System
| Link Classification | Average Mean Absolute Speed Error
(mph) | Minimum Mean Absolute Speed
Error (Mph) | Maximum Mean Absolute Speed
Error (Mph) |
| I-495 (daytime) | 7.2 | 0.1 | 23.9 |
| I-495 (evening) | 9.2 | 0.5 | 22.8 |
| Arterials | 6.8 | 0.1 | 23.2 |
The results presented in Table 5 indicate general
agreement with the results of WLT-based system sample size adequacy analysis:
the WLT-based system sample size is frequently larger than the sample size requirement
on the allowable error of ±10 mph, but frequently smaller
than the requirement on the allowable error of ±5 mph. One will note that the maximum absolute mean
error reaches 23.9 mph for a 10-minute interval. This level of error is far greater than can be accepted for even
the least rigorous traffic management or traveler information application. Finally, Figure 1 presents a histogram of WLT-based
system mean errors, considering the “sign” of the error (indicating over- or under-estimation).
As seen in the figure, there are more than twice as many 10-minute intervals
in which the WLT-based system underestimates the mean speed as opposed to overestimating
the speed. This may indicate that the
WLT-based system is susceptible to including speed samples of stopped or slow
moving vehicles located near the link in question.
Figure 1. Interval Error Histogram
Finally,
hypothesis tests were conducted to investigate the significance of the differences
between WLT-based system speed estimates and the baseline data. The WLT- based system proved to be relatively
effective for the I-495 freeway links during daytime periods – with only 4.2%
of the intervals demonstrating significantly different mean speeds than the population
mean. On the other hand, the WLT-based
system produced poor performance for freeway links at night and arterial links.
In the case of I-495 links at night, 27.3% of the WLT-based system speed
estimates were significantly different from the population mean speed. Similarly, for arterials, 20.7% of the WLT-based
system speed estimates were significantly different from the population mean speed.
While
these results may appear to be fairly positive, one must consider the tests in
more detail. As stated earlier, the average
WLT-based system sample size is quite small (less than 4) for a number of 10-minute
intervals, especially on I-495 nighttime and arterials. On many occasions, the
WLT-based system only produced 1 or 2 samples of vehicle speeds on a link in a
10-minute interval. As a result, the small sample size significantly reduces the
“reliability” of hypothesis test results. In other words, it is important to recall the
underlying premise of a hypothesis test. When
one “fails to reject” the null hypothesis, this is not equivalent to stating that
that the mean speeds are equal. Essentially, there was just insufficient evidence to reject this
claim. In many cases, the evidence is
insufficient due to the very small sample sizes.
CHALLENGES POSED TO TRAFFIC ENGINEERING
Considering
the results of the case study, one can reach two relatively simple conclusions. First, based on the experience of this field
test, early-generation WLT-based traffic monitoring systems are not ready to provide
the accuracy and availability needed by modern traffic management systems (and
to a large extent, traveler information systems). Second, the ability of an early generation
system to generally measure speeds within an error of 10 percent on major routes
points to the potential of WLT-based systems.
Readers
will certainly note that the evaluation methodology required a number of rather
strong assumptions in order to compare the WLT-based system to traditional point
detector data. These assumptions were
necessary due to the fact that sampling with a WLT-based system (or any probe-based
system) is fundamentally different from sampling with point sensors.
This section explores these differences and the challenges that WLT-based
monitoring systems pose to traffic engineering.
First, the limitations that WLT-based monitoring shares with other probe-based
systems are reviewed. Next, the issues
specific to WLT-based systems are summarized.
Sampling Issues
One
of the difficulties of using any probe-based system is determining the relationship
between the characteristics of the probe sample and the entire population.
Using Equation 2, one can apply the Central Limit Theorem to determine
the number of samples needed to estimate the average link speed within some level
of allowable error at some confidence level.
However, this method assumes that the data within each measurement interval
is stationary and that the measured speeds for each vehicle are free of error.
The
stationary data assumption may be grounds for concern especially during peak hour
periods and incidents when speeds change rapidly with time. Shortening the measurement interval would reduce
the variation experienced during the interval. However, shortening the measurement interval would also increase
the total number of samples needed over some unit length of time, using more computational
resources. Non-stationary data is a problem
that is inherent in any sort of travel time or speed estimation system that relies
on data measured over some time interval and it must be understood when interpreting
results from Equation 2.
The
exact data assumption is also troubling in a WLT system. While technology may be improving, there will
always be some inaccuracies in measuring travel times and speeds. These inaccuracies are caused by wireless position
estimation error, vehicle map-matching error (including errors in distinguishing
between phones being used on the road and those being used in office buildings,
parking lots, etc.), and underlying error in the road position data on the map.
Such errors must be taken into account when using Equation 2 to determine
minimum sample size. Future research will
need to address these issues before WLT-based systems can become generally accepted.
Role
of Speed Variance
Another
issue encountered by probe-based systems is how to account for changes in population
speed variance. As this evaluation showed,
the speed variance of the population plays an important role in the minimum sample
size required to determine mean speed. A number of papers have attempted to investigate
how to deal with this issue in probe-based monitoring systems.
Chen
and Chien (2000) modeled a section of Interstate 80 in New Jersey with CORSIM
in an attempt to determine the minimum sample size needed for travel time estimation.
They found that, on account of heavy volumes and intense weaving situations,
some links may not have normally distributed travel times.
For those that did have normally distributed travel times, Chen and Chien
determined the number of probes required for five different volume levels, ranging
from 50% of the average flow rate to 150% of the average flow rate.
Their simulation showed that at 80% and 100% of the average flow rate,
only 3 samples were needed per 5-minute interval.
However, at very low and very high flow rates (when travel time variance
would presumably be greatest), as many as 12 samples were required per 5-minute
interval. This shows the importance of variance in determining
the minimum required sample size.
Holdener
and Turner (1996) used real data from toll tag systems in Houston to determine
the minimum sample size needed for vehicle speed estimation. They found that, for 5-minute collection periods,
the necessary sample size was between 1 and 4 for a 90th percentile
confidence interval and it was between 1 and 6 for a 95th percentile
confidence interval.
A
number of other papers have attempted to specify minimum required numbers of probe
vehicles using simulation and real data. However,
the differences seen in their results just go to emphasize the importance of speed
variance in determining minimum sample size and the location-specific nature of
that variance. While one particular road
in a particular area may require a certain number of samples most of the time,
another road in another area may have a completely different requirement.
Thus, minimum sample size determination must be performed on a location-by-location
basis.
Issues Specific to WLT-based Monitoring
While
there are similarities, WLT-based systems differ substantially from probe-based
traffic monitoring systems that use AVI tags.
With AVI tag systems, the location of an individual vehicle can be determined
with no ambiguity. WLT-based systems,
on the other hand, do not produce data at fixed points, and often have a considerable
amount of error in the position estimates that are reported.
This positional error can have a direct impact on whether a WLT-based system
can match a vehicle to a road, and also on the speed estimates generated for a
vehicle. These errors could potential require larger
sample sizes for WLT-based systems than probe systems based on AVI readers.
Procedures to deal with these issues need to be developed.
There
are additional sample size concerns that result from the need to reduce computational
overhead in a real-time system. In order
to provide travel time estimates based on wireless location data in real time,
the WLT requirements must not exceed the capabilities of the computational resources
that exist. Furthermore, systems that
assess a cost each time a phone is located (which appears to be a likely business
model for location-based services) make sample minimization a cost-saving measure
as well. Hence it is important to understand the minimum necessary sample
size so that the minimum number of vehicle tracks can be sampled to conserve both
processing and financial resources.
CONCLUSIONS
WLT-based traffic monitoring is an extremely appealing conceptual approach
to collecting traffic information. In
fact, the appeal of this concept has led to the relatively large field trials
conducted using WLT while the technical aspects of the technology and traffic
information derivation were still in their infancy. As seen in the results of the VDOT/MSHA/USWC
trial presented in this paper, an early generation WLT-based system produced link
speed estimates of moderate quality. This,
along with the theoretical capabilities of WLT, give reason for much optimism
for use of this technology in future traffic monitoring applications.
In order to make this optimism a reality, however, there is a need for concerted
effort to address the significant sampling challenges raised by such systems.
AVI tag-based probe vehicle systems provide a starting point, but WLT raises
a number of additional challenges. Based on the results of previous field tests
and the critical analysis presented in this paper, it is recommended that a basic
research program commence addressing the complex sampling and map matching challenges
that must be surmounted to make accurate, reliable WLT-based system monitoring
a reality.
REFERENCES
Cayford, R., and T. Johnson. “Operational
Parameters Affecting the Use of Anonymous Cell Phone Tracking for Generating Traffic
Information.” Proceedings of the Transportation Research Board 82nd Annual Meeting. Transportation Research Board, Washington,
D.C., 2003.
Chen, Mei and Steven I. J. Chien. “Determining
the Number of Probe Vehicles for Freeway Travel Time Estimation by Microscopic
Simulation.” Transportation Research
Record. No 1719: 61-68. 2000.
Smith,
B. L., Ulmer, J. M. “Traffic Flow Rate Measurement: An Investigation Into the Impact of the Measurement
Time Interval.” ASCE Journal of Transportation Engineering.
Vol. 18, No. 3. 2002. Vol.
129, No. 3. 2003.
TRB. Highway Capacity Manual. The
National Research Council, Washington, D.C. 2000.
Turner, Shawn M. and Douglas J. Holdener. Probe Vehicle Sample Sizes for
Real-Time Information: The Houston Experience. Proceedings of the Sixth International
Conference on Vehicle Navigation and Information Systems, Seattle, Washington.
1995.
University of Maryland Transportation Studies Center. Final
Evaluation Report for the CAPITAL-ITS Operational Test and Demonstration Program.
University of Maryland, College Park, MD, 1997.
Ygnace, J-L., J-G. Remy, J-L Bosseboeuf, and V. Da Fonseca. Travel
Time Estimates on Rhone Corridor Network Using Cellular Phones as Probes: Phase 1 Technology Assessment and Preliminary
Results. INRETS Report.
Arcueil, France. November, 2000.
Yim, Y.B.Y., and R. Cayford. Investigation of Vehicles as Probes Using Global
Positioning System and Cellular Phone Tracking: Field Operational Test. Report UCB-ITS-PWP-2001-9. California PATH Program, Institute of Transportation
Studies, University of California, Berkeley, CA, 2001.
