Number and Location of Sensors for
Real-Time Network Traffic
Estimation and Prediction
Sensitivity Analysis
Stacy M. Eisenman, Xiang Fei, Xuesong Zhou, and Hani S. Mahmassani
Installing and maintaining sensors in a transportation network can be
expensive. The motivation for this research is finding the best way to
deploy finite resources and generate a network detection system in a
manner that produces minimal estimation errors. The analysis uses a
simulation-based real-time network traffic estimation and prediction
system based on dynamic traffic assignment (DTA) methodology to ana-
lyze different levels of detection and different sensor locations in a por-
tion of the Chesapeake Highway Advisories Routing Traffic (CHART)
network (between Washington, D.C., and Baltimore, Maryland). This
study provides a conceptual framework of the sensor location problem
and a theoretical description of the objectives associated with the sensor
location problem. A sensitivity analysis of the estimation and prediction
quality with the DYNASMART-X real-time DTA system in relation to
sensor number and location is conducted in the Maryland CHART net-
work. The analysis considers both randomly generated location scenar-
ios and scenarios based on engineering judgment. The analysis reveals
the importance of providing detection in specific locations of the network
and the dependence of the value of additional detection on the specific
location selected.
Availability of real-time information on system status, enabled by
sensor deployment and advanced communications, is an integral ele-
ment of intelligent transportation systems. Information from various
types of sensors deployed to monitor network conditions is used for
traffic management, operations, and control purposes. Installing and
maintaining sensors and the associated communications infrastruc-
ture in a transportation network can be expensive. The motivation for
the research described is to make cost-effective decisions on deploy-
ing sensors in a network to support traffic management functions,
particularly real-time traffic status estimation and prediction.
The research is conducted with particular reference to the detec-
tion requirements of a simulation-based real-time network traffic
estimation and prediction system based on dynamic traffic assign-
ment (DTA) methodology. It provides an investigation of the sensi-
tivity of the quality of estimation and prediction to the number and
location of sensors in a traffic network. The specific network con-
sidered is part of the Chesapeake Highway Advisories Routing Traf-
Maryland Transportation Initiative, Department of Civil and Environmental Engi-
neering, University of Maryland, 3130 Jeong H. Kim Engineering Building, College
Park, MD 20742-3021.
Transportation Research Record: Journal of the Transportation Research Board,
No. 1964, Transportation Research Board of the National Academies, Washington,
D.C., 2006, pp. 253–259.
253
fic (CHART) network between Washington, D.C., and Baltimore,
Maryland. The study also provides a conceptual framework for the
sensor location problem and a theoretical description of the objectives
associated with the sensor location problem.
The sensitivity analysis of the estimation and prediction quality
is conducted with the DYNASMART-X (1) real-time DTA system.
The analysis considers both randomly generated location scenarios
and scenarios based on engineering judgment. The latter considers
placing detectors on high-volume links on the main freeways and
arterials. Taken together, the two sets of scenarios provide a useful
insight into the robustness of real-time DTA estimation and predic-
tion and the effect of location-specific considerations on estimation
and prediction quality.
BACKGROUND
The sensor location problem is a complex optimization problem that
can be difficult to solve because of its size and the fact that an optimal
solution may not exist. Previous work by Yang et al. (2) on the choice
of detector location focused on bounding the origin–destination (O-D)
error in the estimation of network performance. The idea is to limit
the O-D estimation error by locating traffic-counting stations so that
a certain portion of the trips between any O-D pair is observed for
at least one link of their path. This principle is referred to as the O-D
covering rule.
Extension of this work by Yang and Zhou (3) framed the problem
as an optimization problem focused on maximizing the total net
observed traffic flows. This optimization is then constrained by the
O-D covering rule (2). This approach is referred to as the maximal
flow-intercepting rule and link independence rule.
Additional work by Bianco et al. (4) proposes an iterative two-
stage procedure that focuses on maximizing coverage in terms of
geographical connectivity and size of O-D demand population.
The general approach used to address this problem relies on heuris-
tics, especially greedy algorithms (2). These algorithms essentially
seek to find the most important location first and locate a sensor there.
Then a second-most-important location is found and the algorithm
continues until reaching a prespecified termination criterion (number
of sensors or no significant improvement).
CONCEPTUAL FRAMEWORK
The sensor location problem could be viewed from the perspective
of the value of information. Sensors continuously provide infor-
mation that helps characterize the status of the network. Using this
254
information in conjunction with knowledge (i.e., historical data, pre-
vious estimation or prediction outputs) could enhance a model’s
estimation and prediction performance (see Figure 1). Adding sen-
sors to the network at specific locations could be evaluated for the
additional value that these sensors would provide to the ability to
estimate and predict network flow patterns (e.g., O-D demand, path
flows, link flows, point speeds), provide travel time information, or
provide better control strategies.
DYNASMART-X is a simulation-based real-time network traf-
fic estimation and prediction system that relies on DTA methodol-
ogy. This model system is dependent on two types of information:
a priori information (historical data, previous estimation or predic-
tion outputs) and real-time information (sensor data). The model is
a continuous learning model in that it receives real-time information
and uses an iterative process to incorporate this information and
improve the quality of estimation and prediction (5, pp. 305–328).
The algorithmic component of the DYNASMART-X system con-
sists of five modules: status estimation, status prediction, O-D esti-
mation (ODE), O-D prediction (ODP), and consistency checking and
updating. The purpose of the status estimation module (RT-DYNA)
is to estimate the current traffic status in the network. The status pre-
diction module (P-DYNA), in contrast, provides future network
traffic status for a predefined horizon. The ODE module is respon-
sible for estimating the coefficients of a time-varying polynomial
function that describes the evolution of the O-D demand process
from the current stage to the next stage. The ODP module utilizes
time to calculate the demand that is generated from each origin to
each destination at each departure time interval of the current and
future stages. Finally, the consistency-checking module is responsi-
ble for minimizing the deviation or discrepancy between what is
estimated by the system and what is occurring in the real world in
an effort to control error propagation. DYNASMART-X imple-
ments two levels of consistency checking: short term and long term.
The short-term consistency checking (STCC) compares the link
densities and speeds of the simulator to evaluate the consistency of
the flow propagation with the real world and correct the simulated
speeds. The long-term consistency checking (LTCC) calculates
scaling factors that are applied to time-dependent O-D demand in
the next running instance of RT-DYNA. An updating function goes
in parallel with the STCC and LTCC tasks.
Ideally, one would want sensors on all the links in the network. This
abundance would reduce the error associated with the status estima-
tion to the system error. Focusing on the sensor location problem, the
goal is to identify the locations that provide the most value given a
Sensor
Information
Estimation
“Knowledge”
Prediction
FIGURE 1 Estimation and prediction enhancement information.
Transportation Research Record 1964
limiting constraint on the number of detectors. In DYNASMART-X,
the sensor information is used in the ODE and ODP modules and
in STCC and LTCC tasks. One of the most important capabilities
of a real-time traffic simulation system is to estimate and predict
time-varying O-D demand adaptively with incoming real-time traf-
fic sensor data. Given status-mapping matrices, Kalman filter process
noise variance–covariance matrices, and measurement noise variance–
covariance matrices, the current implementation of O-D estimation
and prediction in DYNASMART-X aims to incorporate additional
information from newly added detectors to minimize the network per-
formance uncertainty. Thus, the importance of a sensor depends on the
value of information or knowledge that it can bring to the problem.
Given a priori information, one can estimate additional information
as follows:
D = D− + ξ (1)
where
D= estimate,
D− = a priori mean,
ξ= estimation error, and
ξ∼ Normal (0, P −), where P − is the a priori variance–covariance
matrix.
Depending on the desired estimate, a priori mean and variance
estimates, D− and P−, could be calculated from multiple data sources.
Whether one is interested in estimating O-D, path flows, link flows,
point-to-point travel times, or point speeds, this information could
come from static estimation and real-world observations from the
existing detectors. Regardless of the underlying estimation formu-
lations or methods (ordinary least squares or Kalman filtering) and
the original information sources, the current estimate can be also
characterized by its mean and variance statistics.
In essence, the sensor location problem is more than a simple cov-
erage problem. Even if the entire network could be covered by only
a few detectors, one would not be able to infer all information about
the network because the inferred values might be unreliable esti-
mates as a result of a limited number of observations. In this case,
more sensors must be located at critical points to enhance estimation
reliability.
In the Kalman filter framework, the simulated sensor matrix (H) can
be used to map sensor measurements to the estimate (D). The goal
of the sensor location problem would then be to determine a set
of links with additional measurements (C ) that minimize the un-
certainty of the resulting covariance matrix (P +). The set of links (C)
that have sensors (observation links) can be defined as follows:
C = H × D+ (2)
where the objective is to find C that minimizes P + and
H = simulated link proportions;
 = measurement error,  ∼ Normal (0,R);
R = measurement variance, and
P+ = (posterior) variance–covariance matrix.
In summary, the sensor location problem can be viewed as a traf-
fic status learning process that needs more sensors to add valuable
information that can be used to update estimates (in terms of mean
and variance) of the network traffic status. The key question here is
how to characterize the value of additional information from a new
detector in traffic status estimation and prediction.
Eisenman, Fei, Zhou, and Mahmassani
The updating stage in traffic status estimation and prediction
involves the calculation of the weighting matrix (gain factor):
K = P − H T ( HP − H T + R )
−1
(3)
Considering the weighting matrix that is produced in the updat-
ing processes, one should note that this process provides a way to
produce the a posteriori mean and covariance estimates:
D + = D − + K (C − HD − ) (4)
P + = ( I − KH ) P − (5)
where I is the identity matrix.
The foregoing covariance updating formula clearly links the a pri-
ori uncertainty and a posteriori uncertainty, and KH measures the
degree of uncertainty reduction due to inclusion of new measure-
ments. New measurements can come from a single detector or multi-
ple detectors. KH can be viewed as a matrix specifying the value of
additional information.
There are several details to note about the weighting matrix K. First,
it is the product of P −H. The a priori variance–covariance matrix (P −)
indicates the current uncertainty level for each unknown variable
and the covariance between different unknown variables. The sensor
matrix (H ) connects link observations to the unknown values. If a link
can intercept a value with a large variance or it can intercept more than
one unknown variable, the product of P −H is more likely to be large
and more uncertainty reductions are obtainable.
The second detail about the weighting matrix K is the inverse of
HP −H T. If one only considers HH T for multiple possible sensors, the
inverse of HH T specifies the correlation of measurements among
multiple links. Furthermore, HP −H T describes the measurement cor-
relation on the basis of existing estimated variance and covariance.
If HP −H T is large, meaning that either new sensor data could be highly
correlated with each other or they are correlated with the current esti-
mate, the inverse is small and the weight factor becomes insignificant.
It should be recalled that generalized linear regression (GLS) has
the same term (HP −H T )−1, indicating the extent of information or
knowledge obtained from observations.
The third detail to note is the inverse of R. Clearly, the more mea-
surement error, the less uncertainty reduction there will be for the
estimates.
This theoretical discussion is intended primarily to frame the
analysis conducted in the next section and to provide a framework
for thinking about and understanding the sensor location problem of
interest. Attempts to solve the optimization problem implied in this
discussion would be prohibitive from a computational standpoint in
anything other than small contrived networks, hence the motivation
to explore systematically the properties of the solution set to the
problem to inform practical decisions on where to locate a limited
number of sensors in a network.
SENSOR LOCATION ANALYSIS
A sensitivity analysis is presented of estimation and prediction
quality in relation to both sensor location and level of sensor cov-
erage in a network. The analysis is conducted with the simulation-
based DYNASMART-X real-time network traffic estimation and
prediction system on the Maryland CHART network. The purpose
is to explore the significance of adding detectors to the network and
255
also to learn how their location affects estimation and prediction
performance.
Two types of analyses were conducted. First a number of random
detector location scenarios were generated and analyzed. This set of
analyses illustrates how the number of detectors in the network can
influence the estimation and prediction results and also how distri-
bution in the network can produce various results. A second set of
detector location scenarios was generated by using engineering judg-
ment to place detectors on high-volume links, focusing on the main
freeways and arterials.
There are four components to this section. First, to clarify the
analysis, a description of the development and features of the Mary-
land CHART network is presented. In the second section the proce-
dure used to construct the detector information when observation data
were not available is explained. The third section reviews the differ-
ent measures that are used in the analyses, and the fourth presents
descriptions and results pertaining to the different scenarios that were
analyzed.
Maryland CHART Network Description
The applicable portion of the CHART network in Maryland was
developed for use in real-time traffic management. The study area
is concentrated around the I-95 corridor between Washington, D.C.,
and Baltimore. The network is bounded by I-695 to the north, I-495
to the south, US-29 to the west, and I-295 to the east. The network
includes four main freeways (I-95, I-295, I-495, and I-695) as well
as two main arterials (US-29 and US-1). The Maryland CHART
network reduces to 2,182 nodes, 3,387 links, and 111 zones. It also
includes 262 signals. Figure 2 shows the Maryland CHART network
and signal locations.
Within the study are only 14 existing detectors. Twelve of the detec-
tors are located on I-95 and two are located on I-495. These detectors
collect and report data in 5-min intervals. This detector information
can be obtained from the Center for Advanced Transportation Tech-
nology (CATT), the Maryland Department of Transportation, and
the Maryland State Highway Administration. Information describing
I-695
I-95
US 29
I-295
I-495
Rte 1
FIGURE 2 Maryland CHART network.
256
detector location as well as detector data are available from the CATT
laboratory website (6).
Each detector data file contains timestamp information, detector
location, traffic direction, vehicle counts, vehicles per hour, speeds,
and percent occupancy. Sensors collect 24-h data in 5-min intervals.
The percent occupancy refers to the percentage of time the detector
was occupied during the 5-min interval. The speed is the average
speed recorded over the 5-min interval. The vehicle count is the
number of vehicles observed during the 5-min interval. Vehicles per
hour is the 5-min vehicle count converted to an hourly flow rate
[e.g., for a 5-min vehicle count of 120 vehicles/h: (120 vehicles/
5 min) * (60 min/h) = 1,440 vehicles/h].
The DYNASMART-X prototype is calibrated and evaluated
according to its overall system functionality rather than its individ-
ual modules by using the available data, with possible enrichment
from other sources. The primary areas of calibration and evaluation
are traffic estimation, traffic prediction, consistency checking and
updating, and O-D estimation and prediction. Calibration and eval-
uation are performed at the overall system level. Calibration itself is
separated into two types: a priori calibration of structural relations
and real-time adaptive updating of the calibrated models and param-
eter values. For these purposes a set of real-time preprocessed data
was developed from the CATT laboratory databases by the Maryland
Transportation Initiative DTA group. Necessary checking and judg-
ment were exercised to retain consistency between the raw data and
the preprocessed data. Data for the 14 links with reliable real-time
data were processed for October 28 and November 1–5, 2004. These
data were used as the basis of the network calibration and validation.
Experiment Data Synthesis
For the experiments conducted in this research, real-time data avail-
able were limited. Therefore experimental data that are used to
mimic real-time sensor information were synthesized with a DTA
methodology (i.e., DYNASMART-P).
To start, there is an O-D demand table, which includes actual link
counts and a historical static demand table to estimate the O-D traffic
demand matrix. This matrix is treated as the ground truth for exper-
imental purposes. The ground truth O-D demand is loaded onto the
network with a DTA simulation program to generate both link counts
and density (simulated link measurements). The values become the
sensor data or observations in the synthetic data set.
To ensure the internal consistency between link flow measure-
ments and density measurements, this study uses simulated link mea-
surements as estimation input instead of the actual link observations
from the field data.
Analysis Measures
To interpret the influence of a given set of sensors on the ability to
estimate and predict network flow patterns, the root-mean-square
error (RMSE) of the link densities will be calculated for all of the
links in the network. (Generation links will not be included in these
calculations.) The calculation is as follows:
∑ (C − Cl′,t )
2
l ,t
RMSE =
l ,t
L ×T
Transportation Research Record 1964
where
Cl,t = observed density for link l during time interval t (ground
truth output),
C′l,t = simulated density for link l during time interval t (simulated
output),
L = set of links used in statistical calculations,
⎟ L⎟ = total number of links in set, and
T = number of time intervals.
In a given scenario, the RMSE is calculated across all links and
across all time intervals.
Detector Analysis Results
Each set of experiments was performed with a 6-h simulation, from
4:00 a.m. to 10:00 a.m. To ensure that network loading and network
discharge did not unduly influence the results, the analysis period
was reduced to 5 h (4:30 to 9:30 a.m.).
In the development of the sensor location scenarios, a few con-
straints were placed on the selection process. First the links sorted
on the basis of flow and the links with higher flow were considered
to be more attractive. In addition, when consecutive links did not
have access points between them, only one of the links was selected.
Also, if a link was selected for sensor location, the two upstream and
downstream links were selected for sensor placement. The two rules
were implemented to reduce correlation in the selected links and
produce larger coverage of the network. The adjacency rules were
not applied to ramps that connected to freeway links that had sensors.
These selection constraints will be referred to as filters.
Random Sensor Location Analysis
The first set of experiments focused on 20 scenarios in which sensors
were placed in the network on the basis of random selection. (The
selection of the sensor locations was not entirely random in that they
were selected at random from a subset of filtered links. This subset
included the top 220 links sorted by link flow and filtered to meet the
selection constraints.) The 20 scenarios are described as follows:
• Scenarios 1–5: 5 runs with 20 detectors chosen randomly from
the top 220 filtered detectors,
• Scenarios 6–10: 5 runs with 30 detectors chosen randomly
from the top 220 filtered detectors,
• Scenarios 11–15: 5 runs with 40 detectors chosen randomly
from the top 220 filtered detectors, and
• Scenarios 16–20: 5 runs with 80 detectors chosen randomly
from the top 220 filtered detectors.
Figure 3 shows a plot of the RMSE for the estimation and pre-
diction in each of the scenarios. From the plot one can observe the
effects of the random location selection. Within each level of detec-
tion (i.e., 20, 30, 40, and 80 detectors) the random locations clearly
produced variations in the results. In terms of estimation, Scenario 1
(20 detectors) is performing the worst, followed by Scenario 6
(30 detectors) and Scenario 2 (20 detectors). Also in terms of esti-
mation, Scenario 16 (80 detectors) is performing the best, fol-
lowed by Scenario 3 (20 detectors) and Scenario 5 (20 detectors).
Figure 4a–f shows the locations of these detectors in the network.
The fact that two of the top three best and worst scenarios in this
Eisenman, Fei, Zhou, and Mahmassani
RMSE (density)
Scenario Label
RMSE (Est. density) RMSE (Pred_5min)
RMSE (Pred_15min) RMSE (Pred_20min)
FIGURE 3 RMSE for randomly selected sensor locations.
analysis have 20 detectors emphasizes the value of good detector
placement. Given the ability to place 20 detectors in the network, one
would aim to place them to achieve the best results and not misplace
them and obtain the worst.
In the case of the three worst scenarios (Figure 4a–c), each lacks
significant coverage on I-95 (the freeway with the most traffic). Sce-
nario 6 is performing the best out of the three and has the most cov-
erage on I-95 as well as 10 additional detectors. An examination of
the three best scenarios also reveals a few commonalities. The most
obvious is that each of these scenarios provides significant detection
on I-495 (the east–west freeway at the southern edge of the network).
In addition, each of these scenarios appears to provide detection at or
around freeway access points throughout the network.
Judgment-Based Sensor Location Analysis
The second set of detector location scenarios was generated by using
engineering judgment to place detectors on high-volume links,
focusing on the main freeways and arterials. This set of scenario
analyses should reveal the benefits of adding additional detectors to
specific areas in the network. This analysis includes the nine sce-
narios described as follows (each of the scenarios conforms to the
filtering criteria):
• Scenario 21: top 10 links on I-95 southbound (SB) and I-95
northbound (NB) (20 links total);
• Scenario 22: top 5 links on I-95 SB, I-95 NB, I-295 SB, and
I-295 NB (20 links total);
• Scenario 23: top 10 links on I-95 SB and I-95 NB and top 5 links
on I-295 SB and I-295 NB (30 links total);
• Scenario 24: top 5 links on I-95 SB, I-95 NB, I-295 SB, I-295 NB,
US-1 SB and US-1 NB (30 links total);
• Scenario 25: top 10 links on I-95 SB and I-95 NB and top 5 links
on eastbound and westbound cross roads (30 links total);
• Scenario 26: top 5 links on I-95 SB, I-95 NB, I-295 SB, I-295 NB,
US-1 SB, US-1 NB, US-29 SB, and US-29 NB (40 links total);
257
RMSE (Pred_10min)
• Scenario 27: top 10 links on I-95 SB and I-95 NB and top 10 links
on eastbound and westbound cross roads (40 links total);
• Scenario 28: top 10 links on I-95 SB and I-95 NB, top 5 links on
I-295 SB and I-295 NB, and top 5 links on eastbound and westbound
cross roads (40 links total); and
• Scenario 29: top 10 links on I-95 SB, I-95 NB, I-295 SB,
I-295 NB, US-1 SB, US-1 NB, US-29 SB, and US-29 NB (80 links
total).
Figure 5 shows a plot of the RMSE for the estimation and predic-
tion for each of the scenarios. The results of this analysis are interest-
ing. The set of scenarios was developed to allow for the exploration
of trade-offs in locating the detectors on different freeways, arterials,
and cross roads. With this in mind, comparing Scenario 21 (10 links
on I-95 SB and I-95 NB) and Scenario 22 (5 links on I-95 SB, I-95 NB,
I-295 SB, and I-295 NB), both 20-detector scenarios, one can con-
clude that locating detectors on I-95 is more valuable than placing
them on I-295. This result is consistent with the trends observed in
the random selection analysis.
The next sets of comparisons provide less obvious insights. A com-
parison between Scenario 21 and Scenario 23 or 24 shows that Sce-
narios 23 and 24 produce no significant changes in performance even
though there are more detectors. The additional detectors in this sce-
nario were placed on an arterial with much lower volume, and the
detectors could not produce the same level of performance even though
there were more of them. A similar result is obtained with Scenario 26,
whose performance does not improve over that of Scenario 25. This
result can be credited to the reduction in detectors on I-95.
Scenarios 22 and 24 are subsets of Scenario 26. Scenario 26 per-
forms the worst but has the most detection, whereas Scenario 22 per-
forms the best and has the least detection. The explanation for this
finding is that the additional detectors were placed on arterials with
much lower volume and the model is in conflict in trying to match
both the freeway and arterial detector information. An approach that
can be used to accommodate this conflict and the model’s ability to
best manage this situation would be to provide a weighting scheme
that placed a high value on links with higher volume.
258 Transportation Research Record 1964

(a) (b)

(c) (d)

(e) (f)

FIGURE 4 Sensor locations for (a) Scenario 1, (b) Scenario 6, (c) Scenario 2, (d) Scenario 16, (e) Scenario 3, and (f ) Scenario 5.
Eisenman, Fei, Zhou, and Mahmassani
13
12.5
12
RMSE (density)
11.5
11
10.5
10
9.5
9
21 22 23
RMSE(Est Density)
RMSE(Pred_15min)
FIGURE 5 RMSE for judgment-based sensor locations.
Scenario 29 is a scenario with 80 detectors. Scenario 26 is a subset
of this scenario. As is expected, Scenario 29 outperforms Scenario 26.
In this case, critical freeway detectors were added on I-95 and I-295
in addition to the detectors on the minor arterials US-1 and US-29.
Scenarios 25, 27, and 28 all consider the addition of detectors to
east–west cross roads. Again, the results imply that the addition of
detectors on lower-volume arterials produces a decline in estimation
performance.
Overall, these results suggest that high-volume freeways are more
valuable as sensor locations than low-volume arterials. The analy-
sis also suggests that increasing the number of detectors on freeways
is valuable.
Joint Analysis Results
Looking at the results from both analyses, the random selection
method of sensor location produced a lower RMSE. There are several
reasons for this occurrence. First, a random selection of the sensor loca-
tions is likely to provide less correlation than the scenarios that were
developed on the basis of engineering judgment. The second reason
for the better performance in the random analysis scenarios is that they
included detectors on I-495 and I-695, whereas the judgment-based
analysis did not. Freeways I-495 and I-695 are high-volume freeways
that can greatly influence the estimation performance in the network,
and possibly the model would have performed better in the judgment
scenarios had these freeways not been excluded.
CONCLUDING COMMENTS
Increasingly, sensors or detectors are being deployed to monitor net-
work conditions. Installing and maintaining sensors in a transporta-
tion network can be expensive. Ways were explored in this research
259
24 25 26 27 28 29
Scenario Label
RMSE(Pred_5min) RMSE(Pred_10min)
RMSE(Pred_20min)
to allocate resources so as to generate a network detection system
in a manner that produces minimal estimation errors and minimal
equipment costs.
The conceptual framework presented interprets the sensor loca-
tion problem as a value-of-information problem, which leads to
interpretation-with-learning process models. The analysis provided
several valuable insights about the process of selecting the locations
for sensors in a network. The difficulty of determining the best loca-
tions on the basis of judgment alone is an important caveat of the
study. A second valuable result was the heavy emphasis on improve-
ment added by placing sensors on high-volume freeways. Additional
analysis and methodological approaches to solve the problem articu-
lated remain part of the research agenda as ongoing research.
REFERENCES
1. Maryland Transportation Initiative. DYNASMART-X Version 0.930 User’s
Guide. U.S. Department of Transportation, 2003.
2. Yang, H., Y. Iida, and T. Sasaki. An Analysis of the Reliability of an
Origin-Destination Trip Matrix Estimated from Traffic Counts. Trans-
portation Research, Vol. 25B, 1991, pp. 351–363.
3. Yang, H., and J. Zhou. Optimal Traffic Counting Locations for Origin-
Destination Matrix Estimation. Transportation Research, Vol. 32B, 1998,
pp. 109–126.
4. Bianco, L., G. Confessore, and P. Reverberi. A Network Based Model for
Traffic Sensor Location with Implications on O-D Matrix Estimates.
Transportation Science, Vol. 35, 2001, pp. 50–60.
5. Mahmassani, H. S., and X. Zhou. Transportation System Intelligence:
Performance Measurement and Real-Time Traffic Estimation and Pre-
diction in a Day-to-Day Learning Framework. In Festschrift in Honor of
Professor Pravin Varaiya (E. Abed, ed.), Springer-Birkhauser, 2005.
6. Center for Advanced Transportation Technology, University of Maryland,
College Park, Md., July 2004. http://www.cattlab.umd.edu/cf/index.
cfm?js=enabled&bin=trafficData.
The Planning and Environment Group sponsored publication of this paper.