Travel Time Estimation with Correlation

Analysis of Single Loop Detector Data

Huairui Guo and Jionghua (Judy) Jin

In this paper, the average travel time over a link is considered as a random where g is the car length factor that is a function of the lengths of
variable following an identical probability distribution as the arrival vehicles and detectors. The average travel time (T) is estimated by
process. A new estimation method of the average travel time uses a cross-
correlation analysis of traffic flow measurement data. This method requires s
T= (2)
only traffic flow information, which is available from the measurements v
of single loop detectors upstream and downstream from one link. Dif-
ferent from the existing maximum cross-correlation analysis method, where s is the length of a given link.
the proposed method considers average travel time as a random variable, Although the model of Equations 1 and 2 provides a simple way
with its mean value estimated from all significant cross-correlation coef- to estimate average travel time, its accuracy is sensitive to the car
ficients rather than from only the maximum cross-correlation coefficient. length factor g, which is further elaborated later. In addition, the model
Therefore, the inherent variability of average travel time among different does not consider travel time variability among vehicles.
vehicles can be considered. Moreover, different from the existing opti- Recently, some other techniques were developed to estimate travel
mization method, the proposed method uses the statistical t-test of the time. For example, Coifman used instant velocity measurement by
significant cross-correlation coefficients to determine automatically and
a dual loop detector (4). This method is effective for links that have
adaptively the fitting range of the probability density function of the aver-
been equipped with dual loop detectors. However, it is not suitable
for links that are equipped with only single loop detectors, because a
age travel time. Thus, it avoids using the approximated car length factor
single loop detector cannot measure speed directly.
and has no need to predetermine the range of the average travel time as
Another technique is to use a cumulative flow plot to calculate total
required by the optimization method. Details of the average travel time
travel delay and estimate average travel time (5, 6). This technique
estimation procedures are presented, and the effectiveness of the proposed
requires the priori knowledge about the number of vehicles existing
method is demonstrated through both simulation study and a case study
in the link initially, which usually is not available in practice. In addi-
of real traffic data.
tion, there is no systematic way to choose the start time to plot the
departure curves. Thus, implementation of this technique is limited.
The accurate estimation of link travel time is essential to the develop- Nam and Drew report use of a cumulative flow plot to estimate the
ment of an intelligent transportation system. Travel time estimation travel time in a detailed analysis and present the new development
provides not only valuable information for traveler routing or trans- of the flow–density–speed relationship of their models (7). How-
portation scheduling but also the potential capability for incident ever, this method has the same problem as other methods that use a
detection. Single loop detectors have been widely used in arterial roads cumulative flow plot (5, 6).
and freeways for traffic condition monitoring and control. Generally, Some different methods were developed to estimate travel time
by applying advanced signal processing and intelligent classification
such detector data can be used to measure traffic flow directly but
techniques such as state space modeling, neural network classifica-
not vehicle speed and travel time.
tion, and car identification and signature matching. Park et al. use
Various models have been developed to estimate travel time. Among
neural networks to forecast real-time travel time for 5 min ahead (8).
them, the most simple and fundamental model is to use the basic
Hoogendoorn et al. propose a method that describes state space dynam-
relationship among speed, volume, and occupancy (1–3). The basic
ics of traffic with recurrent neural networks (9). Both methods require
idea of this model is briefly reviewed as follows.
the appropriate selection of training data sets to conduct extensive
For a given link, the relationship of average speed (v), flow (q), and
training of the neural network models. Coifman and Cassidy devel-
occupancy (o) is expressed as
oped an algorithm to estimate travel time through vehicle reidenti-
fication, in which vehicle length is used as a “signature” to match the
q
v= (1) vehicle in the downstream to the vehicle in the upstream (5). Thus,
oig the vehicle length must be measured or estimated accurately to ensure
the vehicles are correctly reidentified.
To consider the stochastic nature of traffic, some stochastic models
H. Guo, Reliasoft Corporation, 115 South Sherwood Village Drive, Tucson, AZ 85710. of traffic flow have been developed in the past decade. Dailey uses
J. Jin, Industrial and Operations Engineering Department, University of Michigan, cross-correlation analysis to estimate travel time, assuming that the
1205 Beal Avenue, Ann Arbor, MI 48109-2117. average travel time is a deterministic variable (10). Because it uses
Transportation Research Record: Journal of the Transportation Research Board,
only the maximum cross-correlation coefficient, it is effective when
No. 1968, Transportation Research Board of the National Academies, Washington, different vehicles have close travel speeds. More recently, Dailey
D.C., 2006, pp. 10–19. modifies the model of Equation 1 by considering the measurement

10
Guo and Jin 11

errors of the car length and occupancy (11). The modified model REVIEW: ESTIMATING TRAVEL TIME WITH
considers the occupancy, speed, and vehicle length as random vari- FLOW MEASUREMENTS
ables following certain probability distributions. Then, a Kalman filter
is used to estimate the actual speed of each vehicle. This method Probabilistic Regression Model
requires a prior understanding of the probability distribution function
of various random variables involved in the model. A significant The proposed travel flow model is based on the probabilistic regres-
estimation error—even instability or nonconvergence—may occur sion model proposed by Petty et al. (12). It assumes that the arrivals
if the mean or variance of the random variables were not properly measured at the upstream point within a given time interval (called
estimated. the estimation window) have the same probability density function as
In addition, by considering the random characteristics of average the average travel time over the link. If at sampling time t the numbers
travel times, Petty et al. developed a probabilistic regression model of upstream and downstream arrivals during the sampling interval
with the assumption that average travel time is a random variable Δ are x(t) and y(t), respectively, then the relationship between these
upstream and downstream arrivals can be modeled as
with the same probability distribution function over a given link for
a given estimation window of travel time (12). Although this method a2
is robust to congestion conditions, its effectiveness depends highly on y ( t ) = ∑ [ x ( t − i ) fi ] (3)
prior knowledge of the fit range of the probability density function i = a1

of the average travel time. Petty et al. adaptively estimate the center
of the fit range on the basis of Equation 1 with a fixed value of g. The where fi is the probability density function of the average travel time
change of the car length factor due to the instant change of occupancy over the link, and a1 and a2 correspond to the shortest and longest
is not considered; this could lead to an estimation error on the fit average travel times, respectively, over all upstream vehicles; (a1, a2)
center of the probability density function. Moreover, the selection is called the fit range of the probability density function. Petty et al.
of a fit range also affects estimation performance. The influence of estimate fi to minimize the sum of squares of the regression residual
these two factors on the estimation of average travel time are further errors by solving the following constrained optimization problem:
illustrated later.
⎪⎧ 2 ⎡ ⎤ ⎫⎪
2
In summary, the existing travel time estimation methods limited b a2

by the use of traffic flow data from single loop detectors can be clas- min ⎨ ∑ ⎢ y ( t ) − ∑ x ( t − i ) fi ⎥ ⎬ (4)
⎩⎪ t = b1 ⎢⎣ ⎥⎦ ⎭⎪
fi
i = a1
sified in three categories. The first method is rooted in the model
of Equation 1 for travel speed estimation, which is sensitive to the
car length factor estimation error. The second method considers the subject to
stochastic nature of the traffic flow and uses the maximum cross-
correlation coefficient between the upstream and downstream flow ⎧ a2
⎫
data to estimate average travel time, which ignores travel time vari- ⎨ f: fi ≥ 0 ∑f i = 1⎬
ability among different vehicles. The third method develops a prob- ⎩ i = a1 ⎭
abilistic regression model to estimate the probability density function
where b1 and b2 are the start time index and end time index of the
of the average travel time, which requires prior determination of the
estimation window (b1, b2) (12). Thus, the mean estimate of the
fit range of the estimated probability density function of the average
average travel time is
travel time. Therefore, the aim of this paper is to improve existing
methods of travel time estimation on the basis of flow measurement a2
data from upstream and downstream single loop detectors. Tˆ = Δ ∑ fi i i (5)
In this paper, a new method is developed to estimate travel time i = a1

by integrating the cross-correlation analysis technique with the prob-

abilistic model of the random average travel time. Resembling the From Equation 4, it can be seen that if the predefined fit range of
method of Petty et al. (12), a probabilistic regression model is used to (a1, a2) is too wide, there will be too many parameters fi to be estimated;
consider the randomness of the average travel time. The advantage of this may lead to an unstable solution in the optimization solver. Mean-
the proposed analysis method is to improve estimation by automat- time, if the fit range is too narrow, the resolution and accuracy of
ically determining the fit range of the probability density function of the estimated probability density function may be too poor to get an
the average travel time adaptively. Moreover, the proposed estimation appropriate probability distribution function. Therefore, it is critical
method does not require the estimation of a car length factor, which to define an appropriate fit range of (a1, a2). As reported by Petty
may lead to significant estimation error. The only information needed et al. (12), the fit range is fixed as 20 s, independent of the occupancy
is the downstream and upstream flow measurement. change, and the center of the mean travel time is adaptively estimated
In the following sections, the probabilistic regression model is on the basis of Equation 2 with a fixed car length factor of g in Equa-
briefly reviewed, and the impact of errors in car length factor estima- tion 1. The following analysis illustrates how this assumption of
tion on the existing travel time estimation method is discussed. Then, fixed car length affects the fit center of (a1, a2) used in the travel time
a new cross-correlation analysis method is proposed to consider the estimation of Equation 4.
travel time variability among different vehicles. Statistical hypothesis
testing is used to automatically determine the fit range of the prob- Impact Analysis of Car Length Factor
ability density function of the average travel time. Next, a simulation Estimation Error
study using VISSIM software and a case study of real traffic data
are conducted to demonstrate the analysis procedures and the effec- Hall and Persaud thoroughly discuss the relationship between
tiveness of the proposed method. Finally, findings are summarized speed, occupancy, and car length factor g (2). Their results show a
and conclusions presented. strong correlation between occupancy and average car length. With
12 Transportation Research Record 1968

increasing occupancy, average car length also increases. So, car length ΔT − Δ l
= (8)
should be adaptively estimated for different occupancies. Now, how Tˆc l
the estimation error of using a fixed car length factor affects the
accuracy of travel time estimation is further illustrated. From Equation 8 it can be seen that the estimation error of the
From Equations 1 and 2, the fit center, T̂c, of (a1, a2) used in the average travel time is proportional to the estimation error of the
Petty et al. estimation method (12) is obtained as average car length. If the overall average car length of 6.52 m is used
for the value of l in Equation 6, the range of Δl/l will be (−0.29, 0.36)
sio in the above simulation. The effect of the estimation error of the aver-
Tˆc = (6)
qil age car length on the estimated average travel time T̂c is obvious. So,
if this inaccurate T̂c value is further used as the center of the fit range
where l (which represents the average effective car length, or the sum (a1, a2), an estimate of average travel time could be biased because
of the average car length and the width of the loop detector) is equal the predefined fit width Δ = a2 − a1 may not be sufficiently cover all
to 1/g. When dual loop detectors are used, traffic flow (q) and traffic possible travel time values under different occupancy conditions. The
occupancy (o) are directly obtained and adaptively used to estimate influence of using a fixed car length on the travel time estimation is
T̂c, which is used as the center of the fit range of (a1, a2) in every esti- further illustrated through a simulation analysis later, in the section
mation window of (b1, b2). In this case, a fixed car length is used in on the comparison of different travel time estimation methods.
Equation 6. The resulting deviation error Δl due to incorrectly using In the following sections, a new method is proposed to automatically
the fixed average car length can cause the estimation error ΔTc as determine the fit range of (a1, a2) on the basis of statistical correlation
analysis and t-test.
sio
ΔTc = − Δl (7)
l2 i q
PROPOSED CORRELATION ANALYSIS METHOD
The impact of Δl on ΔTc is further investigated through a simula-
tion study with VISSIM simulation software. The conditions used in Upstream Flow and Downstream Flow
the simulation study are set as follows: link distance s = 300 m, aver-
age traffic flow q = 1,000 vehicles/h (0.28 vehicles/s). The estima- Two data series x(t) and y(t) denote travel flow upstream and down-
tion window is 1 min. For each estimation window, the occupancy stream, respectively. The cross-covariance γyx(k) of x(t) and y(t) and
and the corresponding average car length are calculated from simu- the autocovariance γx(k) of x(t) with lag k are defined as
lation data. The relationship between occupancy and average car
length is plotted in Figure 1. The overall average estimated car length γ yx ( k ) = E {[ y ( t ) − u y ][ x ( t − k ) − ux ]} (9)
is 6.52 m, and the corresponding occupancy is 0.12. Dividing Equa-
tion 7 by Equation 6 yields γ x ( k ) = E {[ x ( t ) − ux ][ x ( t − k ) − ux ]} (10)

8
Average Car Length (meters)

6.52

0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24
Occupancy (x 100%)

FIGURE 1 Correlation between average car length and occupancy.

Guo and Jin 13

Thus, the cross-correlation ρyx(k) and autocorrelation coefficients Hypothesis Testing of

ρx(k) are Cross-Correlation Coefficients

γ yx ( k ) The aforementioned analysis indicates that the discretized average

ρ yx ( k ) = (11) travel time distribution can be estimated directly from the nonzero
σ xσ y
cross-correlation coefficients. In practice, the true value of ρyx(k)
is unknown. Thus, the sample cross-correlation function ρ̂ yx(k) is
γ x (k )
ρx ( k ) = (12) used, which is calculated from the flow data within the estimation
σ x2 window (b1, b2):

where σx and σy are the standard deviation of x(t) and y(t), respec- b2

tively, and x(t) and y(t) are assumed to be stationary series within the ∑ [ x ( t − k ) − x ][ y ( t ) − y ]
ρ̂ yx ( k ) =
t = b1 + k
estimation window of (b1, b2). From Equation 3, (17)
⎪⎧ 2 ⎫
∑ {[ x (t ) − x ] }
b2 − k
2⎪
b

⎨ ∑ [ y (t ) − y ] ⎬
2

a2 t = b1 ⎪⎩ t = b1 + k ⎭⎪
u y = ∑ fi ux (13)
i = a1 where

where uy = E(yt) and ux = E(xt). On the basis of Equations 3 and 9–13, ⎡ b2 ⎤

⎢∑ x (t )⎥
⎢ t =b ⎥⎦
1 ⎧ ⎡ a2 a2
⎤ ⎫ x= ⎣ 1
ρ yx ( k ) = i E⎨
⎢ ∑ fi x ( t − i ) − ∑ fi ux ⎥ [ x ( t − k ) − ux ]⎬ ( b2 − b1 + 1)
σ xσ y ⎩ ⎣ i = a1 i = a1 ⎦ ⎭
and
a2
1
=
σ xσ y
i
∑ f γ (i − k )
i = a1
i x
⎡ b2
⎢∑ y (t )⎥
⎤
⎢ t =b ⎥⎦
y= ⎣ 1
σx a2
( b2 − b1 + 1)
=
σy
i
∑ f ρ (i − k )
i = a1
i x (14)
To determine the nonzero cross-correlation coefficients, the
The unknown parameters fi [i ∈ (a1, a2)] can be estimated through following statistical hypothesis test is used:
the correlation analysis of upstream and downstream flow. H0:
Although the proposed approach uses the correlation analysis to ρ yx( k ) = 0
estimate the travel time, there is a major difference from Dailey’s
maximum cross-correlation method (10). Dailey assumes that the H1 :
average travel time is one fixed value and uses only the maximum ρ yx( k ) ≠ 0 (18)
cross-correlation coefficient for average travel time estimation and
thus does not consider the variability of average travel time due When the estimation window is appropriately selected to include
to different travelers. In the proposed method, the average travel enough number of data, ρ̂yx(k) under H0 approximately follows a
time is assumed to be a random variable over different travelers, and normal distribution with a mean of 0, and the standard deviation σ ρ̂ is
Equation 14 is used to estimate a set of fi as the discretized proba-
1
bility density function by the sampling interval Δ, in which multiple σ ρ̂ = (19)
significant correlation coefficients of ρyx(k) will be used as follows. n
Under free-flow conditions, upstream x(t) are assumed to be
where n [the number of data points used to estimate ρ̂yx(k)] is equal
independent samples, thus
to b2 − b1 − k + 1 (13). Thus, a t-test can be used to check whether
ρxy(k) = 0 under a given α value. The statistic tρ,k is defined as
⎧⎪1 i=k
ρ x (i − k ) = ⎨ (15) ρˆ yx ( k )
⎪⎩0 others t ρ ,k = (20)
σ ρˆ
On the basis of Equations 14 and 15,
Under the given α error, if the condition of tρ,k > t1−α/2,n holds, H0 is
rejected. Thus, the cross-correlation at lag k is significant, and fk cal-
σy culated from Equation 16 is the estimated probability corresponding
fk = ρ yx (k ) (16)
σx to the average travel time k  Δ.
Under severe congestion conditions during the estimation window,
From Equation 16, fk can be determined solely by the cross- no significant cross-correlation coefficients will be found. In fact,
correlation coefficients of ρyx(k). The task of estimating fk is trans- under congestion, it is more important to monitor traffic condition than
ferred to find the nonzero cross-correlation coefficients, which is to estimate travel time. The lack of significant cross-correlation coef-
simpler and straightforward. The proposed method does not need ficients is an effective method of detecting congestion. One example
to predetermine the car length factor and the fit ranges of (a1, a2) of how to automatically determine significant cross-correlation coeffi-
for fi. Moreover, the proposed method using Equation 16 can avoid cients is presented in the section about the simulation and case studies.
the unstable solution problem that occurs in the optimization The proposed model is based on cross-correlation values of up-
approach. stream and downstream volume, which require real-time flow infor-
14 Transportation Research Record 1968

mation from each sampling interval. Traditional flow-density mod- So, the mean estimate of the average travel time based on these 100P%
els do not have this information and thus cannot be used to validate upstream vehicles is
the proposed model analytically. Therefore, VISSIM is used to con-
cm
duct a series of simulations and study how the traffic volume will
μˆ T = Δ i ∑ [( fi P ) i i ] (21)
affect the cross-correlation values. A 300-m-long link with a capac- i = c1
ity of 1,900 vehicles/h is used. The estimation window is 10 min.
According to the above analysis, t0.95,600 = 1.96. Therefore, the thresh- The mean estimate using Equation 21 ignores the effect of the
old value of rejecting H0 is 0.08. The traffic flow is set initially at small probability of the vehicles having either extremely long travel
200 vehicles/h and ends at 1,900 vehicles/h, with step increases of time (or never arriving at the downstream detector) or extremely short
100 vehicles/h. For each setting, the simulation lasted 1 h. Averages travel time; this provides a more realistic value of the mean estimate
of all the maximum cross-correlation values of each estimation win- of the average travel time. With the sample cross-correlation function
dow are calculated and plotted in Figure 2. of Equation 17 substituted into Equation 21, the mean estimate of the
Under the free-flow condition, increases in traffic volume can lead average travel time can be obtained:
to decreases in the maximum cross-correlation coefficients. However,
for all the volumes that are less than the road capacity, the cross- cm

correlation values are >0.08, which indicates a strong cross-correlation ∑ {⎡⎣ρˆ (i )⎤⎦ i i}
i = c1
yx

between upstream and downstream traffic flows. Therefore, the pro- μˆ T = Δ i cm

(22)
posed method can be used effectively to estimate travel time under
the free-flow condition.
∑ ⎡⎣ρˆ yx (i )⎤⎦
i = c1

Mean Estimate of Average Travel Time

SIMULATION STUDY AND CASE STUDY
The estimation of fi using Equation 16 considers only those nonzero
probability values of fi that correspond to a certain percentage of Description of Simulation Conditions
vehicles that have the same average travel time over the link. The
The four simulation studies conducted are discussed in the following
summation of all estimated fi represents the percentage of upstream
subsections. VISSIM simulation software was used to generate the
vehicles within the estimation window of (b1, b2) used for estimating
travel flow data and obtain references of the true travel time values
average travel time, that is,
for the simulation studies. The total simulation time was 5 hours, but
cm only data generated in the past 4 hours were used for the analysis.
∑f
i ∈c1
i =P The common conditions used for the four simulation studies are

• traffic flow q = 1,000 vehicles/h,

Therefore, among the upstream vehicles used for the travel time esti- • speed v = 45 to 60 km/h (12.5 to 16.67 m/s),
mation, 100 fi /P% of vehicles have the travel time ci  Δ, i = 1, . . . , m. • link distance s = 300 m,

0.4

0.35

0.3
Cross-Correlation Values

0.25

0.2

0.15

Threshold = 0.08
0.1

0.05
200 400 600 800 1000 1200 1400 1600 1800 2000
Flow Setting

FIGURE 2 Cross-correlation between x and y under different flow conditions.

Guo and Jin 15

• data sampling interval of flow measurement Δ = 1 s, and 25

• width of the estimation window b2 − b1 = 10 min. Δ i ∑ { ⎡⎣ρˆ xy ( i ) ⎤⎦ i i}

μˆ T = i = 20
25
= 22.35
∑ ⎡⎣ρˆ xy (i )⎤⎦
i = 20
Cross-Correlation Analysis and t-Test

This simulation study illustrates the procedures of the cross-correlation This analysis indicates that, with the t-test, the proposed estimation
analysis and the t-test discussed in the section on hypothesis test- method has no requirement on the preselection of the fit range of
ing of cross-correlation coefficients. As an example, upstream (a1, a2) and the optimization solver is not needed.
and downstream data are selected from an arbitrary estimation
window. The sample cross-correlation function ρ̂ yx (k) (0 < k ≤ 35)
is calculated on the basis of Equation 17, as shown in Figure 3. Comparison of Different Travel Time
Because n = b2 − b1 − k + 1 is large, the value of t1−α/2,b2−b1−k+1 will not Estimation Methods
depend on the lag k. If α error is selected as 5%, then t1−α/2,b2−b1−k+1 =
t0.95,600−k+1 = 1.96. So, on the basis of Equation 20, the threshold η This simulation is used to compare the performance of different
for determining the significant cross-correlation coefficients is methods of travel time estimation. As is discussed earlier, two param-
obtained as eters need to be set using the optimization method of Petty et al. (12).
One is the width of the fit range, equal to a2 − a1 = 20 s. The other is
t 0.95 ,600 − k +1 the fixed average car length (used in Equation 6 for the estimation
1.96 1.96
η= = ≈ = 0.08 of the center of the fit range). For comparison of the optimal perfor-
n 600 − k + 1 600 mance of the Petty et al. method (ignoring the bias in the estimation
of average car length), the true average car length of 6.5 m, obtained
From Figure 3, it is easy to see that the cross-correlation coefficients from VISSIM software, is used in the simulation. The travel time
from lag 20 to lag 25 are significant because they are larger than the estimation results—obtained by using the optimization method, a
threshold, 0.08. So, on the basis of Equation 2, the mean estimate of single (maximum) cross-correlation coefficient, and the proposed
the average travel time over the link is obtained as method—are all compared with true travel time in Figure 4.

0.2

0.15
Cross-Correlation Coefficent Values

1.96 σp̂
0.1

0.05

-0.05
1 5 10 15 20 25 30 35
Lag k

FIGURE 3 Cross-correlation analyses between upstream and downstream.

16 Transportation Research Record 1968

25

Optimization method

24 Single cross-correlation method

Proposed method

23
Travel Time (second)

22

21

Reference travel time

20

19
0 5 10 15 20 25
Sampling Time (10 mins)

FIGURE 4 Comparison of different travel time estimation methods.

For further comparison, the estimation errors of each method ure 5 shows that under the given simulation conditions, 600 s is a
(i = 1, 2, 3) are calculated by the deviation ei of the estimated travel relatively good estimated window length for all three estimation
time μ̂ Ti from the true travel time μ̂ T0, that is, ei = μ̂ Ti − μ̂ T0. The aver- methods.
age and standard deviation of the estimation errors of each method When the maximum cross-correlation method is compared with the
are compared in Table 1, which indicates that the performance of the proposed method and the estimated window length is <300 s, esti-
proposed method is better than that of the other two methods. mation performance is similar for both methods (Figure 5). However,
as the estimated window length increases, the performance of the pro-
posed method is better than that of the maximum cross-correlation
Effect of Window Length Estimation on
method, because more upstream arrivals with different travel times
Travel Time Estimation
are included in longer estimated window lengths. The maximum
This simulation is used to study how the length of the estimation win- cross-correlation method, which relies on only a single maximum
dow affects the performance of each estimation method. Five estimated cross-correlation coefficient, cannot consider traveler variability
window lengths (b2 − b1) are used in the simulation: b2 − b1 = 90, 150, and is more sensitive to noise influence in determining the maximum
300, 450, and 600 s. The other simulation parameters are kept the same cross-correlation coefficient. The proposed method, which uses all
as in the section on the effect of window length estimation on differ- significant cross-correlation coefficients for travel time estimation, can
ent travel time estimation methods. The three methods of travel time consider traveler variability with different travel times. Moreover,
estimation are compared in Figures 5a and 5b, which correspond to the use of the multiple significant cross-correlation coefficients can
the average of the estimation errors and the standard deviation of the smooth the estimation error of the cross-correlation coefficients. That
estimation errors, respectively. is why the proposed estimation method is better than the maximum
The plots in Figure 5 clearly indicate that for all estimation cross-correlation method with longer estimated window lengths.
methods, the standard deviations of the estimation errors are sig-
nificantly reduced when the window length estimation is ≥300 s. Case Study
However, if too large an estimation window is selected, the dynamic
behavior of the change in traffic condition will be lost. So, selec- A case study illustrates how the proposed method estimates travel
tion of an appropriate window length estimation is important. Fig- time under real traffic conditions. The flow data used in this study

TABLE 1 Comparison of Estimation Errors

Optimization Single Cross- Proposed Multi-Cross-

Value Method Correlation Method Correlation Method

Mean of error 0.2464 0.4547 0.2427

Standard deviation of error 1.238 1.1759 0.4814
Guo and Jin 17

0.8

Mean Value of Estimation Error

0.7

0.6

0.5 Optimization method

0.4

0.3

0.2

0.1 Proposed method Single cross-correlation method

0
100 150 200 250 300 350 400 450 500 550 600
Length of Fitting Window
(a)
Standard Deviation of the Estimation Error

3.5
Optimization method
3

2.5 Single cross-correlation method

2

1.5 Proposed method

1

0.5

0
100 150 200 250 300 350 400 450 500 550 600
Length of Fitting Window
(b)

FIGURE 5 Comparison of estimation errors under different fitting window lengths.

were collected at the intersection of Speedway Blvd. and Tucson by Detector 2 and used to estimate travel time; the link distance is
Blvd. in Tucson, Ariz., from 11:30 a.m. to 12:30 p.m. on December 2,689 ft (817 m), and the speed limit is 35 mph (15.6 m/s). Travel
13, 2002, by single loop detectors used to measure traffic flow in time estimation was updated every 10 min.
each direction. Figure 7 shows the cross-correlation analysis results in each esti-
The detector locations for the case study are illustrated in Figure 6. mation window, and travel time estimates are plotted in Figure 8. A
Speedway Blvd. has three lanes of traffic. Traffic data for the cen- reasonable trend of increased estimated travel time is observed
ter lane from Tucson Blvd. to Country Club Blvd. were collected approaching the noontime lunch break.
Country Club

Compell
Tucson

1 1 detector

speed limit 35miles/hr

2 Speedway 2
(15.6m/sec)

3 3

4 4

5 5

6 6

2689 feet (817 m)

FIGURE 6 Single loop detector locations at Speedway Blvd. in Tucson, Ariz.

18 Transportation Research Record 1968

0.4 0.4
1.96 standard error
0.2 0.2

0 0

-0.2 -0.2

0 20 40 60 80 0 20 40 60 80
11:30-11:40 AM 11:40-11:50 AM
(a) (b)
Cross-Correlation Coefficients

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

0 20 40 60 80 0 20 40 60 80
11:50-Noon Noon-12:10 PM
(c) (d)

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

0 20 40 60 80 0 20 40 60 80
12:10-12:20 PM 12:20-12:30 PM
(e) Lag k (f)

FIGURE 7 Cross-correlation analyses at different time periods.

55

50

45
Average Travel Time (seconds)

40

35

30

25

20

15
11:30-11:40AM 11:40-11:50AM 11:50-Noon 12:00-12:10PM 12:10-12:20PM 12:20-12:30PM
Time

FIGURE 8 Estimated average travel time using the proposed method.

Guo and Jin 19

CONCLUSION 3. Jacobson, L. N., N. L. Nihan, and J. D. Bender. Detecting Erroneous

Loop Detector Data in a Freeway Traffic Management System. In Trans-
This paper proposed a new correlation analysis method for travel time portation Research Record 1287, TRB, National Research Council,
Washington, D.C., 1990, pp. 151–166.
estimation. With this method, only travel flow information measured 4. Coifman, B. Estimating Travel Times and Vehicle Trajectories on Free-
by single loop detectors is required. Therefore, estimation errors ways Using Dual Loop Detectors. Transportation Research, Part A,
caused in the measurement of occupancy and vehicle length are Vol. 36, No. 4, 2002, pp. 351–364.
avoided. By using the t-test, the proposed method can automatically 5. Coifman, B., and M. Cassidy. Vehicle Reidentification and Travel Time
Measurement on Congested Freeway. Transportation Research, Part A,
determine the center and width of the estimation window of the
Vol. 36, No. 10, 2002, pp. 899–917.
probability distribution function of average travel time. 6. Nam, D. H., and D. R. Drew. Traffic Dynamics: Method for Estimating
Several comparison analyses between the proposed method and Freeway Travel Times in Real Time from Flow Measurements. Journal
two existing methods were conducted as simulation studies, and a of Transportation Engineering, Vol. 122, No. 3, 1996, pp. 185–191.
case study shows the effectiveness of the proposed method for the 7. Nam, D. H., and D. R. Drew. Automatic Measurement of Traffic Variables
for Intelligent Transportation System Applications. Transportation
data analysis of real traffic. Research, Part B, Vol. 33, No. 6, 1999, pp. 437–457.
8. Park, D., L. R. Rilett, and G. Han. Spectral Basis Neural Networks for
Real-Time Travel Time Forecasting. Journal of Transportation Engi-
ACKNOWLEDGMENTS neering, Vol. 125, No. 6, 1999, pp. 515–523.
9. Hoogendoorn, S. P., J. W. C. van Lint, and H. J. van Zuylen. Freeway
The authors thank the reviewers for insightful comments and sug- Travel Time Prediction with State-Space Neural Networks: Modeling
gestions that significantly improved the quality and readability of the State-Space Dynamics with Recurrent Neural Networks. In Transporta-
paper. The authors gratefully acknowledge financial support from tion Research Record: Journal of the Transportation Research Board,
No. 1811, Transportation Research Board of the National Academies,
the U.S. Department of Transportation and Bureau of Transporta- Washington, D.C., 2002, pp. 30–39.
tion Statistics; the Federal Highway Administration and Arizona 10. Dailey, D. J. Travel Time Estimation Using Cross-Correlation Techniques.
Department of Transportation; and the National Science Foundation. Transportation Research, Part B, Vol. 27, No. 2, 1993, pp. 97–107.
11. Dailey, D. J. A Statistical Algorithm for Estimating Speed from Single
Loop Volume and Occupancy Measurement. Transportation Research,
REFERENCES Part B, Vol. 33, No. 5, 1999, pp. 313–322.
12. Petty, K., P. Bickel, M. Ostland, J. Rice, F. Schoenberg, J. Jiang, and
Y. Ritov. Accurate Estimation of Travel Times from Single-Loop Detec-
1. Athol, P. Interdependence of Certain Operational Characteristics Within
tors. Transportation Research, Part A, Vol. 32, No. 1, 1998, pp. 1–17.
a Moving Traffic Stream. In Highway Research Record 72, HRB, National
13. Shumway, R. H., and D. S. Stoffer. Time Series Analysis and Its Appli-
Research Council, Washington, D.C., 1965, pp. 58–87.
cations. Springer-Verlag, New York, 1999.
2. Hall, F. L., and B. N. Persaud. Evaluation of Speed Estimates Made with
Single-Detector Data from Freeway Traffic Management Systems. In
Transportation Research Record 1232, TRB, National Research Council, The Artificial Intelligence and Advanced Computing Applications Committee spon-
Washington, D.C., 1989, pp. 9–16. sored publication of this paper.