Content uploaded by Edward Chung
Author content
All content in this area was uploaded by Edward Chung on Feb 19, 2020
Content may be subject to copyright.
This may be the author’s version of a work that was submitted/accepted
for publication in the following source:
Bhaskar, Ashish,Chung, Edward, & Dumont, AndreGilles
(2011)
Fusing loop detector and probe vehicle data to estimate travel time statis
tics on signalized urban networks.
ComputerAided Civil and Infrastructure Engineering,26(6), pp. 433450.
This ﬁle was downloaded from: https://eprints.qut.edu.au/53147/
c
Consult author(s) regarding copyright matters
This work is covered by copyright. Unless the document is being made available under a
Creative Commons Licence, you must assume that reuse is limited to personal use and
that permission from the copyright owner must be obtained for all other uses. If the docu
ment is available under a Creative Commons License (or other speciﬁed license) then refer
to the Licence for details of permitted reuse. It is a condition of access that users recog
nise and abide by the legal requirements associated with these rights. If you believe that
this work infringes copyright please provide details by email to qut.copyright@qut.edu.au
Notice:Please note that this document may not be the Version of Record
(i.e. published version) of the work. Author manuscript versions (as Sub
mitted for peer review or as Accepted for publication after peer review) can
be identiﬁed by an absence of publisher branding and/or typeset appear
ance. If there is any doubt, please refer to the published source.
https://doi.org/10.1111/j.14678667.2010.00697.x
Page 1 of 38
Fusing loop detector and probe vehicle
data to estimate travel time statistics on
signalized urban networks
Ashish Bhaskar*
Laboratory of Traffic Facility
Swiss Federal Institute of Technology
Lausanne, Switzerland
Ph: +41 21 693 2341;
Fax: +41 21 693 63 49
ashish.bhaskar@epfl.ch
Prof. Edward Chung
School of Urban Development
Queensland University of Technology
Brisbane, Australia
Ph: +61 731381143
Fax: +61 3138 1827
edward.chung@qut.edu.au
Prof. AndréGilles Dumont
Laboratory of Traffic Facility
Swiss Federal Institute of Technology
Lausanne, Switzerland
Ph: +41 21 693 2345
Fax: +41 21 693 63 49
andregilles.dumont@epfl.ch
First submission: 22nd June 2009
Final submission: 29th April 2010
*Corresponding author
Abstract
This paper presents a methodology that integrates cumulative plots with probe vehicle data for
estimation of travel time statistics (average, quartile) on urban networks. The integration reduces
relative deviation amongst the cumulative plots so that the classical analytical procedure of defining
the area between the plots as the total travel time can be applied. For quartile estimation, a slicing
technique is proposed. The methodology is validated with real data from Lucerne, Switzerland and it is
concluded that the travel time estimates from the proposed methodology are statistically equivalent to
the observed values.
Page 2 of 38
1. Introduction
Travel time is defined as the time needed to travel from a point upstream (u/s) to a point downstream
(d/s) on the network. It quantifies congestion and is easily perceived by all road users and operators. It
is an important network performance measure and a decision making variable. For instance, travel
time information is utilized by the operators to develop traffic control strategies for reducing congestion
on both spatial and temporal scales.
Literature is abundant with models for obtaining travel time values. Majority of the literature is limited to
freeways (Nam and Drew, 1996; Dharia and Adeli, 2003; Jintanakul et al., 2009) and cannot be
applied on urban networks. The development of a travel time model on urban networks is more
challenging than freeways due to number of reasons. For instance, interruption in traffic flow due to
traffic signals; non conservation of traffic flow on urban link due to midlink sources and sinks (e.g.
parking, side street) etc. Travel time models can be differentiated into estimation models and
prediction models. Estimation models provide experienced travel time whereas, prediction models
(Park et al., 1999; You and Kim, 2000; Zhang and Rice, 2003; Vlahogianni et al., 2004; Vlahogianni et
al., 2008; Hamad et al., 2009) provide expected (forecasted) travel time in future. The future can be
immediate (i.e., for trips just starting) to several minutes (say 30 minutes) ahead.
Traffic flow on an urban network is in stopandgo running condition i.e., vehicles have to the stop at
the intersection during signal red phase and queue of vehicles is formed. The individual vehicle travel
time on urban link depends on number of factors. For instance, traffic demand; signal parameters;
vehicles’ entry time on the link relative to downstream intersection signal red phase; and number of
vehicles queued in front of it when it reaches the downstream signal etc. The distribution of travel time
from different vehicles departing within an estimation interval (Here, referred as within interval
distribution.) is bimodal (or even multimodal) with modes corresponding to the vehicles that can pass
the link without stopping at intersection and for vehicles that experience delay at intersection.
Average travel time is an important indicator for network performance measure and generally most of
the inpractice models are applicable for average travel time estimation (Review provided in Section 2).
Due to the above mentioned bimodal distribution; none of the vehicle can encounter average travel
time. Hence, for better understanding of the network performance it is important to estimate other
statistics, such as upper quartile of travel time in addition to the average travel time. Little research is
performed on the estimation of within interval distribution mainly due to unavailability of individual
vehicle travel time data. Robinson (2005) has applied kNN approach, on the Automatic Vehicle
Identification (AVI) data from central London, to estimate travel time variance for within interval
distribution. He has observed around 23% variance in travel time between vehicles traversing the link
during 15 minutes estimation interval. This is more than three times of the variance observed by Part
et al., (1999) on Houston freeways.
The objective of this paper is to develop and validate a methodology for estimation of travel time
statistics (average and quartiles) on signalized urban networks. The estimates are for certain time
period that is integral multiple of signal control cycle. For instance, average travel time during five
Page 3 of 38
signal control cycles. The develop methodology should be robust with respect to urban complexities,
such as midlink sources and sinks and detector counting error. The proposed methodology is named,
CUmulative plots and PRobe Integration for Travel timE estimation (CUPRITE) (Bhaskar, 2009). The
methodology is based on classical analytical procedure for travel time estimation using cumulative
plots. Analytical modeling is performed through integrating or fusing cumulative plots with probe
vehicle data for accurate estimation of travel time statistics (average travel time and quartile of travel
time). The proposed methodology can be applied for real time application, where the estimated travel
time is the experienced travel in the last estimation interval.
The paper is organized as follows: section 2 provides the literature review for the average travel time
estimation on urban networks. The classical analytical procedure for travel time estimation and its
vulnerability for application on urban environment are introduced in section 3. Thereafter, the
proposed methodology is developed in section 4, followed by its validation with real data in section 5.
Finally, the conclusions are presented in section 6.
2. Literature review
Advancement of technology has resulted in different traffic data retrieval systems from traditional loop
detectors to advanced electronic systems onboard a vehicle, such as Vehicle Information and
Communications Systems (VICS). Fixed sensors, such as loop detectors provide traffic flow and
occupancy at the specific location on the network whereas; mobile sensors, such as probe vehicle
provide data for the entire journey of the vehicle. Based on the type of data available, different models
are proposed to estimate average travel time for all the vehicles traversing the road. Moreover, the
availability of different data systems provide avenue for application of data fusion techniques for more
reliable and robust travel time estimation. Thereafter, here we classify the literature into: i) fixed
sensor; ii) mobile sensor; and iii) data fusion based models for average travel time estimation.
Fixed sensor based:
The initial motivation for the development of travel time estimation models was to consider effect of
congestion in conventional traffic assignment step used in fourstep transportation modeling. Several
travel time functions (or volume delay functions) are proposed that define relationship between link
travel time and traffic intensity (flow/capacity ratio). These include: Bureau of Public Roads (BPR,
1964); Davidson’s function (Akçelik, 1978; Tisato, 1991); conicalvolume delay function (Spiess, 1990)
etc. Webster delay model (Webster and Cobbe, 1966) is a pioneer model for estimating average
deterministic delay at undersaturated signalized intersection. Researchers have followed Webster’s
work to suit different field conditions and modified models are proposed, such as Akçelik (Akçelik,
1988; Akçelik, 1991; Akcelik and Rouphail, 1993; Akcelik and Rouphail, 1994) and Highway Capacity
Manual’s procedure for delay estimation (TRB, 1998; TRB, 2000).
The simplicity of these travel time function make them favorable candidate for transport planning and
policy analysis. They are not suited for ITS applications where more accurate and reliable analysis
especially for variable traffic conditions in real time is required.
Page 4 of 38
Regression analysis based models to estimate link travel time, as a function of site characteristics and
detector data are also proposed. Wardrop (1968) has defined regression relationship between
average journey speed in central urban area as a function of average traffic flow, width of the road, the
number of controlled intersections per miles and average proportion of green time. Researchers (Gault,
1981; Young, 1988; Sisiopiku and Rouphail, 1994; Sisiopiku et al., 1994) have observed a regression
relationship between average travel time and certain ranges of detector occupancy, for midlink
detectors, with queue that does not persist over detector location. Zhang (1999) has proposed a
regression equation for journey speed as a function of volume to capacity ratio and midlink detector
flow and occupancy.
Generally, the regression models are site specific and require calibration to suite different environment.
These models should not be generalized and the effect of parameters, such as detector location,
effective green time, progression quality, link length, opposing flow from permissive phasing, traffic
composition etc. should not be overlooked.
Researchers have also applied machine learning algorithm, such as kNearest Neighbor (Robinson
and Polak, 2005) and Artificial Neural Networks (Palacharla and Nelson, 1999; Liu et al., 2005), for
travel time estimation and prediction (Hamed et al., 1995). Such models require measured link travel
time values and the corresponding detector data for a training period. The training data set should
properly represent the required extend of the solution space and the model should be applied well
within the limits for which it is trained.
Skabardonis and Geroliminis (2005) have proposed a model for travel time estimation based on
upstream detector data and signal parameters. The queueing at the signal is considered by applying
kinematic wave theory. The required detector should be sufficient upstream from the intersection
stopline, so that the flow and occupancy measures from the detector are not affected by the presence
of queue at the signals. The model involves calibrating the fundamental diagram (flowdensity
relationship) for links using detector data.
Recent advancement in sensor technology has produced Advance Inductive Loop Detectors (AILD)
that can provide magnetic vehicle signatures. The signatures from upstream and downstream
detectors can be correlated to reidentify the vehicles (or platoon) at downstream location for travel
time estimation. Ritchie et al., (2002; 2005) have demonstrated the potential application of vehicle
signature for travel time estimation on urban arterial. However, this approach is still in initial research
stage, and further study is needed to increase the accuracy, reliability and reidentification rate.
Moreover, for the implementation of reidentification technique, existing infrastructure should be
upgraded with AILD and a high bandwidth in the data communication channel.
Mobile sensor based:
Mobile sensors, such as probe vehicle is a vehicle equipped with vehicle tracking equipment (e.g.,
GPS) and can provide data for the vehicles’ trajectory (time stamp and position coordinates) and
hence its travel time. In practice, only a small fraction of all the vehicles traversing the link are probe
vehicles. Average travel time for all the vehicles traversing the link can be estimated by applying
Page 5 of 38
statistical sampling techniques on the travel time obtained from the probe vehicles (Hellinga and Fu,
2002; Long Cheu et al., 2002). Researchers (Srinivasan and Jovanis, 1996; Long Cheu, Xie and Lee,
2002) have shown interest to determine minimum number of probes required for statistically significant
travel time estimation.
Data fusion based:
Researchers have also applied data fusion techniques to fuse data from different sources, specifically
detector and probe vehicles data for travel time estimation. El Faouzi (2004) provides an overview of
the application of data fusion techniques in road traffic engineering. Dailey et al., (1996) summaries
ITS data fusion projects.
Berka et al. (1995) has applied weighted average based fusion technique where the fused average
travel time is the weighted average of the estimated average travel time from detectors and estimated
average travel time from probe vehicles. The weights are defined by considering variables, such as
the standard deviation of the travel time from detector data and from probe data, respectively; weights
assigned to detector travel time in data screening; the sum of weights of reasonable probe reports etc.
A similar weighted average based data fusion approach for travel time estimation is proposed by El
Faouzi (2004).
Choi and Chung (2002) have applied the data fusion technique for 5 minutes average travel time
estimates using detector and probe vehicle data. The algorithm first estimates spacemean speed
from detector counts and occupancy using Dailey (1999) equation, which provides travel time
estimates for each minute. Each minute travel time estimates are aggregated using Voting Technique
for 5 minutes average travel time (TTd). Average 5 minutes travel time (TTg) from GPS probes is
obtained using Fuzzy regression. Finally, fused link travel time is obtained by applying Bayesian
Pooling Method on TTd and TTg. The algorithm is tested for undersaturated traffic condition and should
be tested for oversaturated traffic condition too. They quote that “a different level of service might
produce totally different weights of each data collection mechanism. In such cases, a different data
fusion method and/or a revision of the proposed algorithm may be needed”.
Xie et al., (2004) have applied two independent neural network methodologies: MultiLayer Perception
(MLP) and MultiLayer regression (MLR) models to fuse average travel time estimates from detector
data and probe vehicles. The average travel time from detector is the sum of the free flow travel time
and signal delay. The signal delay is estimated using Singapore model (Xie et al., 2001). Average
travel time from probe samples are considered only if the sample size during estimation interval is
more than 10 vehicles or is more than the minimum required sample size determined by central limit
theorem.
Page 6 of 38
3. The classical analytical procedure for travel time estimation
3.1 The procedure
Cumulative plot is a plot of cumulative count of vehicles versus time at a specific location on the
network. The classical analytical procedure for travel time estimation considers cumulative plots U(t)
and D(t) at upstream (u/s) and downstream (d/s) locations, respectively (Daganzo, 1997). It defines
the total travel time from u/s to d/s as the area (Refer to Figure 1a) between the plots. Say, time t1 and
time t2 correspond to the start and end of the U(t) represented in the area, respectively. Similarly t3 and
t4 are time corresponding to the start and end of D(t) represented in the area, respectively. Then,
mathematically, the average travel time
TT
is represented as follows:
1 1 1 1
1 1 1
( ) ( ) ( ) ( )
N N N
i i i
D i U i D i U i
TT NN
(1)
2 1 4 3
( ) ( ) ( ) ( )N U t U t D t D t
(2)
Here N is the number of vehicles that depart downstream (arrives upstream) during the time interval
from t3 to t4 (t1 to t2).
3.2 The Relative Deviation (RD) issue with the procedure
The area between the plots is the total travel time from upstream to downstream as long as all the
vehicles represented in U(t), from time t1 to t2, and in D(t), from time t3 to t4, are same. Therefore, the
plots should be based on only those vehicles that traverse from upstream to downstream.
Cumulative plots are defined based on the detector counts at a specific location. Practically, detectors
are not perfect and one can easily observe 5% error in detector counting. Moreover, urban network
has midlink sources and sinks, such as parking or sidestreet. This results in non conservation of
vehicles (loss or gain of vehicles) between upstream and downstream locations. Due to detector
counting error; non conservation of vehicles between plots location; and any such combinations over
time, there is relative deviation (RD) amongst the plots (also termed as “drift”).
Let us explain RD with the help of an example. Consider a scenario where upstream detector is
overcounting. In Figure 1b: U(t) is the cumulative plot observed from the overcounting upstream
detector; U’(t) is from a perfect detector. U(t) deviates from U’(t), or there is a relative deviation
between U(t) and D(t). The observed cumulative plots are U(t) and D(t) and if the classical procedure
is applied then the error in the estimation of travel time, during TEI travel time estimation interval, is
represented as the shaded region in the figure. If RD is left unchecked then the error can exponentially
grow with time. Hence, the RD issue is critical in the application of the classical procedure.
Note: U(t) and D(t) will eventually “diverge” from each other if: upstream detector is overcounting; or
downstream detector is undercounting; or there is midlink sink. U(t) and D(t) will eventually “cut” each
Page 7 of 38
other if: upstream detector is undercounting; or downstream detector is overcounting; or there is mid
link source. If the plots “diverge” then the travel time is highly overestimated and if the plots “cut” then
travel time estimates are negative. In practice, there is complex combination of detector errors, mid
link sources and midlink sinks over time, which defines the relative deviation for each estimation
interval.
Figure 1: Classical analytical procedure and its vulnerability to relative deviation (RD) amongst
the plots.
Traffic flow direction
Area (A)
N
Average
Travel Time
= A/N
D(
t
)
U(
t
)
Time
Cumulative counts
u/s
d/s
Road
t3
t4
t1
t2
D(
t
)
Time
Cumulative counts
U’(
t
): from perfect detector
U(
t
): from
overcounting
detector
U(
t
) deviated from U’(
t
)
Relative deviation
between U(
t
) and D(
t
)
(a)
(b)
Example: Upstream detector is overcounting
TEI
Error in travel
time estimation
i
t
tti Horizontal distance
(temporal separation)
between the curves for
rank i.
n Vertical distance
(counts separation)
between the curves at
time t.
Page 8 of 38
3.3 How RD issue can be addressed?
Refer to Figure 1a: The vertical distance (counts separation) between the plots (at time t) is the
number of vehicles (n) between upstream and downstream locations. The horizontal distance
(temporal separation) between the plots (for rank i) is an estimate of travel time, tti, for the ith vehicle
under FIFO assumption. Therefore, the knowledge about the number of vehicles between upstream
and downstream locations; or the travel time of individual vehicle can be applied to address the RD
issue.
The number of vehicles between upstream and downstream locations is difficult to obtain. Alternatively,
we can consider the knowledge about the queue length. Theoretically, the queue length can be
defined as follows:
( ) ( )
ff
Queueattimet U t t D t
(3)
Here, tff is the free flow travel time of the link. This is further discussed in section 4.4.
There is an increasing use of probe vehicle, which can provide its travel time. Hence, in this paper we
propose a methodology that integrates probe vehicle with cumulative plots to resolve the RD issue;
and applies slicing technique for estimation of travel time quartiles.
4. The proposed methodology
4.1 Probe vehicle data
Here, probe vehicle is a vehicle equipped with vehicle tracking equipment. There are issues related to
the probe vehicle data, such as map matching, frequency of probe data etc. To address these issues
is beyond the scope of this paper. We assume that the time when the probe vehicle is at upstream (tu)
and downstream (td) locations is accurately obtained and its travel time is td – tu.
4.2 Architecture of CUPRITE
The proposed methodology integrates cumulative plots with probes. The basic concept for the
integration is introduced in Section 4.3. The architecture (see Figure 2) of the proposed algorithm is as
follows (the details for which are provided in the following sections):
Step 1 Cumulative plots U(t) and D(t) are defined. Here, if the detector data is individual
vehicle data (pulse data), then the cumulative counts can be obtained by
cumulative the vehicles. However, if detector data is not a pulse data but an
aggregated traffic counts during certain detection interval (for instance counts per
60 seconds), then cumulating the counts for each detection interval will not reflect
the actual traffic fluctuations within the detection interval. These fluctuations can
be captured by integrating the detector counts with signal timings, where the
counts during the signal red phase is assigned to be zero, and counts during the
signal green phase is segregated into counts from the saturation flow and counts
Page 9 of 38
from non saturation flow. Refer to Bhaskar et al., (2008) for the methodology for
integration of signal timings with aggregated traffic counts from detector data for
accurate representation of cumulative plots.
Step 2 Probe vehicle data, list of [tu] and [td], is defined (Refer to Sections 4.1 and 4.3).
Moreover, if the conditions for virtual probe (Section 4.4.1) are satisfied then the
list [tu] and [td] is appended with additional elements corresponding to the virtual
probe i.e., tu= tGEtff; td = tGE (where tGE and tff are time corresponding to the end of
signal green phase and link free flow travel time, respectively). Else, only real
probes are considered.
Step 3 Points through which U(t) should pass are defined (Section 4.5).
Step 4 U(t) is redefined by first vertical scaling and shifting the plots (Section 4.6) so that
it passes through the above defined points (Step 3); and
Step 5 Finally, for each estimation interval: a) average travel time (Section 4.7) is
estimated using classical analytical procedure; and b) quartile for travel time
(Section 4.8) is estimated using slicing technique.
Page 10 of 38
Figure 2: CUPRITE architecture for estimation of travel time statistics.
4.3 Integration of cumulative plots and probes
Suppose, there is no RD and both U(t) and D(t) are perfect. Due to nonFIFO traffic behavior, even in
the absence of RD, the rank of a vehicle in upstream and downstream cumulative plots may not be
same i.e, U(tu) ≠ D(td) (Figure 3a). We fix the vehicle to downstream cumulative plot, i.e., we fix the
rank of the vehicle in the cumulative plots as D(td) (Refer to Figure 3b) and define a parameter Δt :
1( ( ))
du
t U D t t
(4)
Where U1(D(td)) is the time when the vehicle is represented at U(t) given that we fix its rank to D(td).
Signal data
U(t) D(t)
Probe
vehicle
Probe data
Lists of [tu] & [td]
Redefine upstream
cumulative plot
Refefined
U(t)
Classical analytical procedure
No
Points through
which U(t)
should pass
Define points through which
U(t) should pass
Virtual probe
conditions
satisfied*
Yes
*If the virtual probe
conditions are not satisfied
then only real probes are
considered
Slicing technique
Quartile of
travel time
Average
travel time
Detector
data
Is pulse
data?
Yes
Integrate signal timings
with detector data
Page 11 of 38
Figure 3: a) Illustration for the relationship between probe data and cumulative plots; b) Fixing
of probe data to D(t).
If all the vehicles in U(t) and D(t) are same then
t
from all the vehicles should be zero (5). This is
an important property and is the explanation for the area between the plots to be the total travel time.
0
i
i
t
(5)
If there is presence of RD then the equation (5) is not satisfied. Therefore, RD issue can be addressed
by correcting the cumulative plots such that equation (5) is satisfied.
Equation (5) is satisfied when we are considering all the vehicles. Each vehicle has an equal
probability of being a probe vehicle and only a fraction of vehicles are randomly selected as probes.
We make a hypothesis that we can remove or at least reduce the RD by redefining U(t) such that
t
from all the probes is zero.
D(
t
)
U(
t
)
Time
Cumulative counts
td
tu
(a)
U(tu)
D(td)
D(
t
)
U(
t
)
Time
Cumulative counts
td
tu
(a)
U(tu)
D(td)
U1(D(td))
Δt
Probe Data
tu
td
Probe fixed
to D(t)
U(tu) may not be
equal to D(td)
Page 12 of 38
In practice, we do not know which plot is responsible for RD issue. It can be U(t), D(t) or both. It is
complicated to correct both U(t) and D(t) simultaneously. As the deviation amongst the plots is relative
therefore, we can correct either U(t) or D(t). Here, we redefine U(t) because we fix the rank of the
probe considering D(t). Alternatively, we can redefine D(t), if we fix the rank of the probe considering
U(t).
4.4 Virtual probe
Virtual probe (Figure 4) is defined as a virtual vehicle that, during undersaturated traffic flow, departs
from the downstream at the end of signal green phase (at time tGE) and its travel time is freeflow travel
time (tff) of the link. The probe is not real and is defined with the aim to reduce RD.
For undersaturated traffic condition, the vehicle queue formed during the signal red phase should be
completely served during the signal green phase i.e., the queue length at time tGE should be zero.
Considering equation (3): U(tGE  tff) = D(tGE) i.e., the travel time of the vehicle that enters the
intersection at time tGE should be close to tff. Therefore, under such conditions we can define virtual
probe (see Figure 4) such that it is observed at upstream and downstream locations at time tGE  tff and
time tGE, respectively (i.e. for virtual probe tu = tGE  tff and td = tGE.).
Note: Virtual probe is only defined if the following conditions for virtual probe are satisfied.
4.4.1 Conditions for virtual probe
i. As the travel time of a virtual probe is defined as freeflow travel time of the link,
therefore on the link the sources for significant midlink delay, such as midlink
intersections and onstreet bus stop should be absent.
ii. Virtual probes are defined only for undersaturated condition with logic of zero queue
length at the end of signal green phase. Traffic condition is defined as undersaturated if
counts during the signal cycle (or more specifically during signal green time) are less
than the corresponding capacity (Figure 4) i.e.,
( ) ( ) *
GE GE
D t D t c s g
(6)
Where: s and g are saturation flow rate and effective signal green time, respectively;
s*g is the capacity and ∆ is a calibration parameter to take into account the error in the
estimation of capacity.
To define equation (6) it is assumed that there is no spillover from downstream link. If
there is spillover, then vehicles are restricted to flow resulting in low counts at stopline
detector. Capacity is generally not corrected to account for the spillover from
downstream link. Due to which equation (6) is satisfied and system can falsely indicate
undersaturated situation for spillover cases. Though under such situation the queue
may not vanish and hence virtual probe should not be defined.
Page 13 of 38
iii. Virtual probe is defined with the aim to reduce RD. Hence, it should only be defined if
there is presence of RD i.e., the following equation should be satisfied:
1( ( )) [ , ]
GE GE ff ff
U D t t t t
(7)
Where: δ is a calibration parameter taking into account the variation in the estimation of
tff. It can be considered equal to the standard deviation of the estimate of tff.
Figure 4: Illustration of a virtual probe fixed to D(t).
4.5 How to define the points through which U(t) should pass?
Say, we have n probe vehicles and the database for the probe is defined as list of [tu] and list of [td],
where the size of each list is n. The value of jth element in the list represents the data from the jth
probe. These lists are appended with additional elements satisfying the conditions for virtual probe
(Section 4.4.1). If the conditions are satisfied, then time tGE is appended to the list [td]; and time
(tGE  tff) is appended to the list [tu].
4.5.1 Grid technique
Consider an example, in Figure 5a, where we have four probes fixed to D(t). The U(t) should pass
within the region satisfying the following constrain (Refer to the rectangular region in Figure 5a)
min[ ] max[ ]
uu
probes probes
t t t
(8)
min[ ( )] max[ ( )]
dd
probes probes
D t counts D t
(9)
Saturation flow rate (s)
tGE
Green (g)
Red (r)
Cycle (c)
(tRS ,D(tRS))
(
tGE
, D(
tGE
))
(
tGE

tff
, D(
tGE
))
Virtual probe
(tff)
D(t)
Time
Cumulative counts
D(tGE)D(tGE  c)
s*g
D(tGE)D(tGE  c) < s*g
Queue is served
tGE  c
Page 14 of 38
We can define a grid with rows corresponding to D(td) and columns corresponding to tu within the
above region (Refer to Figure 5b). If U(t) passes through the diagonal nodes of the above grid then
∑∆t = 0 is satisfied. Therefore, the required points to pass are the diagonal nodes of the grid and can
be obtained from the following algorithm:
Step 1 Sort list [td] in ascending order of its values. This is required as the rank of the
probe is defined considering D(t).
Step 2 Sort list [tu] in ascending order of its values. This is required to make sure that the
redefined U(t) is monotonically increasing and satisfies the property of ∑∆t = 0.
Step 3 The required points through which U(t) should pass are (tuj, D(tdj)); where tuj and tdj
are jth value in the sorted list of [tu] and [td], respectively.
Figure 5: Points through which the U(t) should pass.
Time
Cumulative counts
Time
tu2 tu1 tu4 tu3
D(t)
td1 td2 td3 td4
tu2 tu1 tu4 tu3 td1 td2 td3 td4
D(t)
Point through which U(t) should pass
∆t1
∑∆ti=0
∆t2
∆t3
∆t4
∆t1=∆t2; ∆t3=∆t4
Region
(b)
Cumulative counts
(a)
Grid
Page 15 of 38
4.6 How to redefine U(t)?
4.6.1 Reference point
We define reference point as the point in which we have confidence that it is a correct point on the
cumulative plot. Initially, U(t) and D(t) are two independent cumulative plots. When the traffic condition
is freeflow (for instance during night) then counts for cumulative plots can be initialized to zero. This is
the initial reference point (P0). Say [P1, P2, P3, …, Pn] is the list of n points from where U(t) should
pass, then for redefining U(t) for point Pi, the reference point is Pi1 and so on.
4.6.2 Vertical scaling and shifting technique
The RD issue is the result of: a) the error in the cumulative counts due to error in detector counting;
and b) inconsistency between the cumulative count at upstream and downstream locations  due to
midlink sources and sinks. Therefore, to address the issue, the cumulative counts (vertical axis)
should be corrected. For this, we apply the following vertical scaling and shifting technique.
Say, a) point (tRef, U(tRef)) is a reference point; and b) point (tp, Yp) is a point through which U(t) should
pass (Section 4.5). The redefined U(t) should pass through both these points.
Refer to Figure 6, we define
p
t
Y
= U(tp)  U(tRef) and
p
t
y
= Yp  U(tRef).
Figure 6: Illustration of the abbreviations for vertical scaling.
a) For time ≤ tRef
Time
Reference Point
(tRef, U(tRef))
Point to pass
(tp, Yp)
tp
tRef
Cumulative counts
(tp, U(tp))
t
p
t
y
p
t
Y
p
t
Ref
Ref
Ref
()
( ) ( )
()
( ) ( )
p p p
p
p
t t t p p
tp
tp
t
Y y U t Y
Y U t U t
y Y U t
Y U t U t
t
Y
U(t)
Page 16 of 38
The reference point is the correct point on the cumulative plot; therefore no correction on the
cumulative plot is required for time less than and equal to tRef.
b) For tRef < time ≤ tp
We perform vertical scaling on U(t) such that it passes through the point (tp, Yp). The scale is defined
as follows:
Ref Ref
Ref
Ref
() ( ) ( )
( ) ( )
1 ( ) ( )
p
p
tp
p
tp
p
yY U t if U t U t
Y U t U t
scale
if U t U t
(10)
The value of the scale reflects the net effect of the RD on the cumulative plots:
i. scale >1: The plots are diverging. For diverging plots, the error can exponentially grow
with time. For instance, there is a midlink sink. The vehicles from the sink are
observed at upstream and not at downstream.
ii. scale <1: The plots are converging. For converging plots, U(t) can cut D(t) resulting in
negative travel time estimation. For instance, there is a midlink source. The vehicles
from the source are observed at downstream but not at upstream.
iii. scale =1: RD is absent.
Here, the relative deviation in the cumulative count at time t, (
t
) is the result of the accumulation of
the relative deviation since time tRef (Refer to Figure 6). Equation (11) defines the RD (
p
t
) at time tp.
()
p p p
t t t p p
Y y U t Y
(11)
The proportion of this relative deviation to the cumulative counts (
p
p
t
t
Y
) is assumed to be constant(12).
;
p
p
ttRef p
tt
t t t
YY
(12)
Where:
t
is the relative deviation at time t; Yt = U(t)  U(tRef).
The above equations can be rearranged to define
t
as follows:
(1 )* ;
t t Ref p
scale Y t t t
(13)
c) For time > tp
All the points on U(t) beyond time tp are shifted vertically so that the redefined U(t) is continuous. The
magnitude of the shift is the relative deviation observed at time tp (Eq (11)).
Page 17 of 38
The above is summarized as follows: we redefine U(t) (14) by applying correction (15) on it such that
all points on the plot: i) before time tRef have no correction; ii) between tRef to tp are scaled vertically;
and iii) beyond tp are shifted vertically.
( ) ( ) t
U t U t
(14)
Ref
0
(1 )*( ( ) ( ))
()
Ref
t Ref p
p p p
tt
scale U t U t t t t
U t Y t t
(15)
Figure 7 represents an example, where P0 is the initial reference point; and points P1 and P2 are two
points through which U(t) should pass (refer to section 4.5). First, the correction for point P1 is
performed with P0 as the reference point (Refer to Figure 7b). Thereafter, P1 becomes the reference
point for P2 and correction for P2 is performed (Refer to Figure 7c). The redefined U(t) considering
points P1 and P2 is represented in Figure 7c.
Page 18 of 38
Figure 7: Example for redefining U(t) based on vertical scaling and shifting technique.
Time
Cumulative counts
Time
Cumulative counts
Time
Cumulative counts
P0
P1
P2
P0:Reference point
P1
P1:Reference point
P2
(a)
(b)
(c)
Vertical
Scaling
Vertical
Shifting
No
Correction
Vertical
Scaling
Vertical
Shifting
No
Correction
Correction for P1
with P0 as
reference
Correction for P2
with P1 as
reference
U(t)
Parallel
curves
Parallel
curves
Redefined U(t)
Page 19 of 38
4.7 Average travel time estimation
The classical procedure (see section 3) is applied between redefined U(t) and D(t) to estimate average
travel time.
4.8 Quartiles of travel time estimation
By definition, quartile is any value that divides the sorted data into equal parts:
i. Q1: the first quartile is the 25th percentile and 25% of the data is lower than Q1.
ii. Q2: the second quartile is the 50th percentile (or median) and it divides the data into two
equal parts.
iii. Q3: the third quartile is the 75th percentile and 75% of the data is lower than Q3.
To obtain quartiles of travel time, we need either individual vehicle travel time data or grouped vehicle
travel time data. The later is the data consisting of representative travel time from different group of
vehicles. For FIFO systems, the horizontal distance (temporal separation) between the cumulative
plots is an estimate for individual vehicle travel time. For both FIFO and non FIFO systems, the area
between the plots is an estimate for total travel time for a group of vehicles represented within the
plots. We propose the following slicing technique where, within an estimation interval, we slice the
area between the cumulative plots to obtain the grouped vehicle travel time data.
4.8.1 Slicing technique
Cumulative plot is a two dimensional piece wise linear graph with coordinates (ti, i) as its nodes; where
ti is the time when the ith vehicle is observed. In Figure 8, we illustrate a study link between u/s and d/s
locations. In the figure, links L1, L2 and L3 are three upstream links that contribute to the flow on the
study link. Each of these links has a signal phase Phase1, Phase2 and Phase3 that permit the flow
towards the study link, respectively. Time tgs1, tgs2 and tgs3 correspond to the start of signal green time
for Phase1, Phase2 and Phase3, respectively.
Upstream cumulative plot, U(t), is defined by the flow contributions from different upstream links (L1, L2
and L3). We define a “cut node” as the node corresponding to the start of each signal green time (tgs1,
tgs2 and tgs3) for the upstream signal phases that permits the flow towards the study link. The “cut
nodes” in the U(t) is marked in the figure. Here a node is a “cut node” if the following is satisfied:
1 2 3
1 1 1
(( ) ( ) ( ))
( , ) " "
i gs i i gs i i gs i
i
if t t t or t t t or t t t
thennode t i is a cut node
(16)
Similarly, for downstream cumulative plot, we define a “cut node” as the node corresponding to the
start of the signal green time for the downstream signal phases that permits the departure of the
vehicles from the study link.
Page 20 of 38
Figure 8: Simplified illustration of how a “cut node” is defined for upstream cumulative plot
with flow contribution from different upstream links.
We define, Mu as a matrix of the “cut nodes” for the upstream cumulative plot and similarly MD for
downstream cumulative plot. Within an estimation interval, the total area A between the cumulative
plots, is fragmented into different areas (Ai) (see Figure 9), by horizontal cuts corresponding to the
nodes at MU (“cut node” matrix for U(t)), MD (“cut node” matrix for D(t)) and with the following
constraint:
For each fragmented area Ai, if the counts Ni, are above a certain threshold number, Nthreshold, then the
area (Ai) is further fragmented by a horizontal cut into two fragmented areas: Ai1 and Ai2 with counts
Nthreshold and NiNthreshold, respectively. The threshold value provides an upper limit on the number of
vehicles for each fragmented area.
The process is repeated until each fragmented area satisfies the constraint. Finally, each fragmented
area (Ai) represents the total travel time for the Ni number of vehicles. We assume that Ni number of
u/s
L2
L3
L1
d/s
Study link
tgs1
tgs2
tgs3
Phase1
Phase2
Phase3
Time
u/s Signal phases† associated with the flow towards the study link
Time
Cumulative counts
tgs1
tgs2
tgs3
cut nodes
U(t)
† For simplicity of presentation, the green time illustrated here is effective green.
Page 21 of 38
vehicles experience similar travel time (
i
TT
) equal to the Ai/Ni. This defines a grouped vehicle travel
time data.
Figure 9: Illustration for slicing the area between cumulative plots for defining travel time for
different group of vehicles within an estimation interval.
Note: if we define Nthreshold =1 vehicle, then the travel time estimates from the above procedure is
equivalent to obtaining individual vehicle estimates as the horizontal distance between the plots.
An estimate for the quartiles of travel time is obtained by sorting the travel time values obtained from
all the sliced areas and corresponding number of vehicles as follows:
For an estimation interval, say we have a two dimensional array with first column as list of Ai/Ni (i.e.,
list LA/N) and second column as list of Ni (i.e., list LN). Following steps are followed:
Step 1 Sort the array with respect to the values in the list LA/N;
Step 2 Define a cumulative frequency list (Lf) by cumulating the values in the list LN;
Step 3 Define N, as total number of vehicles in the estimation interval. This is the last
element of the above cumulative frequency list;
Time
Cumulative counts
N
TEI
Time
Cumulative counts
Ni
A
MU
MD
Ai
For each fragmented area Ai
Ni
Ai
Ni Nthreshold
Ai1
if Ni > Nthreshold
Ai2
Nthreshold
Horizontal cut
New Horizontal cut
For Ni vehicles travel time is:
i
i
i
A
TT N
Page 22 of 38
Step 4 Define the index for the quartiles as follows:
Q1_index = 0.25*N
Q2_index = 0.5*N
Q3_index = 0.75*N
Step 5 Quartiles are defined as the value corresponding to the jth element of the sorted
list LA/N where j is the rank of Lf such that:
if(j=0) and ( Lf[j] ≥ Q3_index),
then Q3 = LA/N[0]
if(j>0) and ( Lf[j1] < Q3_index) and (Lf[j] ≥ Q3_index),
then Q3 = LA/N[j]
Similarly, for Q2 and Q3;
Note: here the elements of the list start from rank 0.
For better understanding of the above algorithm a self explaining example is presented in the Figure
10, where Table A is the original list of LA/N and LN ; Step 1 and Step 2 are executed in Table B; Step 3
and Step 4 are performed in Table C; and finally, the quartiles are defined by performing Step 5 in
Table B.
Page 23 of 38
Figure 10: An example for quartile estimation using slicing technique.
5. Validation of the methodology with real data
The methodology is validated using real data from Lucerne, Switzerland. The traffic at the site is
controlled by a fully actuated signal controller named VSPLUS (VSPLUS) that provides the detector
counts and signal timings. Ground truth individual vehicle travel time, is obtained from manual number
Table A
Table B
Original List
Step 1
Step 2
Sorted w.r.t Ai/Ni
[Cumulative Ni]
LA/N
[Ai/Ni]
LN
[Ni]
LA/N
[Ai/Ni]
LN
[Ni]
Lf
122.14
2.0
122.13
2.0
2.0
192.84
2.0
122.14
2.0
4.0
176.64
4.0
130.96
5.0
9.0
130.96
5.0
154.90
1.0
10.0
122.13
2.0
164.88
2.0
12.0
< Q1_index
198.54
5.0
Q1 →
166.08
5.0
17.0
≥ Q1_index
200.68
4.0
176.64
4.0
21.0
191.27
1.0
177.60
2.0
23.0
164.88
2.0
188.31
5.0
28.0
< Q2_index
234.54
1.0
Q2 →
191.27
1.0
29.0
≥ Q2_index
217.51
5.0
192.84
2.0
31.0
166.08
5.0
198.54
5.0
36.0
154.90
1.0
200.68
4.0
40.0
< Q3_index
228.88
3.0
Q3 →
213.28
4.0
44.0
≥ Q3_index
188.31
5.0
217.51
5.0
49.0
177.60
2.0
228.88
3.0
52.0
253.28
4.0
234.54
1.0
53.0
213.28
4.0
253.28
4.0
57.0
Table C
Step 3
N=
57.0
Step 4
Q1_index =0.25*N =
14.25
Q2_index =0.5*N =
28.50
Q3_index =0.75*N =
42.75
Page 24 of 38
plate survey performed from 3:00 pm to 6:00 pm on 15th April, 2008 (working day). The probes are
randomly selected from the survey individual vehicle data.
CUPRITE is applied on the link (see Figure 11) from Intersection A (“Kaserneplatz”) to intersection D
(“Pilatusplatz”). From A to B, there is minor side street acting as both source and sink; from B to C
there is onstreet bus stop; and from C to D, there are two different movements (left and through)
associated with the link. Around 15% of the vehicles are lost in the side street in between intersection
C to D. The detectors at the site are also not perfect.
The four stopline detectors at A (as1,as2, as3, as4) provide total cumulative plot at the upstream (UT).
The downstream cumulative plots for through movement (DThru) and left movement (DLft) are obtained
from stopline detectors (ds1,ds2) and detector (ds3), respectively. UT is scaled vertically using the
average turning ratio of 55% for through movement and 30% for left movement to define the initial
arrival cumulative plot for each movement.
Figure 11: Illustration of the link characteristics between intersections A and D.
5.1 Average Travel Time estimation
5.1.1 Ground truth travel time
The number plate survey captures the sample of vehicles traversing the link. We are interested in
actual average travel time for all the vehicles departing the link during travel time estimation interval.
Say the mean and standard deviation of the travel time obtained from the survey be
s
X
and Ss,
respectively. We define the confidence bounds in the actual average travel time (µs) of the vehicles as:
Loss
A
C
Figure not to scale
B
Bus stop
Detectors
DThru
DLft
Bus Lane
as1
as2
as3
as4
ds1
ds2
ds3
D
Page 25 of 38
/2, 1 /2, 1
ss
ss
s n s s n
ss
SS
X t X t
nn
(17)
Where:
/2, 1
s
n
t
is the tstatistic with α level of significance and ns1 degrees of freedom; ns is number
of survey vehicles in an estimation interval.
5.1.2 CUPRITE application
As the survey vehicle data is available for a fixed time period and the probe data required for
CUPRITE application is randomly selected from the survey vehicle data. Therefore, for each
estimation interval CUPRITE is applied for nC times with different values of the seed for random
number generator to randomly select probe vehicles. Hence, the application of CUPRITE provides
different travel time estimates for a given estimation interval. Say for an estimation interval the mean
and standard deviation of the estimates be
C
X
and SC, respectively. Then we define the confidence
bounds for the travel time estimate by CUPRITE as:
/2, 1 /2, 1
CC
CC
C n C C n
CC
SS
X t X t
nn
(18)
Where:
µC is the mean of the population of estimates from CUPRITE application;
/2, 1
C
n
t
is the tstatistic at α level of significance and nC1 degrees of freedom.
Figure 12 illustrates an example for the presentation of results. For each estimation interval, the black
box represents the confidence bounds for the ground truth average travel time (see Figure 12a) and
the orange box represents the confidence bounds for the travel time estimates from the CUPRITE
(Figure 12b). Note: In the results, if the mean from the CUPRITE is in within the confidence bounds
from the survey then we can say that the travel time estimates from CUPRITE are statistically
equivalent to that from the survey.
Accuracy of the estimates from CUPRITE is defined as following:
(%) (100 )Accuracy MAPE
(19)
1
ni
i
Error
MAPE n
(20)
( )*100
ii
i
sC
is
XX
Error X
(21)
Page 26 of 38
Where: Errori is the absolute percentage error for ith estimation interval;
i
s
X
and
i
C
X
are the mean of
survey travel time and mean of travel time estimates from CUPRITE application during ith estimation
interval, respectively; n is the number of estimation intervals; and MAPE is the Mean Absolute
Percentage Error obtained from the CUPRITE application for different estimation intervals during
survey period.
Here, the estimation interval is five continuous signal cycles. During the analysis period, the cycle time
varies between 96 s to 116 s. The level of significance for tstatistics considered is 0.05 (=α).
The results presented here are from for A→DLft. Figure 13, Figure 14 and Figure 15 illustrate results
with one, two and three probes per estimation interval (Sn). The orange box overlaps with black box,
indicating that the CUPRITE can estimate the true actual travel time. It can be seen that even the
short term oversaturation in the system can be accurately estimated. For instance, in Figure 13: fourth,
fifth, sixth and seventh estimation intervals (time from 15:30 hr to 16:00 hr) have significant variation in
average travel time between the periods. This fluctuation is also accurately captured by CUPRITE.
For A→DLft: the accuracy (19) of the CUPRITE model increases from 92.3% to 94.6% with increase in
number of probes from one probe per estimation interval (see Figure 13) to three probes per
estimation interval (see Figure 15), respectively.
Figure 12: Systematic representation of the results for CUPRITE validation.
Time
Travel time
Estimation interval
From Survey data
Time
Travel time
Estimation interval
From CUPRITE application
(b)
(a)
/2, 1
s
s
sn
s
S
Xt n
/2, 1
C
C
Cn
C
S
Xt n
/2, 1
s
s
sn
s
S
Xt n
/2, 1
C
C
Cn
C
S
Xt n
s
X
C
X
Page 27 of 38
Figure 13: Results for A→DLft with Sn = 1.
Figure 14: Results for A→DLft with Sn = 2.
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (seconds)
Time (hr:mm)
Survey
CUPRITE
Individual vehicle
A →DLft (Sn=1)
Accuracy (%) = 1MAPE= 92.2%
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (seconds)
Time (hr:mm)
Survey
CUPRITE
Individual vehicle
A →DLft (Sn=2)
Accuracy (%) = 1MAPE= 93.9%
Page 28 of 38
Figure 15: Results for A→DLft with Sn = 3.
5.2 Quartile for travel time estimation
Here slicing technique is applied and quartile Q3 (75th percentile) is estimated. The results are
presented in Figure 16, Figure 17 and Figure 18 for one, two and three probes per estimation interval,
respectively. Here the accuracy of the estimates from CUPRITE is defined as following:
(%) (100 )Accuracy MAPE
(22)
1
ni
i
Error
MAPE n
(23)
( )*100
ii
i
sC
i
s
QQ
Error Q
(24)
Where: Errori is the absolute percentage error for ith estimation interval;
i
s
Q
is the Q3 of survey travel
time during ith estimation interval (As CUPRITE is applied nc number of times (Refer to section 5.1.2),
therefore
i
C
Q
is the 75th percentile of the different values of Q3 of travel time obtained from CUPRITE
application during ith estimation interval.); n is the number of estimation intervals.
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (seconds)
Time (hr:mm)
Survey
CUPRITE
Individual vehicle
A →DLft (Sn=1)
Accuracy (%) = 1MAPE= 94.6%
Page 29 of 38
It is observed that the accuracy increases from 92.4% to 94.7% for increase in Sn from one to three
probes per estimation interval. The results are similar to what we have observed for application of
CUPRITE for average travel time estimation
The above analysis indicates the potential of CUPRITE for quartile travel time estimation in addition to
the average travel time estimation.
Figure 16: Q3 estimation using CUPRITE for route from A→DLft (Sn=1).
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (secomds)
Time (hr:mm)
A→DLft (Sn=1)
Accuracy (%) = 1MAPE= 92.4%
Survey
CUPRITE
Individual vehicle
Page 30 of 38
Figure 17: Q3 estimation using CUPRITE for route from A→DLft (Sn=2).
Figure 18: Q3 estimation using CUPRITE for route from A→DLft (Sn=3).
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (secomds)
Time (hr:mm)
A→DLft (Sn=2)
Accuracy (%) = 1MAPE= 94%
Survey
CUPRITE
Individual vehicle
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (secomds)
Time (hr:mm:ss)
A→DLft (Sn=3)
Accuracy (%) = 1MAPE= 94.7%
Survey
CUPRITE
Individual vehicle
Page 31 of 38
5.3 Discussion on probe as percentage of vehicles traversing the link (Sp)
The results presented in the previous section are with fixed number (Sn) of probes per estimation
interval. In order, to capture the effect of probe market penetration we apply the model with probes as
percentage (Sp) of vehicles traversing the route during three hour of survey period. Here during an
estimation period there can be no probe (Sn=0) or at least one probe (Sn>0) (Refer to Figure 19). In
the present analysis around 60% of the estimation intervals have no probe (Sn=0) for Sp=1% and the
percentage of estimation intervals with Sn=0 decreases with increase in Sp.
Figure 20, Figure 21 and Figure 22, illustrates results for average travel time estimation for Sp equals
1%, 2% and 3%, respectively. The results indicate that even with 1% of the probes CUPRITE can
capture the fluctuations in time series of travel time. The accuracy of the estimation increased from
83.5% to 92.3% with increase in Sp from 1% to 3%, respectively. Similar, Figure 23, Figure 24 and
Figure 25 illustrate results for Q3 estimates for Sp equals 1%, 2% and 3%, respectively. The accuracy
of the Q3 estimation is close to 90%.
Figure 19: Percentage of estimation intervals versus Sn for route A→DLft for different Sp.
0%
10%
20%
30%
40%
50%
60%
0
1
2
3
4
5
6
7
8
Percentage of estimation intervals
Number of probes per estimation interval (Sn)
SP=1
SP=2
SP=3
Sp=1% : Around 60% of the estimation periods
with no probe (Sn=0)
Sp=1%
Sp=2%
Sp=3%
Page 32 of 38
Figure 20: Results for A→DLft with Sp = 1%.
Figure 21: Results for A→DLft with Sp = 2%.
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (seconds)
Time (hr:mm)
Survey
CUPRITE
Individual vehicle
A →DLft (Sp=1%)
Accuracy (%) = 1MAPE= 86.7%
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (seconds)
Time (hr:mm)
Survey
CUPRITE
Individual vehicle
A →DLft (Sp=2%)
Accuracy (%) = 1MAPE= 89.24%
Page 33 of 38
Figure 22: Results for A→DLft with Sp = 3%.
Figure 23: Q3 estimation using CUPRITE for route from A→DLft (Sp=1%).
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (seconds)
Time (hr:mm)
Survey
CUPRITE
Individual vehicle
A →DLft (Sp=3%)
Accuracy (%) = 1MAPE= 91.9%
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (secomds)
Time (hr:mm)
A→DLft (Sp=1%)
Accuracy (%) = 1MAPE= 89.9%
Survey
CUPRITE
Individual vehicle
Page 34 of 38
Figure 24: Q3 estimation using CUPRITE for route from A→DLft (Sp=2%).
Figure 25: Q3 estimation using CUPRITE for route from A→DLft (Sp=3%).
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (secomds)
Time (hr:mm)
A→DLft (Sp=2%)
Accuracy (%) = 1MAPE= 90.2%
Survey
CUPRITE
Individual vehicle
0
50
100
150
200
250
300
350
15:00 15:30 16:00 16:30 17:00 17:30 18:00
Travel time (secomds)
Time (hr:mm)
A→DLft (Sp=3%)
Accuracy (%) = 1MAPE= 90.5%
Survey
CUPRITE
Individual vehicle
Page 35 of 38
6. Conclusions
A majority of research on travel time estimation provide average travel time. The distribution of travel
time from different vehicles departing within an estimation interval is generally bimodal and hence it
can happen that none of the vehicle can experience average travel time. For better understanding of
the network performance statistics, such as quartiles should be explored.
The methodology proposed and validated in this paper provides travel time statistics (average and
quartiles). The methodology is based on the classical analytical procedure for travel time estimation.
The procedure is vulnerable to the relative deviation issue. This issue is addressed by integrating
cumulative plots with probe vehicle data. For this, the probe data is fixed to the downstream
cumulative plot (D(t)) and upstream cumulative plot (U(t)) is redefined: First, the points through which
U(t) should pass are defined by applying grid technique thereafter, the U(t) is redefined by applying
vertical scaling and shifting technique. The average travel time is estimated by applying classical
procedure between redefined U(t) and D(t). For estimation of quartiles, slicing technique is proposed.
The methodology is validated using real data from Luzern city, Switzerland. The application site
represents a typical urban network with: a) mixed traffic (with buses); b) onstreet bus stops;
c) midlink sinks and sources; and d) detector counting error. The validation of the methodology on
real network demonstrates its potential for practical application. The methodology requires few probes
per estimation interval for accurate estimation. The current market penetration of probe is low, and
with limited number of probes per estimation interval, it can considerably enhance the accuracy of
travel time estimation on urban networks.
Though, the development of methodology is based on urban networks, but it can be equally applied to
freeway facilities. It can be easily integrated with traffic monitoring system to simultaneously monitor
both urban and freeway networks.
The probe vehicle data for the methodology is the time when the probe is at upstream and
downstream locations. Advanced loop detectors with the capacity to provide vehicle signatures can be
explored for vehicle reidentification. The reidentified vehicle can be a proxy for probe vehicle data.
For this, even with low reidentification rate, the methodology can accurately estimate travel time.
The proposed methodology accurately estimates travel time, which is the experienced travel time. It
should be extended further, for shortterm travel time prediction, by exploring forecasting techniques,
such as time series analysis, Artificial Neural Network applications etc.
The methodology integrates cumulative plots with probe vehicles. The cumulative plot considered is
two dimensional i.e., both U(t) and D(t) are represented in the same figure. An avenue for extension of
this research is to consider three threedimensional (3D) representations of cumulative plots i.e., to
consider cumulative counts; time; and location from upstream to downstream as three different axis,
respectively. The integration of 3D representation of cumulative plots with the trajectory of a probe
vehicle should be explored for detailed modeling of individual vehicle trajectories.
Page 36 of 38
References
Akçelik, R. (1978), A New Look at Davidson’s Travel Time Function, Traffic Engineering and Control,
19(10), 459463.
Akçelik, R. (1988), Highway Capacity Delay Forumula for Signaliyed Intersections, ITE Journal
(Institute of Transportation Engineers), 58(3), 2327.
Akçelik, R. (1991), Travel Time Functions for Transport Planning Purposes: Davidson's Function, Its
TimeDependent Form and an Alternative Travel Time Function Australian Road Research
21(3), 4959.
Akcelik, R. & Rouphail, N. M. (1993), Estimation of Delays at Traffic Signals for Variable Demand
Conditions, Transportation Research Part B: Methodological, 27(2), 109131.
Akcelik, R. & Rouphail, N. M. (1994), Overflow Queues and Delays with Random and Platooned
Arrivals at Signalized Intersections, Journal of Advanced Transportation, 28(3), 227251.
Berka, S., Tarko, A., Rouphail, N. M., Sisiopiku, V. P. & Lee, D.H. (1995), Data Fusion Algorithm for
Advance Release 2.0, Advance Working Paper Series, No. 48, University of Illinois at Chicago,
Chicago, IL.
Bhaskar, A. (2009), A Methodology (Cuprite) for Urban Network Travel Time Estimation by Integrating
Multisource Data, School of Architecture, Civil and Environmental Engineering, Ph.D Thesis,
The Ecole Polytechnique Fédérale de Lausanne (EPFL). http://library.epfl.ch/theses/?nr=4416.
Bhaskar, A., Chung, E., de Olivier, M. & Dumont, A. G. (2008), Analytical Modelling and Sensitivity
Analysis for Travel Time Estimation on Signalised Urban Network, Transportation Research
Board 87th Annual Meeting, Washington, D.C.
BPR. (1964), Bureau of Public Roads: Traffic Assignment Manual, in U.S. Dept. of Commerce, U. P.
D., Washington D.C (ed.).
Choi, K. & Chung, Y. (2002), A Data Fusion Algorithm for Estimating Link Travel Time, ITS Journal:
Intelligent Transportation Systems Journal, 7(34), 235260.
Daganzo, C. F. (1997), Fundamentals of Transportation and Traffic Operations, Pergamon, Oxford.
Dailey, D. J. (1999), A Statistical Algorithm for Estimating Speed from Single Loop Volume and
Occupancy Measurements, Transportation Research Part B: Methodological, 33(5), 313322.
Dailey, D. J., Harn, P. & Lin, P.J. (1996), Its Data Fusion, Research Project T9903, Task 9 ITS
Research Program, University of Washington
http://www.its.washington.edu/pubs/fusion_report.pdf (Last accessed January 2010).
Dharia, A. & Adeli, H. (2003), Neural Network Model for Rapid Forecasting of Freeway Link Travel
Time, Engineering Applications of Artificial Intelligence, 16(78), 607613.
El Faouzi, N. E. (2004), DataDriven Aggregative Schemes for Multisource Estimation Fusion: A Road
Travel Time Application, in Dasarathy, B. V. (ed.), Proceedings of SPIE  The International
Society for Optical Engineering, Orlando, FL, pp. 351359.
El Faouzi, N. E. (2004), Data Fusion in Road Traffic Engineering: An Overview, in Dasarathy, B. V.
(ed.), Proceedings of SPIE  The International Society for Optical Engineering, Orlando, FL, pp.
360371.
Gault, H. E. (1981), An onLine Measure of Delay in Road Traffic Computer Controlled Systems,
Traffic Engineering and ControL, 22 (7), 384389.
Hamad, K., Shourijeh, M. T., Lee, E. & Faghri, A. (2009), NearTerm Travel Speed Prediction Utilizing
HilbertHuang Transform, ComputerAided Civil and Infrastructure Engineering, 24(8), 551
576.
Hamed, M. M., AlMasaeid, H. R. & Bani Said, Z. M. (1995), ShortTerm Prediction of Traffic Volume
in Urban Arterials, Journal of Transportation Engineering, 121(3), 249254.
Hellinga, B. R. & Fu, L. (2002), Reducing Bias in ProbeBased Arterial Link Travel Time Estimates,
Transportation Research Part C: Emerging Technologies, 10(4), 257273.
Page 37 of 38
Jintanakul, K., Chu, L. & Jayakrishnan, R. (2009), Bayesian Mixture Model for Estimating Freeway
Travel Time Distributions from Small Probe Samples from Multiple Days, Transportation
Research Record, 2136, 3744.
Liu, H., Van Zuylen, H. J., Van Lint, H., Chen, Y. & Zhang, K. (2005), Prediction of Urban Travel Times
with Intersection Delays, IEEE Conference on Intelligent Transportation Systems, Proceedings,
ITSC, Vienna, pp. 10621067.
Long Cheu, R., Xie, C. & Lee, D. H. (2002), Probe Vehicle Population and Sample Size for Arterial
Speed Estimation, ComputerAided Civil and Infrastructure Engineering, 17(1), 5360.
Nam, D. H. & Drew, D. R. (1996), Traffic Dynamics: Method for Estimating Freeway Travel Times in
Real Time from Flow Measurements, Journal of Transportation Engineering, 122(3), 185191.
Palacharla, P. V. & Nelson, P. C. (1999), Application of Fuzzy Logic and Neural Networks for Dynamic
Travel Time Estimation, International Transactions in Operational Research, 6 (1), 145160.
Park, D., Rilett, L. R. & Han, G. (1999), Spectral Basis Neural Networks for RealTime Travel Time
Forecasting, Journal of Transportation Engineering, 125(6), 515523.
Ritchie, S., Park, S., Oh, C., Jeng, S.T. & Tok, A. (2005), Field Investigation of Advanced Vehicle
Reidentification Techniques and Detector Technologies  Phase 2, California Partners for
Advanced Transit and Highways (PATH). Research Reports: Paper prr20058.
Ritchie, S., Park, S., Oh, C. & Sun, C. (2002), Field Investigation of Advanced Vehicle Reidentification
Techniques and Detector Technologies  Phase 1, California Partners for Advanced Transit
and Highways (PATH). Research Reports:UCBITSPRR200215.
Robinson, S. (2005), The Development and Application of an Urban Link Travel Time Model Using
Data Derived from Inductive Loop Detectors, Ph.D thesis, Imperial College London, London,
UK.
Robinson, S. & Polak, J. W. (2005), Modeling Urban Link Travel Time with Inductive Loop Detector
Data by Using the KNn Method, Transportation Research Record, 1935(1), 4756.
Sisiopiku, V. P. & Rouphail, N. M. (1994), Toward the Use of Detector Output for Arterial Link Travel
Time Estimation: A Literature Review, Transportation Research Record(1457), 158165.
Sisiopiku, V. P., Rouphail, N. M. & Santiago, A. (1994), Analysis of Correlation between Arterial Travel
Time and Detector Data from Simulation and Field Studies, Transportation Research
Record(1457), 166173.
Skabardonis, A. & Geroliminis, N. (2005), RealTime Estimation of Travel Times Along Signalized
Arterials, 16th International Symposium on Transportation and Traffic Theory, Maryland, USA.
Spiess, H. (1990), Conical VolumeDelay Functions, Transportation Science, 24(2), 153158.
Srinivasan, K. & Jovanis, P. (1996), Determination of Number of Probe Vehicles Required for Reliable
Travel Time Measurement in Urban Network, Transportation Research Record: Journal of the
Transportation Research Board(1537), 1522.
Tisato, P. (1991), Suggestions for an Improved Davidson Travel Time Function, Australian road
research, 21(2), 85100.
TRB. (1998), Highway Capacity Manual, Special Report 209, National Research Council, Washington
D.C.
TRB. (2000), Highway Capacity Manual, in Board, T. R. (ed.), National Research Council, Washington,
D.C.
VICS. Http://Www.Vics.Or.Jp/English/Vics/Index.Html (Last Accessed January 2010).
Vlahogianni, E. I., Golias, J. C. & Karlaftis, M. G. (2004), ShortTerm Traffic Forecasting: Overview of
Objectives and Methods, Transport Reviews, 24(5), 533557.
Vlahogianni, E. I., Karlaftis, M. G. & Golias, J. C. (2008), Temporal Evolution of ShortTerm Urban
Traffic Flow: A Nonlinear Dynamics Approach, ComputerAided Civil and Infrastructure
Engineering, 23(7), 536548.
Page 38 of 38
VSPLUS. Http://Www.VsPlus.Com/E/Vsintro.Htm (Last Accessed January 2010).
Wardrop, J. G. (1968), Journey Speed and Flow in Central Urban Areas, Traffic Engineering and
Control, 9, 528532.
Webster, F. V. & Cobbe, B. M. (1966), Traffic Signals, Road Research Technical Paper No. 56, Her
Majesty's Stationery Office, London, England.
Xie, C., Cheu, R. L. & Lee, D. H. (2001), CalibrationFree Arterial Link Speed Estimation Model Using
Loop Data, Journal of Transportation Engineering, 127(6), 507514.
Xie, C., Cheu, R. L. & Lee, D. H. (2004), Improving Arterial Link Travel Time Estimation by Data
Fusion, 83rd Annual Meeting of the Transportation Research Board.
You, J. & Kim, T. J. (2000), Development and Evaluation of a Hybrid Travel Time Forecasting Model,
Transportation Research Part C: Emerging Technologies, 8(16), 231256.
Young, C. P. (1988), A Relationship between Vehicle Detector Occupancy and Delay at Signal
Controlled Junctions, Traffic Engineering and Control, 29(3), 131134.
Zhang, H. M. (1999), LinkJourneySpeed Model for Arterial Traffic, Transportation Research
Record(1676), 109115.
Zhang, X. & Rice, J. A. (2003), ShortTerm Travel Time Prediction, Transportation Research Part C:
Emerging Technologies, 11(34), 187210.