ARTICLE IN PRESS

Transportation Research Part C xxx (2010) xxx–xxx

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Identification of oversaturated intersections using high-resolution

traffic signal data
Xinkai Wu a,1, Henry X. Liu a,*, Douglas Gettman b,2
a
Department of Civil Engineering, 500 Pillsbury Drive S.E., Minneapolis, MN 55455, United States
b
Kimley-Horn & Associates, 7878 N. 16th St., Suite 300, Phoenix, AZ 85020, United States

a r t i c l e i n f o a b s t r a c t

Article history: Conceptually, an oversaturated traffic intersection is defined as one where traffic demand
Received 20 April 2009 exceeds the capacity. Such a definition, however, cannot be applied directly to identify
Received in revised form 24 January 2010 oversaturated intersections because measuring traffic demand under congested conditions
Accepted 25 January 2010
is not an easy task, particularly with fixed-location sensors. In this paper, we circumvent
Available online xxxx
this issue by quantifying the detrimental effects of oversaturation on signal operations,
both temporally and spatially. The detrimental effect is characterized temporally by a
Keywords:
residual queue at the end of a cycle, which will require a portion of green time in the next
Traffic signal systems
Oversaturation severity index (OSI)
cycle; or spatially by a spill-over from downstream traffic whereby usable green time is
Shockwave reduced because of the downstream blockage. The oversaturation severity index (OSI), in
Residual queue either the temporal dimension (T-OSI) or the spatial dimension (S-OSI) can then be mea-
Queue-Over-Detector (QOD) sured using high-resolution traffic signal data by calculating the ratio between the unus-
able green time due to detrimental effects and the total available green time in a cycle.
To quantify the T-OSI, in this paper, we adopt a shockwave-based queue estimation algo-
rithm to estimate the residual queue length. S-OSI can be identified by a phenomenon
denoted as ‘‘Queue-Over-Detector (QOD)”, which is the condition when high occupancy
on a detector is caused by downstream congestion. We believe that the persistence dura-
tion and the spatial extent with OSI greater than zero provide an important indicator for
measuring traffic network performance so that corresponding congestion mitigation strat-
egies can be prepared. The proposed algorithms for identifying oversaturated intersections
and quantifying the oversaturation severity index have been field-tested using traffic signal
data from a major arterial in the Twin Cities of Minnesota.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction

With traffic congestion continuing to grow in urban areas of the United States and throughout the world, more and more
signalized intersections are operated under oversaturated conditions. Although a significant amount of literature has been
devoted to how to manage oversaturated traffic signal systems, our understanding of the characteristics of oversaturation,
particularly identification of oversaturation and the transition process from under-saturated condition to oversaturation, is
limited. This research takes one step further toward a better understanding of oversaturation, by providing a coherent meth-
odology to detect the onset of oversaturation at signalized intersections.

* Corresponding author. Tel.: +1 612 625 6347; fax: +1 612 625 7750.
E-mail addresses: wuxxx273@umn.edu (X. Wu), henryliu@umn.edu (H.X. Liu), doug.gettman@kimley-horn.com (D. Gettman).
1
Tel.: +1 612 625 0249; fax: +1 612 625 7750.
2
Tel.: +1 602 906 1332.

0968-090X/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trc.2010.01.003

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
2 X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx

An oversaturated intersection, generally, is defined as one where the demand exceeds the capacity. The degree of satu-
ration, i.e. the volume/capacity ratio, is defined as:
vi
Xi ¼ ð1Þ
ci
where Xi is the degree of saturation for lane group i; vi and ci are demand flow rate and capacity for lane group i, respectively.
A lane group is oversaturated when Xi > 1.
For a single intersection with two competing streams, Gazis (1964) expanded this concept for oversaturation by propos-
ing the following inequality:
qa =sa þ qb =sb > 1  ðL=CÞ ð2Þ
where qa and qb are arrival rates for two conflicting directions; sa and sb are saturation flow rates for two directions; L is the
total lost time and C is the cycle length.
Direct application of the above definitions to detect the onset and quantify the duration and extent of oversaturation,
however, is difficult, partly because of the uncertainty of the capacity and saturation flow, and partly due to the difficulty
to measure the arrival flow using current data collection systems under oversaturated situations (the very condition that
we are trying to identify). Most of the existing detection systems, particularly with inductance loop detectors, provide obser-
vations of traffic flows at a fixed point on a link when they are not fully occupied. Traffic demand is simply not measurable
when a fixed-location detector is occupied with a vehicular queue.
Alternatively, oversaturation has also been characterized as ‘‘a stopped queue that cannot be completely dissipated dur-
ing a green cycle” (Gazis, 1964), or ‘‘traffic queues persist from cycle to cycle either due to insufficient green splits or because
of blockage” (Abu-Lebdeh and Benekohal, 2003). Both studies, however, did not provide a methodology on how to measure
queue length or to identify the existence of such situations.
Another measure that can be utilized potentially to detect oversaturation is green phase utilization. If a signal phase is
oversaturated with a long queue, once the green light starts, vehicles will continue to discharge at the saturation flow rate
until the end of the effective green. The ratio between the ‘‘used” green time and the green phase duration can be treated as
an indicator of saturation for a particular approach. This concept has been utilized by several adaptive traffic control systems,
for example, Sydney Coordinated Adaptive Traffic System (SCATS) (Sims and Dobinson, 1980) and Adaptive Control Soft-
ware-Lite (ACS-Lite) (Luyanda et al., 2003; Gettman et al., 2007). This method, however, cannot estimate the degree to which
a certain traffic phase is oversaturated. The indicator simply identifies that green time is insufficient to serve the traffic de-
mand, but the severity of oversaturation is unknown. In addition, it is necessary to point out that high green utilization of a
given phase is not necessarily an indicator of oversaturation. In some cases, well-coordinated of signalized intersections can
also generate high values of green utilization.
So far, to the best of our knowledge, previous research studies using traffic data from signal systems to diagnose and iden-
tify oversaturation are mostly qualitative and incomplete. Conceptual definitions discussed above are either not applicable in
the real world or having other deficiencies. Since detection of the onset of oversaturation as well as quantifying the severity
of oversaturation is a critical step before appropriate mitigation strategies can be applied, it becomes imperative to have an
implementable and quantifiable measure of oversaturation and a coherent methodology to identify such situation. We hope
that this research fills in this gap.
This paper is organized as follows: The next section discusses a quantifiable measure of oversaturation. Such a measure is
in agreement with conceptual understanding of oversaturation qualitatively, but allows quantitative measurements from a
typical vehicle-actuated traffic signal system. Methodologies for identification and quantification of oversaturation are then
described in Section 3, followed by the results from a field test. Section 5 concludes this study with directions for future
research.

2. A quantifiable measure of oversaturation

Since the general definition of oversaturation, i.e. traffic demand exceeding the capacity of a facility, cannot be applied
directly to detect the occurrence of oversaturation, we propose a measure of oversaturation by quantifying its detrimental
effects. The detrimental effect of oversaturation can be described in temporal and/or spatial dimensions, both of which lead
to the reduction of usable green time in a cycle for a signalized approach.
The detrimental effect of oversaturation in the temporal dimension is characterized by a residual queue at the end of a
cycle. These residual vehicles, which were part of the discharging platoon but did not pass through the intersection in the
current cycle, cannot be discharged due to insufficient green splits, thereby creating detrimental effects on the following cy-
cle by occupying a portion of the green time. The portion of green time that is utilized by the residual queue becomes unus-
able for the traffic arrivals during that cycle.
Detrimental effect of oversaturation in the spatial dimension can be characterized by a spill-over from downstream traf-
fic. When spill-over happens, a downstream link of the intersection is blocked and vehicles cannot be discharged from the
intersection even in green phase. Therefore a portion of the green time becomes unusable, creating detrimental effects. The
most common case causing spill-over is that the downstream link is fully occupied by vehicular queue, but oversaturated
conditions occur only if spill-over from a downstream intersection leads to unusable green time.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx 3

Therefore the condition of oversaturation is characterized by a residual queue at the end of a cycle creating detrimental
effects on the following cycle, and/or by a downstream spill-over within a cycle creating detrimental effects on upstream
traffic facilities. To quantify the detrimental effects in either temporal or spatial dimensions, we introduce the oversaturation
severity index (OSI) by using the ratio between unusable green time and total available green time in a cycle. OSI will be a
non-negative percentage value between 0 and 100, with 0 indicating no detrimental effect for signal operation, and 100
being the worst that all available green time becomes unusable.
We further differentiate OSI into T-OSI and S-OSI. T-OSI describes the detrimental effects created by a residual queue, i.e.
the detrimental effect in temporal dimension; and S-OSI describes the detrimental effects caused by spill-over, i.e. the det-
rimental effect in spatial dimension. Although both T-OSI and S-OSI can be calculated using the ratio between unusable
green time and total available green time, the meanings of ‘‘unusable” are different. For T-OSI, the ‘‘unusable” green time
is the equivalent green time to discharge the residual queue in the following cycle therefore vehicles are discharged at sat-
uration flow rate during that time period; while for S-OSI, the ‘‘unusable” green time is the time period during which an
downstream link is blocked therefore the discharge rate is zero.
Since T-OSI quantifies the detrimental effect of oversaturation on the following cycle, the persistence duration and fre-
quency of T-OSI greater than zero, becomes an important indicator of traffic congestion at the intersection level. On the other
hand, S-OSI describes the detrimental effect of oversaturation caused by downstream queue spillover, indicating the spatial
extent that traffic congestion has spread network-wide.
The differentiation of T-OSI and S-OSI may also help to identify the causal relationship of arterial traffic congestion. For
example, for an intersection, positive T-OSI indicates that the available green time is insufficient for queue discharge and a
residual queue is generated at the end of a cycle. Subsequently, in the following cycles, queue may grow and spillover to
upstream intersection so that S-OSI at the upstream intersection will become positive. Clearly in this case a positive
S-OSI at the upstream intersection is caused by the downstream bottleneck. Note that the downstream bottleneck may lead
to the situation that both S-OSI and T-OSI of the upstream intersection are greater than zero, simply because a portion of the
green time is wasted due to downstream blockage (i.e. S-OSI > 0), and a residual queue may be generated (i.e. T-OSI > 0) due
to the reduction of green time. Therefore positive T-OSI and S-OSI indicates that traffic congestion may start to spread even
further upstream.
With the proposed oversaturation severity index, the focus of the identification algorithms shifts from measuring travel
demand to quantifying detrimental effects in both temporal and spatial dimensions. It is necessary to point out that the classi-
fication and quantification of detrimental effects created by oversaturation are very important for traffic management, be-
cause different oversaturated situations may call for different strategies to mitigate congestion. For example, for an isolated
intersection with positive T-OSI values, extension of green time to discharge residual queue may be sufficient (Quinn, 1992).
While for an arterial corridor with multiple intersections having positive T-OSI and S-OSI values, simultaneous or even neg-
ative offsets (see Pignataro et al., 1978) may be needed to prevent further deterioration of an oversaturated condition. Since
the focus of this paper is to identify and quantify oversaturated conditions, how to map OSI with different mitigation strat-
egies is left for future research.
In the following, we propose two algorithms for the identification and quantification of oversaturated conditions, one for
estimation of residual queue length, and the other for detection of spill-over conditions. Our focus here is not only to identify
oversaturation qualitatively, but also to provide a method that can quantify the severity of oversaturation.

3. Algorithms for identification of oversaturation

The identification algorithms discussed in the following section will work with typical detector (with size of 60  60 ) con-
figurations for a vehicle-actuated signalized intersection, i.e. with either stop-line detector for vehicle presence detection or
advance detector (a few hundred feet upstream from stop line) for green extension, or both. Throughout of this paper, we
will assume advance detectors are available and we will note the necessary changes to be made if the only available detector
is located at the stop-bar. We also assume that high-resolution (i.e. second-by-second or event-based) traffic signal data can
be collected. The availability of high-resolution traffic signal data has been greatly increased in recent years. For example,
second-by-second detector data has been used by ACS-Lite (Luyanda et al., 2003). Continuous event-based signal data
(including both vehicle-detector actuation events and signal phase change events) has been collected and archived by the
SMART-Signal system (Systematic Monitoring of Arterial Road Traffic and Signals) developed at the University of Minnesota
(Liu and Ma, 2009). Although the algorithms presented in this paper are demonstrated by using event-based data from the
SMART-Signal system, they are also applicable to second-by-second signal data coming from any other traffic signal man-
agement system.

3.1. Algorithm for residual queue length estimation

Residual queue at an intersection refers to those vehicles that are part of the discharging platoon cannot pass the inter-
section during the green time. Residual queue also represents the minimum queue length at the end of a cycle. Vehicles in
the residual queue will then occupy a portion of green time in the next cycle for discharging. The ratio between the residual
queue discharging time and the total available green time is then denoted as T-OSI, as we discussed above.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
4 X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx

Estimation of residual queue length requires the reconstruction of queue length profile within a cycle, including the max-
imum queue length that has been reached. As the traditional input–output approach for queue length estimation can only
handle queues that are shorter than the distance between vehicle detector and intersection stop line, in this paper we adopt
the queue length estimation method developed by Liu et al. (2009). Both maximum and minimum queue lengths can be cal-
culated using this method. In the following, for the sake of completeness, we provide a brief discussion of the queue length
estimation algorithm, with a focus on its’ application to oversaturated conditions. Interested readers should refer to Liu et al.
(2009) for more details. We should note here that the queue estimation method discussed in Liu et al. (2009) needs to be
modified for the case of queue spill-over from downstream, as will be discussed in Section 3.2.
The queue estimation method is based on the identification of traffic state changes and the associated shockwaves pre-
sented in a cycle. As indicated in Fig. 1, assuming there is no residual queue from past cycles, at the beginning of red, vehicles
arrive at the intersection are forced to stop which creates a queuing shockwave (v1) propagating backward from the stop line.
At the beginning of the effective green, vehicles begin to discharge at the saturation flow rate (assume there is no blockage
downstream) forming the discharge shockwave (v2), again propagating upstream from the stop-line. The discharge shock-
wave v2 usually has higher speed than v1, so the two waves will meet some time after the start of the green, which is the
time that the maximum queue length is reached. As soon as the two shockwaves meet, a departure shockwave (v3 in
Fig. 1) is generated, propagating toward the stop line. Here the front of the departure shockwave reflects the discontinuity
between the saturated discharging traffic flow and the new traffic arrivals after the maximum queue length is reached. A
residual queue is formed some time after the start of the red light of the next cycle when the departure shockwave v3 meets
the compression shockwave (v4). The compression shockwave (v4) is similar to the queuing shockwave (v1), as both shock-
waves form a stationary queue. The difference is that compression shockwave represents traffic discontinuity from saturated
traffic flow to the ‘‘jammed” traffic condition, while the queuing shockwave represents a change to the jammed traffic con-
dition from arriving traffic, which is not necessary in a saturated state. The shockwave motion described above will repeat
from cycle to cycle.
Intuitively, the queuing profile including the maximum queue length (Lnmax , i.e. the maximum queue length for the cycle n)
and the minimum queue length (Lnmin , i.e., the residual queue length for the cycle n) can be easily derived if the times when
shockwaves cross the detector location, which present the changes of traffic state, can be identified (indicated as ‘‘break
points” A, B, and C in Fig. 1). High-resolution detector data, which provide detailed detector occupancy time and vehicle gaps
between consecutive vehicles, imply the changes of traffic state, therefore can be utilized to identify these break points. Fig. 2
presents an example of detector occupancy time (see Fig. 2a) and vehicle gaps (see Fig. 2b) within a cycle collected from an
advance detector at one intersection in the Twin Cities, Minnesota (the data collection details will be introduced in Section 4).
As shown in the figure, a sudden increase of occupancy time indicates that queue (v1) spills back to the advance detector;
therefore the time instant of break point A (TA) can be identified when occupancy time is significantly increased. Similarly,
the time instant of break point B (TB) can be identified when occupancy time drops to the normal value; as this change indi-
cates that discharge shockwave (v2) propagates back to the advanced detector and vehicle begins to move. Point C represents
the time when departure shockwave v3 crosses the detector line. Before point C appears, vehicles discharge at the saturation
flow rate, i.e. the saturation traffic state (qm, km); and after TC, the traffic condition changes to arrival traffic state (qa, ka). The
change of these two traffic states (from saturation to free flow arrival) is indicated by the variation of time gaps. As indicated
in Fig. 2b, before TC, the time gaps between vehicles are consistently small (less than 2.5 s), meaning that most of vehicles are

Distance
H
Trajectory of Shockwave

v3
Lnmax
v1

Loop A B C A'
Detector v2
D v5
Residual Queue
Ld v1
n
L
min v4

Tgn TA Trn TB Tmax

n
TC Tgn +1 Tmin
n
Trn +1 Time

Fig. 1. Shockwave profile within a cycle.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx 5

(a) Detector Occupancy Time

(Sec)
50

40

30

20

10 Break Point A Break Point B

0 Time
7 : 2 6: 07 7 : 2 6: 50 7 : 2 7: 33 7: 28 : 1 6 7: 29 : 0 0

(b) Time Gap Between Consecutive Vehicles

(Sec)
10

Pattern I: Saturation condition (qm , km )

7. 5
n n
Pattern II: Free flow arrival (qa , ka )

5 Break Point C

2. 5

0
7: 26 : 0 7 7 : 2 6: 50 7 : 2 7: 33 7: 28 : 1 6 7 : 2 9: 0 0 Time

Fig. 2. (a) Detector occupancy profile in a cycle. (b) Time gap between consecutive vehicles in a cycle.

discharged at saturation flow rate. But after TC, the vehicle gaps become much larger and their variances are significantly
increased. Therefore a threshold value of time gap can be used to identify TC.
Once break points (A, B, and C) have been identified, the flow and density of each traffic state (i.e. the arrival traffic state
(qa, ka), saturation traffic state (qm, km), and jammed traffic state (0, kj)) can be calculated based on detector occupancy times
and time gaps between vehicles. Then wave speeds of v1, v2, and v3 can be estimated using following equation:
Dq q2  q1
v¼ ¼ ð3Þ
Dk k2  k1
where q1, k1 are the flow and density of the upstream traffic and q2, k2 are the flow and density of the downstream traf-
fic.Using the estimated shockwave speeds from above equation, the maximum queue (both length Lnmax and time T nmax )
and the minimum queue (both length Lnmin and time T nmin ) during nth cycle can be calculated based on the shockwave profile
(see Fig. 1) by the following equations:
8  
< Ln ¼ Ld þ ðT C  T B Þ= v12 þ v13
max
ð4Þ
: T n ¼ T þ L n  L = v
max B max d 2

8    
< Ln Lnmax n nþ1
min ¼ v 3 þ T max  T g = v13 þ v14
ð5Þ
: Tn ¼ T nþ1 þ Lnmin =v 4
min g

where Ld is the distance from the stop line to the loop detector; and T nþ1
g is the green end of (n + 1)th cycle.
Note that it is necessary to identify whether a residual queue is present at the end of the cycle before the calculation of the
minimum queue length:
8 n
< Lmax þ T nmax < T nþ1 without residual queue
v3 g
ð6Þ
: Lnmax þ T n P T nþ1 with residual queue
v3 max g

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
6 X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx

For severely congested traffic conditions, the break point C may not be found during the green phase. In such cases, the
traffic pattern does not change during the green phase and vehicles keep discharging at the saturation flow rate (see Fig. 3).
Then a residual queue must exist, at least, between the detector location and the stop line. Eq. (4) cannot be applied to cal-
culate the maximum queue length since shockwave speed v3 cannot not be calculated. Under such conditions, the complete
queue profile cannot be recovered from the detector data. However, since the entire green time has been used for queue dis-
charge, the number of vehicles passing detector locations during the green time can be counted (between TB and T nþ1 g ), so

that a minimum of the maximum queue length, i.e. min (Lnmax ), can be estimated by simply taking the end of cycle T gnþ1
as TC (see Fig. 4). As the space headway at jammed traffic conditions (ljam) and the velocity of the discharge wave (v2) can
be assumed constant, the following equation holds:
8  
< min Lnmax ¼ ljam  N þ Ld
If point C cannot be identified :   min ðLnmax Þ ð7Þ
: min T n n
max ¼ T r þ v2

where N is the traffic count between TB and T gnþ1 ; ljam is the space headway at jammed traffic conditions (assumed as a known
constant); and T nr is the end of the red phase of nth cycle.
Then v3 can be calculated by:
 
min Lnmax  Ld
If point C cannot be identified : v3 ¼   ð8Þ
T nþ1
g  min T nmax

The coordinate of the minimum of the residual queue length, i.e. min (Lnmin ) and min (T nmin ) for Point D, can then be esti-
mated using Eq. (5).
If residual queue exists at the end of a signal cycle, some portion of the green time in the following cycle will be utilized to
discharge the residual vehicles, therefore becomes the ‘‘unusable” green time for that cycle. The unusable green time can be
calculated by using the number of vehicles in the residual queue multiplied by the saturation discharge headway (around
2 s). The detrimental effect caused by residual queue therefore can be quantified by the oversaturation severity index in tem-
poral dimension (T-OSI):

(a) Detector Occupancy Time

(Sec)
80

60

40

20
Break Point A Break Point B
0
1 7: 27: 06 1 7: 27: 49 1 7: 28: 32 1 7: 29: 15 1 7: 29: 59 Time

(b) Time Gap Between Consecutive Vehicles

(Sec)
12

No pattern changes, always discharge at saturation

9 flow rate; Point C cannot be identified

0
1 7: 27: 06 1 7: 27: 49 1 7: 28: 32 1 7: 29: 15 1 7: 29: 59 Time

Fig. 3. Break points identification (point C cannot be identified).

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx 7

Distance
H
v3 Trajectory of Shockwave

min( Lnmax )
v1

Loop A B C A' v5
Detector v2
D
Ld min( Lnmin )
v1
v4

Tgn TA Trn TB min(Tmax

n
) Tgn+1 (TC ) min(Tmin
n
) Trn +1 Time

Fig. 4. Calculation of residual queue length when point C cannot be found.

unusable green time Ln =ljam  h

T-OSI ¼  100% ¼ min  100% ð9Þ
total available green time G
where G is the effective green time, and h is the saturation discharge headway.
We should note that, when only stop-line detection is available, the residual queue will occur if ‘‘break point” C cannot be
identified within the green time. Although the length of the residual cannot be measured with the stop-line detection only, it
is sufficient to say that oversaturation may have occurred at this intersection for this cycle, i.e. T-OSI > 0.

3.2. Algorithm for the identification of a spill-over

Spill-over creates detrimental effects for the operation of upstream traffic signals. Identification of spill-over is particu-
larly important because it indicates that traffic congestion has started to spread out in the network involving multiple inter-
sections. To identify spill-over using traffic signal data, we first need to illustrate the concept of Queue-Over-Detector (QOD),
i.e. complete occupation of a detector for a relatively long time due to vehicular queue.
Generally, there are two types of QOD. One is caused by red signal phases. Due to the cyclic signal timing, vehicles slow
down and stop due to the red light, and then resume traveling as the light turns green while the queuing wave propagates
back. If a vehicle in the queue stays on the detector because of a red light, detector occupancy time could increase contin-
uously, creating the first type QOD. The second type QOD is caused by a spill-over. When a queue spills back from down-
stream intersection to a upstream one, the upstream intersection may be blocked and vehicles cannot be discharged even
when the signal is green. Some vehicles will stay on the detector for a relatively long time creating prolonged detector occu-
pancy time after traffic light turns green. Conceptually, the duration of the second type of QOD is equivalent to ‘‘the time
period within a signal cycle in which the vehicles are certainly moving at the location of the detectors during green time
in the absence of disturbances”, which was briefly discussed in Muck (2002). Therefore, the detrimental effect of a spill-over
can be quantified by measuring the duration of the second type of QOD.
It is not difficult to identify QOD using high resolution data; as it is indicated by a relatively large occupancy time (or per-
centage occupancy value keeping at 100% for some time). In our implementation, a threshold value of 3 s (which is roughly
equivalent 5 mph of speed assuming a 22 ft. effective vehicle length), is used for the identification of QOD.
We now need to distinguish between the two types of QOD since the first one does not indicate spill-over. Fig. 5 demon-
strates both types of QOD by drawing each vehicle trajectory starting from upstream to downstream. Since the first type
QOD is caused by red signal, the maximum occupancy time is the red interval. Consider the residual queue from the last cycle
and queue propagation at the green start, the first type QOD can only happen within the range between two shockwaves
(compression shockwave v4 and discharge shockwave v2, which have the same velocity, see Fig. 4). Therefore if QOD occurs
   
between T ng þ Ld =v 4 and T nr þ Ld =v 2 , it is of the first type.
h i
The second type of QOD, which occurs outside of the time interval T ng þ Ld =v 4 ; T nr þ Ld =v 2 , indicates that a spill-over has
happened at a downstream location. This creates unusable green time, meaning that vehicles cannot be discharged during
the green time because of downstream congestion. Therefore when a QOD event h is identified, spill-overi occurs when the
QOD starting time T QOD QOD n n
start or ending time T end falls outside of the time interval T g þ Ld =v 4 ; T r þ Ld =v 2 . An example case of
the second type of QOD is demonstrated in Fig. 5.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
8 X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx

Distance

Downstream

Stop-bar detector

v2
v2
Advance detector
A C S E

QOD caused by QOD caused by

red interval spillover

Upstream Time

Fig. 5. Queue-Over-Detector phenomena.

The oversaturation severity index in space dimension (S-OSI) can then be calculated as:
P  QOD 
unusable green time T end;i  T QOD
start;i
S-OSI ¼  100% ¼  100% ð10Þ
total available green time G

where T QOD QOD

start;i and T end;i are the starting and ending times of the ith occurrence of the second type of QOD.
In order to improve the robustness of the identification of the spill-over condition, the maximum queue length of the
downstream intersection should also be estimated using the method discussed in Section 3.1. If the estimated maximum
queue length is longer than or equal to the link length, then oversaturation is confirmed. This can be used to avoid some
diagnosis errors caused by ‘‘incidents” (for example, the detector is occupied by a broken vehicle for a relatively long time).
These ‘‘incidents” may generate the second type of QOD, but does not necessarily indicate an oversaturated condition.
Here we should note that the queue length estimation method discussed in Section 3.1 cannot be applied directly to an
intersection with queue spillover from the downstream. With spillover, queued vehicles can only be discharged when the
downstream blockage is cleared and signal light remains green. For the example shown in Fig. 6, when traffic light turns

Distance
H QOD Caused by Spillover

v3

Lnmax C
v1

Loop A B C'
Detector v2 A' B'
D

Ld v1 v2 Lnmin
v2 v4

Time
Intersection i Tgn TA Trn TB QOD
Tstart
QOD
Tend Trn +1

v1
v2
v4
Time
Intersection i+1

Fig. 6. Shockwave profile with downstream spillover.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx 9

green at intersection i (T nr ), queued vehicles start to discharge, but the discharging process is disturbed because queue at
downstream intersection i + 1 grows and eventually spills over to upstream intersection i. A second type of QOD will be iden-
tified by the advance detector, starting at time T QOD QOD
start . QOD ends when the spillover is cleared, i.e. at time T end . Under such
saturations, the queue estimation method needs to be modified because the break points A, B, and C need to be updated
as A0 , B0 , and C0 , as shown in the figure. Eqs. (4)–(8) will remain valid after the update.

4. Field-test results

4.1. Data collection

Trunk Highway 55, a major arterial in the Twin Cities, Minnesota, was used as the test site. Fig. 7a illustrates the six coor-
dinated intersections at the test location. Fig. 7b shows the detector layout of four intersections where oversaturation has
been identified to occur in a recurrent fashion. All six intersections shown in the figure are equipped with vehicle-actuated
traffic controllers, with advance detectors on the major approach and stop-bar detectors on the minor approach, for the pur-
pose of traffic signal operations. Stop-bar detectors are used to detect the presence of vehicles and advance detectors are
located about 400 feet upstream from the stop line to detect vehicles for green extension on the coordinated phases. For
the purpose of verification of the estimated queue length, we have also installed stop-bar and link entry detectors along
TH55 at the six intersections (see Fig. 7b for the detector configurations at the four intersections of interest). These additional
detectors are not used for regular traffic signal operations; rather, the collected data from these detectors provide further
evidence for the results presented below.
High-resolution event data including signal phase changes and vehicle-detector actuations are continuously collected
from the six intersections and are archived by the SMART-Signal system and transmitted back to the University of Minnesota
lab in real-time. Fig. 8 shows the sample data collected at the study site. SMART-Signal data provides start and end times of
each vehicle-detector actuation ‘‘event” and every signal phase change ‘‘event”. Therefore, the signal phase duration can be
calculated from the time difference between the start and end of a signal event; and the time interval between the start and
end of a vehicle actuation event is the detector occupancy time. The time interval between the end of a vehicle actuation
event and the start of next vehicle actuation event (from the same detector) is the time gap between two consecutive vehi-
cles crossing the detector.

4.2. Estimation results of residual queue length

Using the event-based data from the SMART-Signal system, the queue length estimation method discussed above was
applied for the estimation of residual queue length. We note that field evaluations of the queue length estimation method

(a)

(b) Boone Ave. Winnetka Ave. Rhode Island Ave.

TH 55 842 ft 1777 ft Phase 6

2635 ft

375 ft

400 ft Phase 2 TH 55
Advance detectors
Stopbar detectors
Additional detectors Glenwood Ave.

Fig. 7. (a) Data collection site (from Google Map). (b) Detector layout.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
10 X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx

Detector #8 on at 08:09:15.012;
08:09:15.012, D8 on, 7.902s Vacant time is 7.902s
08:09:15.481, D8 off, 0.468s
Green Phase #3 off at 08:09:16.761;
08:09:16.761, G3 off, 29.389s
Green duration time is 29.389s
08:09:16.761, Y3 on, 179.021s
08:09:17.620, D9 on, 2.686s
08:09:18.151, D10 on, 2.593s
08:09:18.307, D9 off, 0.687s Detector #9 off at 08:09:18.307;
Occupy time is 0.687s
08:09:18.823, D10 off, 0.671s
Yellow Phase #3 off at 08:09:20.244;
08:09:20.244, Y3 off, 3.482s
Yellow duration time is 3.482s
08:09:21.649, D22 on, 80.953s
08:09:22.008, D22 off, 0.359s
Green Phase #1 on at 08:09:23.242;
08:09:23.242, G1 on, 172.806s
Red duration time is 172.806s

Fig. 8. Sample data collected at the test site.

have been conducted by a Minneapolis-based transportation consulting firm and the evaluation results have been reported
in Liu et al. (2009). The average of the Mean Absolute Percentage Error (MAPE) of the queue length estimation is within 15%.
Fig. 9 presents an oversaturation case based on the data collected by an advance detector on eastbound direction at the
intersection of Glenwood Ave. (see Fig. 7a) on February 28, 2008. As indicated in Fig. 9, residual queues appeared at the end
of the first two cycles, meaning that these two cycles were oversaturated. In this particular case, the reason for the occur-
rence of oversaturation is signal preemption, which created a shorter cycle length in the second cycle (the cycle length was
132 s during the preemption that was 48 s less than the normal cycle length). Due to the insufficient green time, some
queued vehicles could not be discharged until the next cycle, creating detrimental effects on the following cycles.
Using the estimated residual queue length, the oversaturation severity indices for these two cycles are estimated at 7.5%
and 7.0%, meaning that at least 7.5% and 7.0% of green time in these cycles will be used for the discharge of the residual
queue. In the calculation of these severity indices, we have assumed that the space headway for the jammed condition is
25 ft. and the saturation headway is two seconds.
It should be noted that the estimated maximum queue lengths (500–600 ft.) during these three cycles are not long when
compared with the link length (1777 ft. from Glenwood to Rhode Island). However, residual queues occured at the end of the
first two cycles indicating that although the maximum queue lengths (which include stopped vehicles only) were not long,
the traffic volume joining the discharge platoon after the last stopped vehicle started to move was rather high. A portion of
those newly arriving vehicles, which joined the discharge platoon but could not pass the intersection during the green phase,
formed the residual queue.
As we discussed previously, if the entire green time has been used for queue discharge, the departure shockwave cannot
be identified. In this case we can only measure the minimum of the maximum queue length within a cycle. We present such
cases in Fig. 10 below. The data was collected from an advance detector on eastbound direction at the intersection of Boone
Ave. (see Fig. 7) on February 28, 2008. During the first five cycles, the departure shockwaves could not be identified, so we

Case
Queue Length of Over
Profile sat ur at i onGlenwood
at Eastbound
Distance (feet)
700

600

500

400

300

200

100 Residual Queue

0 Time
16: 48 : 00 16: 49 : 26 16: 50: 53 16: 52: 19 16: 53: 46 16: 55: 12 16: 56: 38 16: 58: 05

Fig. 9. Estimation results of residual queue for eastbound approach at Glenwood.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx 11

Distance (feet) Queue Length Profile Queue Tr a j ect or i es

for Eastbound Approach at Boone Ave
1800

1500

1200

900

600

300 Residual Queues

G=98sec G=112sec G=98sec G=114sec G=115sec G=116sec
0
7: 48: 00 7: 50: 53 7: 53: 46 7: 56: 38 7: 59: 31 8: 02: 24 8: 05: 17 Time

Fig. 10. Estimated residual queue length for eastbound approach at Boone Ave.

can only estimate the minimal values of the maximum queue length. The queue lengths in this case were quite long, aver-
aging around 1500 ft. in the first five cycles. The minimum values of the residual queue length were also estimated, as shown
in Fig. 10. The minimal oversaturation severity indices (T-OSI) are estimated at 9.8%, 19.4%, 10.5%, 11.3%, and 10.3% for these
five cycles. The oversaturated condition persisted until the sixth cycle.

4.3. Results of spill-over detection

The spatial detrimental effect caused by oversaturation at the arterial/network level is characterized by spill-overs, which
can be diagnosed by identifying the second type of QOD, as discussed in Section 3.2. In Fig. 11, we present the detector occu-
pancy time within a cycle for an afternoon peak hour on November 17, 2008 for westbound TH55 at Rhode Island Ave. As
shown in the figure, the second type of QOD, i.e. QOD caused by spill-over, is identified. This means that vehicles cannot be
discharged from the intersection although the traffic light is green, due to a spill-over happening in the downstream link.
Oversaturation is therefore identified at this intersection for this cycle.
To further verify that there is a spill-over happening in the downstream link, vehicle trajectories are derived based on the
vehicle events collected by the advance detector at the intersection of Rhode Island Ave. A simplified car-following model,
which is similar as the one proposed by Newell (2001), is used to generate the vehicle trajectories. Vehicles are assumed to
accelerate if their speeds are less than free-flow speed (55 mph, in this study) and decelerate if higher than the free-flow
speed. The distance between any two vehicles should satisfy a safety distance constraint, which is determined by the speeds
of the two consecutive vehicles. The driver’s reaction time is set at 1.0 s and space headway between two vehicles in a sta-
tionary queue is assumed as 25 ft. The maneuver decision of a vehicle to pass the intersection or not during the yellow time
is determined according to its speed, remaining yellow time, and distance from the front vehicle. Lane-changing behaviors
are not taken into account in this model. The estimated vehicle trajectories starting from the advance detector line at the
intersection of Rhode Island Ave. and ending at 500 ft. downstream from the intersection of Winnetka Ave. are presented
in Fig. 12. As clearly indicated in the figure, queue spills back from Winnetka to Rhode Island and blocks the Rhode Island
intersection during green time, resulting in the second type of QOD.
Further investigations indicate that this spill-over started at 17:06:31, stayed about 30 min, and ended at 17:36:31 as the
second type of QOD was found in nine cycles consecutively. This is also confirmed by looking into the queue length profile of

Occupat
Detector Occupancy Time fori on Ti me
Westbound TH 55 at Rhode Island Ave.
Distance (feet)
7800

Stop-bar
7600
Detector
7400 QOD caused by red phase QOD caused by spillover
v2 v2
7200 Advance
Detector
7000
Occupancy Time of Stop-bar Detector
Occupancy Time of Advance Detector
6800
17: 12: 20 17: 13: 03 17: 13: 47 17: 14: 30 17: 15: 13 Time

Fig. 11. Identification of the second type QOD using detector occupancy data.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
12 X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx

Vehicle Trajectories in the Case ofTrSpillover

aj from Winnetka to Rhode Island
Distance (feet)
9000

8500 Intersection
Winnetka

Location of
8000
Advance Detector

Intersection
7500 Rhode Island

Location of
Spillover Advance Detector
7000
17 : 12: 03 17 : 12 : 46 17: 13 : 29 17 : 14 : 12 17 : 14 : 56 17 : 15 : 39 17 : 16 : 22 17 : 17 : 05
Time

Fig. 12. Vehicle trajectories in the case of spill-over from Winnetka to Rhode Island.

Queue Length Profile at the Intersection of Winnetka

Distance (feet)
1800

1500

1200

900

600

300

0 Time
17: 05: 17 17: 12: 29 17: 19: 41 17: 26: 53 17: 34: 05 17: 41: 17

Fig. 13. Queue length profile at the intersection of Winnetka when Rhode Island Intersection is oversaturated.

Table 1a
Oversaturation severity indices (OSI) for Winnetka Ave. intersection.

Cycle start Available green (s) OSI: created by residual queue OSI: created by spillover
Residual queue (ft.) Unusable green (s) T-OSI (%) Unusable green (s) S-OSI (%)
Winnetka Ave.
17:06:14 101 0.0 0.0 0.0 0.0 0.0
17:09:14 101 180.3 0.0 0.0 0.0 0.0
17:12:14 101 178.8 14.4 14.28 0.0 0.0
17:15:14 101 0.0 14.3 14.16 0.0 0.0
17:18:14 101 149.1 0.0 0.00 0.0 0.0
17:21:14 101 157.6 11.9 11.81 0.0 0.0
17:24:14 102 156.4 12.6 12.36 0.0 0.0
17:27:14 106 130.1 12.5 11.81 0.0 0.0
17:30:14 101 153.4 10.4 10.31 0.0 0.0
17:33:14 105 0.0 12.3 11.69 0.0 0.0
17:36:14 102 0.0 0.0 0.00 0.0 0.0

the downstream intersection at Winnetka Ave. during these cycles. As shown in Fig. 13, the maximum queue lengths for
these nine cycles are around 1200–1500 ft., which is significantly longer than the link length (842 ft.). This indicates that,
during these cycles, the Rhode Island intersection must be blocked for a portion of green time (S-OSI > 0). Interestingly, be-
cause of the reduction of usable green time, residual queues were also generated at Rhode Island Intersection for some cycles
(i.e. T-OSI > 0). This demonstrates that oversaturated traffic condition at Winnetka (T-OSI > 0 and S-OSI = 0) has spread to

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003
 ARTICLE IN PRESS
X. Wu et al. / Transportation Research Part C xxx (2010) xxx–xxx 13

Table 1b
Oversaturation severity indices (OSI) for Rhode Island intersection.

Cycle start Available green (s) OSI: created by residual queue OSI: created by spillover
Residual queue (ft.) Unusable green (s) T-OSI (%) Unusable green (s) S-OSI (%)
Rhode Island
17:06:31 136 0.0 0.0 0.0 0.0 0.0
17:09:31 136 0.0 0.0 0.0 3.0 2.2
17:12:31 136 89.6 0.0 0.0 28.0 20.6
17:15:31 136 164.3 7.2 5.3 28.8 21.2
17:18:31 136 0.0 13.1 9.7 15.0 11.1
17:21:31 136 180.4 0.0 0.0 41.7 30.6
17:24:31 135 165.3 14.4 10.7 34.1 25.2
17:27:31 139 138.2 13.2 9.5 25.2 18.1
17:30:31 120 125.3 11.1 9.2 16.3 13.6
17:33:31 141 0.0 10.0 7.1 8.6 6.1
17:36:31 135 0.0 0.0 0.0 0.0 0.0

upstream, leading to insufficient green at Rhode Island to discharge queue (T-OSI > 0 and S-OSI > 0). Please see Tables 1a and
1b for oversaturation severity indices including both T-OSI and S-OSI at Winnetka and Rhode Island. Note for Winnetka, since
there was no downstream blockage, S-OSI was always zero during that time period.

5. Concluding remarks

In this paper, we propose a quantifiable measure of oversaturation by focusing on its detrimental effects in both temporal
and spatial dimensions. The temporal detrimental effect is characterized by the occurrence of a residual queue and the spa-
tial detrimental effect is referred to a spill-over from downstream intersection to upstream. We propose two algorithms to
identify oversaturated signalized intersection, one for the estimation of residual queue length using shockwave based meth-
od and the other for detection of spill-over by identifying long detector occupancy time during green phase. We define an
oversaturation severity index using the unusable green time caused by residual queue or spill-over. Our field-test results
from a major arterial in the Twin Cities area demonstrate that the developed algorithms are very effective in identifying
oversaturated conditions. We should note that, although the identification algorithms focus on a single approach of a signal-
ized intersection, the proposed methodology can be easily expanded to identify oversaturation of an intersection with multi-
ple approaches, an arterial, or a network of intersections.
This research is expected to contribute to the future research on the mitigation of signal oversaturation. As mentioned
before, different types of OSI (T-OSI or S-OSI) and different levels of OSI may require different mitigation strategies. There-
fore, future research may focus on how to map the identified OSI values with various mitigation strategies for better man-
agement of traffic signals under oversaturated conditions.

Acknowledgements

This research was partially supported by the National Cooperative Highway Research Program (NCHRP). The views ex-
pressed in this paper are those of the authors and do not necessarily reflect the positions or policies of the NCHRP Program.
Our sincere thanks go to Steve Misgen and Ronald Christopherson of Minnesota Department of Transportation for their help
on the data collection. The authors would also like to thank the anonymous reviewers for their insightful and constructive
comments.

References

Abu-Lebdeh, G., Benekohal, R., 2003. Design and evaluation of dynamic traffic management strategies for congested conditions. Transportation Research
Part A 37 (2), 109–127.
Gazis, D.C., 1964. Optimal control of a system of oversaturated intersections. Operations Research 12, 815–831.
Gettman, D., Shelby, S.G., Head, L., Bullock, D.M., Soyke, N., 2007. Data-driven algorithms for real-time adaptive tuning of offsets in coordinated traffic signal
systems. Transportation Research Record, 1–9 (No. 2035).
Liu, H., Ma, W., 2009. A virtual probe approach for time-dependent arterial travel time estimation. Transportation Research Part C 17 (1), 11–26.
Liu, H., Wu, X., Ma, W., Hu, H., 2009. Real-time queue length estimation for congested signalized intersections. Transportation Research Part C 17 (4), 412–427.
Luyanda, F., Gettman, D., Head, L., Shelby, S., Bullock, D., Mirchandani, P., 2003. ACS-lite algorithmic architecture: applying adaptive control system
technology to closed-loop traffic signal control systems. Transportation Research Record, 175–184 (No. 1856).
Muck, J., 2002. Using detectors near the stop-line to estimate traffic flows. Traffic Engineering and Control 43, 429–434.
Newell, G.F., 2001. A simplified car-following theory: a lower order model. Transportation Research Part B 36 (3), 195–205.
Pignataro, L.J., McShane, W.R., Crowley, K.W., 1978. Traffic Control in Oversaturated Street Networks. National Cooperative Highway Research Program
Report 194. Transportation Research Board, Washington, DC.
Quinn, D.J., 1992. A review of queue management strategies. Traffic Engineering and Control 3311, 600–605.
Sims, A.G., Dobinson, K.W., 1980. The Sydney coordinated adaptive traffic (SCAT) system philosophy and benefits. IEEE Transactions on Vehicular
Technology 29 (2), 130–137.

Please cite this article in press as: Wu, X., et al. Identification of oversaturated intersections using high-resolution traffic signal data. Trans-
port. Res. Part C (2010), doi:10.1016/j.trc.2010.01.003