Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Spatiotemporal Analysis of Traffic Congestion Caused by Rubbernecking at Freeway Accidents

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/259827547

Spatiotemporal Analysis of Traffic Congestion Caused by Rubbernecking at


Freeway Accidents

Article  in  IEEE Transactions on Intelligent Transportation Systems · September 2013


DOI: 10.1109/TITS.2013.2261987

CITATIONS READS
11 63

2 authors:

Younshik Chung Will Recker


Yeungnam University University of California, Irvine
46 PUBLICATIONS   504 CITATIONS    161 PUBLICATIONS   3,936 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Application of vehicle black box for traffic safety View project

Korea Traffic Information Center build up View project

All content following this page was uploaded by Younshik Chung on 19 February 2014.

The user has requested enhancement of the downloaded file.


1416 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 14, NO. 3, SEPTEMBER 2013

Spatiotemporal Analysis of Traffic Congestion


Caused by Rubbernecking at Freeway Accidents
Younshik Chung and Wilfred W. Recker

Abstract—In this paper, we present a well-specified analytical Center, Virginia Commonwealth University, Richmond, VA,
methodology for estimating capacity reduction that is attributable USA [11], the leading cause of vehicular accidents is, in fact,
to accidents in the opposite direction of accident—the condition rubbernecking. Additionally, rubbernecking caused by vehic-
whereby drivers in the opposite direction of an accident, by virtue
of their curiosity, tend to be distracted by the accident. The ular accidents and other incidents accounted for 16% of all
methodology is based on a binary integer programming formu- vehicular accidents, whereas the total number of outside-the-
lation that is used to identify the spatiotemporal region that is car distractions accounted for 35% [11].
affected by the influence of the accident. Thresholds measured Although secondary problems caused by rubbernecking are
against control sample readings from inductance loop detectors critical, virtually most research regarding delay by incidents
are used to determine the patterns and magnitudes of the delay.
A key feature of the methodology is its ability to separate non- has been focused on quantifying congestion that is occurring
recurrent delay from any recurrent delay that is present on the in the direction of the incident [12]–[19]. On the other hand,
road at the time and place of a reported accident, to estimate the few basic studies have tried to identify the negative impact
contribution of nonrecurrent delay caused by the specific accident. of rubbernecking. For instance, they include the determination
A case study that is based on historical inductance loop detec- of capacity reduction due to the incident [9], [20], [21], the
tor data from six major freeways in Orange County, California,
is presented. Potential factors contributing to delay, including explanation of traffic oscillations due to rubbernecking using
accident characteristics, geometric characteristics, environmental a behavioral car-following model [22], and the identification of
condition, traffic characteristics, and congestion characteristics, secondary incidents due to rubbernecking [2].
are analyzed for their effects by using the semiparametric Cox However, for the purpose of the successful operation of acci-
proportional-hazards model. dent management systems, knowing the basic factors affecting
Index Terms—Binary integer programming (BIP), censored total delay caused by rubbernecking is also crucial. Thus, the
data, Cox’s proportional-hazards (PH) model, freeway accident, objective of this paper is to develop the methodology for spatial
inductive loop detector (ILD), rubbernecking, traffic congestion. and temporal estimation of the congested region caused by
I. I NTRODUCTION rubbernecking at a traffic accident and to identify its major
causal factors. The congested region is identified by binary

A LTHOUGH there is no direct capacity reduction by lane


blockage in the opposite direction of accident, drivers in
the opposite direction of an accident, by virtue of their curiosity,
integer programming (BIP) and is based on basic accident
information and empirical statistics from traffic variables ob-
tained from inductive loop detectors (ILDs). Since a number of
tend to be distracted by the accident—this is commonly called estimated results were censored by time and/or space boundary
the “rubbernecking” or “gawking” phenomenon. Basically, conditions, general statistical approaches were not available.
rubbernecking/gawking refers to the tendency of drivers of An approach based on survival analysis was applied to analyze
vehicles in adjacent lanes or the opposite direction to slow estimated rubbernecking delay. Specifically, a statistical model
down as they pass by an incident to see what is happening [1]– based on the Cox-type proportional-hazard (PH) analysis is
[10]. Since such behavior invites sudden slowdowns in traffic, estimated, which describes rubbernecking delay as a function
it may result in additional congestion in the opposite direction. of day of week, time of day, weather, and observable (e.g.,
Moreover, during rubbernecking, the eyes of drivers may be from emergency calls and/or aerial or on-scene observation)
focused more on the accident scene than on the direction of their characteristics of the accident.
driving. This, in itself, can lead to another vehicular accident. Consequently, the results of this paper will be useful for the
According to a study by the Transportation Safety Training efficient operation of accident management systems and the
evaluation of its performance by quantifying accident conges-
Manuscript received July 24, 2012; revised January 21, 2013 and March 21, tion in terms of total delay to evaluate the benefit of accident
2013; accepted May 1, 2013. Date of publication May 15, 2013; date of management systems accrued from efficient traffic operations.
current version August 28, 2013. The Associate Editor for this paper was
W.-H. Lin.
In addition, they will provide a basis for simulation modeling
Y. Chung is with the Department of Transport Safety and Highway Research, of the rubbernecking phenomenon.
The Korea Transport Institute, Goyang 411-701, Korea (e-mail: tpgist@koti.
re.kr).
W. W. Recker is with the Department of Civil and Environmental Engi- II. P RELIMINARY A NALYSIS
neering, University of California, Irvine, CA 92697 USA (e-mail: wwrecker@
uci.edu). A. Section Definition
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. A freeway section in this study corresponds to a portion of
Digital Object Identifier 10.1109/TITS.2013.2261987 the freeway whose boundaries are defined by the midpoints

1524-9050 © 2013 IEEE


CHUNG AND RECKER: ANALYSIS OF TRAFFIC CONGESTION CAUSED BY RUBBERNECKING AT FREEWAY ACCIDENTS 1417

TABLE II
O BSERVED RUBBERNECKING S PEEDS

Fig. 1. Section definition and the corresponding detector location.

TABLE I
BASE C ASE OF D ISTRIBUTIONAL P ROPERTIES
FOR F REEWAY S ECTION S PEEDS

of two consecutive detector stations, as shown in Fig. 1. It is


assumed that the estimated speed at the ILD station is repre-
sentative of the speed for the corresponding section. Based on
the sections and their corresponding detector stations, estimated
speeds for each section are calculated for each 5-min interval
Fig. 2. Schematic congestion effect by rubbernecking at an accident.
during the one-year analysis period.
Relative to the display of information in Table II, we can
describe the negative effects (i.e., speed reduction) of rubber-
B. Speed Distribution
necking schematically, as shown in Fig. 2. The negative effect
For each section, for each day, for one year, and for tm in of rubbernecking will be propagated from the opposite section
5-min increments, speed sj (tm ) has been established; that is, of the accident to its upstream sections (i.e., to downstream
for every j and t, nominally 52 observations have been con- sections on the basis of the accident section). Such distinct dis-
structed. For example, section j on Monday from tm = 08:10 continuity between noncongested and congested flow is known
to 08:15 for 52 weeks is composed of 52 samples. Thus, the nth as a shockwave [23]. If the dot-shaded area that is affected
speed for any particular day of the week/time interval/section by the shockwave in Fig. 2 is identified, then the temporal
combination can be represented as sjn (tm ). and spatial impacts of rubbernecking will be also determined.
Let Ωjm = Ω(s̄j (tm ), σsj (tm ) ) denote the set of parameters The following section describes the method for distinguishing
defining the distribution of speeds sjn (tm ) corresponding to the the regions between noncongested and congested areas due to
accident-free case. Then, for tm > to , m = 1, 2, . . . (i.e., time rubbernecking at traffic accidents.
intervals after the accident that occurs in the opposite direction
of section i at to ), for all upstream sections that could have been
C. Quantifying Total Delay Caused by Rubbernecking
possibly affected by the accident, we can compose a matrix of
accident-free-case conditions (i.e., conditions in which there is To identify the spatiotemporal congested area caused by rub-
no accident) that can be expected to prevail, as described in bernecking, the methodology developed by Chung and Recker
Table I. [12] was applied. A key feature of their methodology is the
Similarly, the speed distribution in the opposite direction development of a method to separate nonrecurrent delay from
under each traffic accident can be described as in Table II. any recurrent delay that is present on the road at the time
For example, suppose that an accident occurred in the opposite and place of a reported accident, to estimate the contribution
direction of freeway section i at time t = to . Then, we can of nonrecurrent delay caused by the specific accident [12].
observe the corresponding measurements for the rubbernecking While the proposed methodology has been shown to work for
conditions, ŝj (tm ); j = i, i − 1, i − 2, . . .; m = 1, 2, . . .. We estimating the spatiotemporally congested region on freeway
can then compose a matrix of rubbernecking conditions as in sections on which an accident has occurred (i.e., traffic flow is
Table II. in the same direction as the accident), this is not the case for the
1418 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 14, NO. 3, SEPTEMBER 2013

where ŝj (tm ) is any particular speed not drawn from the
distribution of sjn (tm ), s̄j (tm ) and σsj (tm ) are the mean speed
and the standard deviation speed drawn from the distribution of
sjn (tm ), and α is a positive value. Additionally, nmin obs rep-
resents to set a threshold regarding the minimum number of ob-
servations to have some confidence in the statistical calculations
for mean and standard deviation. Since 30 is commonly used as
the minimum number of observations required for the law of
large numbers to apply, it was determined that nmin obs = 30.
Using this procedure, the problem of determining the “best”
set of dot-shaded (or yellow) cells can be formulated by using
BIP as follows:

Min Z = [Pjm · δjm + (1 − Pjm ) · (1 − δjm )]
δjm
∀j,m
s.t.
Fig. 3. Maximum set of freeway sections impacted by rubbernecking. δj+k,m ≤ [1 − (δj,m − δj+1,m )] · R, ∀j, m; ∀k ≤ J − j
δj,m+r ≤ [1 − (δj,m − δj,m+1 )] · R, ∀j, m; ∀r ≤ M − m
opposite direction. The objective of this study is to apply the ≤ [1 + (δj,m − δj+1,m )] · R, ∀j, m; ∀k ≤ M − m
δj,m+k 
methodology proposed by Chung and Recker [12] to the op- 0
posite direction of accident occurrence and, in turn, establish if δjm = (2)
1
the spatiotemporally congested region caused by rubbernecking
can be determined. The following steps highlight the procedure. where δ is the binary variable, R is a large number, J is the max-
1) Step 1—Determining the Maximum Extent of Rubber- imum number of upstream sections, and M is the maximum
necking Influence: The first step to identify the congested number of subinterval time periods that define the maximum
region due to rubbernecking is to estimate the maximum pos- duration assumed for congestion caused by the accident (e.g.,
sible extent of the shockwave by assuming the worst possible for 5-min subintervals and a maximum analysis time period of
conditions—total blockage for some prespecified time period. 4 h, M = 48).
Thus, for any given accident occurring in the opposite direction
of section i at time t1 , we compute the maximum number of up- III. C ASE S TUDY
stream sections that could be affected by the assumed persistent A. Data Description
total blockage at section i at time t1 . Using this spatiotemporal 1) Traffic Flow Data: This research uses one year (from
information, the “maximum area of interest” for any accident March 2001 to February 2002) of historical ILD data from six
occurring at the opposite section of section i at time t1 can major freeways in Orange County, California, i.e., Interstate
be schematically constructed. Based on this interpretation, the 405 (I-405), Interstate 5 (I-5), State Route 22 (SR-22), State
only data relevant to a rubbernecking phenomenon occurring at Route 55 (SR-55), State Route 57 (SR-57), and State Route 91
section i at time t1 are restricted to cells in the shaded (blue) (SR-91). The study area includes 499 mainline loop detector
area in Fig. 3. In Fig. 3, the cells in the dot-shaded (yellow) stations, and the average space between two consecutive de-
area represent speeds, i.e., ŝj (tm ), that have been reduced tector stations is about 0.8 mi. Since the size of the database
due to rubbernecking at the accident. Other shaded (“blue”) aggregated into 5-min intervals is over 52 000 000 records, a
cells represent speeds that are not significantly different from database management system is employed to efficiently manage
nonrubbernecking conditions. the data set, and its application program interface programs
2) Step 2—Determining the Congested Region: The second with C and C++ were employed in most analyses.
step is to distinguish between the shaded (blue) area and the
dot-shaded (yellow) area in Fig. 3. Since the speed of traf- B. Accident Data
fic in sections that are adversely affected by rubbernecking
Accident data were obtained from the Traffic Accident
will be reduced, the basic idea behind distinguishing between
Surveillance and Analysis System (TASAS) maintained by the
these two regions is to compare the rubbernecking speed, i.e.,
California Department of Transportation (Caltrans) for the six
ŝj (tm ), with the distribution of nonrubbernecking speeds, i.e.,
major Orange County freeways in 2001. Approximately 6200
sjn (tm ); n = 1, 2, . . . , nobs ; nobs ≤ T , and assign some level
accidents were included for the study period. The accident data
of confidence that any particular ŝj (tm ) was not drawn from the
include basic information related to accident time and location
distribution of sjn (tm ). Based on this idea, Chung and Recker
in terms of freeway milepost. In addition, data for each accident
[12] used the discriminant variable Pjm to identify affected
include three primary accident characteristics: 1) accident type,
versus nonaffected speed by rubbernecking as follows:
⎧ which is based on the type of collision (rear end, sideswipe, or
⎨ 0, ŝj (tm ) ≤ s̄j (tm )−α·σsj (tm ) ; nobs ≥ nmin obs hit object), the number of vehicles involved, and the movement
Pjm = 1, ŝj (tm ) > s̄j (tm )−α·σsj (tm ) ; nobs ≥ nmin obs of these vehicles prior to the accident; 2) accident location,
⎩ which is based on the location of the primary collision (e.g.,
0.5, nobs < nmin obs
(1) left lane, interior lanes, right lane, right-shoulder area, off-road
CHUNG AND RECKER: ANALYSIS OF TRAFFIC CONGESTION CAUSED BY RUBBERNECKING AT FREEWAY ACCIDENTS 1419

beyond the right-shoulder area); and 3) accident severity, in total delay for 2394 accidents (including censored results)
terms of injuries and fatalities per vehicle. was 2.87 vehicle hours, with the minimum total delay and
the maximum total delay being 0 and 1445.44 vehicle hours,
respectively.
C. Estimation of Total Delay Due to Rubbernecking
If each loop detector in the affected region is correctly
IV. C RITICAL FACTORS TO T OTAL D ELAY
working and reporting without error, the pattern of congestion
C AUSED BY RUBBERNECKING
resulting from an accident should be represented, as shown in
Fig. 3. However, in practice, it is often the case that some loop A. Definition of Candidate Variables
detectors may temporarily be not providing valid data, which
Candidate variables are classified into five groups based on
is usually due to a variety of reasons, including broken cables,
TASAS, traffic data, and estimated congestion regions: 1) ac-
interference from other electronic devices, communication fail-
cident characteristics, including accident type, accident causal
ure, and software error [24]. On the Orange County freeway
factor, truck accident, accident location, accident severity, num-
system, there are many such cases of missing data traced to
ber of vehicles involved, number of persons killed, number of
these reasons, leading to the delay of only 2394 of the 6182
persons injured, and accident time; 2) geometric characteristics
total accidents being successfully estimated.
referring to the opposite direction of the corresponding accident
However, certain results were censored by temporal con-
in terms of the number of lanes; 3) environmental condition
straints (or time boundary conditions). As suggested by Chung
(i.e., whether or not the road surface was wet); 4) traffic charac-
and Recker [12], an upper limit of 4 h after accident occurrence
teristics referring to the opposite direction of the corresponding
was applied in the determination of the spatiotemporal extent of
accident, such as annual average daily traffic (AADT), truck
the congestion region. However, in some cases, congestion was
AADT, and occupancy1 ; and 5) congestion characteristics re-
observed to remain after the maximum number of time period
ferring to the duration time, maximum congested time, and
M . Similarly, calculations for some of the rubbernecking delays
maximum congested length based on the estimated congestion
were cut off in terms of a spatial boundary condition, which is
region in the accident direction. The value of occupancy em-
due either to county lines or to an end of roadway (e.g., freeway
ployed is referenced to the accident section and its opposite
interchange). Another spatial cutoff result is due to detector
section during the 5-min period prior to the accident occurrence
problems (or missing data). Most of such results are caused by
time; this variable represents the mean value for the 5-min
accidents related to fatalities, hazardous materials, secondary
interval.
accidents that occurred either before the first accident was
Some of the nominal variables that can affect the total delay
cleared or while rubbernecking, etc., and they are called cen-
caused by an accident are classified into binary variables.
sored observations. In such cases, since the congestion caused
Specifically, the accident time variable (in terms of time of
by rubbernecking is not cleared, the causal factor analysis for
day) is divided into four time intervals, i.e., 06:01–09:00,
total delay would lead to an erroneous conclusion if censored
09:01–15:30, 15:31–18:00, and 18:01–06:00. The accident time
delays are ignored.
variables, including time of day and week, reflect the general
Having completed the given steps, which determine the
traffic pattern that is present. Table III shows the candidate
region (in time and space) that is negatively affected in the
variables.
opposite direction by any particular accident, we can calculate
the total delay (T D) caused by rubbernecking as
 B. Multivariate Analysis Using Cox’s PH Model
TD =
∀m,j∈dot−shaded cells
1) Estimation of Multivariate Cox Model: Since the result
   that the majority (26.9%) of 2394 observations for rubber-
1 1
× max Lj · − · Vjm , 0 (3) necking have zero values for delay leads to the rubbernecking
ŝj (tm ) s̄j (tm )
delay observations not nicely fitting such formal distributions
where as exponential, Weibull, gamma, lognormal, log-logistic, and
Lj Length of freeway segment j. Gompertz, multivariate effects on rubbernecking delay are an-
Vjm Volume (count) of vehicles in segment j during alyzed by using the semiparametric model (Cox’s PH model)
time m. rather than fully parametric survival models. The Cox model,
ŝj (tm ) Speed affected by rubbernecking in segment j at which assumes that the covariates multiplicatively shift the
time m. baseline hazard function, is by far the most popular choice, due
s̄j (tm ) Annual average speed in segment j at time m. to its elegance and computational feasibility [26]. It has a con-
There are many cases of missing data on the Orange County siderable advantage compared with the parametric approaches
freeway system. Thus, the total delay caused by rubbernecking in that it does not need an assumption about the baseline hazard
was successfully estimated for only 2394 (38.73%) of the 6182 function.
total accidents. Moreover, 432 (18.05%) of 2394 accidents
1 Initially, traffic volume was considered as a candidate variable. However,
resulted in censoring due to the space and time boundary
due to its property of having the same value under two different traffic situations
conditions and/or temporarily not providing valid data issues, (i.e., uncongested and congested conditions), using this variable may result in
as previously described. From the estimated results, the median a biased model. Thus, this variable was ignored.
1420 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 14, NO. 3, SEPTEMBER 2013

TABLE III TABLE IV


C ANDIDATE VARIABLES F ITTED C OX M ODEL FOR RUBBERNECKING D ELAY

Fig. 4. Graph on cumulative hazard of Cox–Snell residuals.

Cox model for rubbernecking delay, including seven variables


identified as being significant at 95% confidence level.
2) Assessment of Model Adequacy: The Cox–Snell residual
plot was used for assessing the overall goodness-of-fit of the
model [27]. If the Cox model fits the data well, then the
Cox–Snell residuals should have a standard exponential dis-
tribution with a hazard function that is equal to 1, and thus,
the cumulative hazard of the Cox–Snell residuals should be a
straight 45◦ line [26]. Since the plot in Fig. 4 is comparatively
close to the 45◦ line, it was concluded that the Cox model for
the total delay caused by rubbernecking at a traffic accident fits
the data fairly well.

C. Interpretation of the Cox Model


When interpreting a Cox model for each variable, a positive
Despite the semiparametric nature, methods for assessment coefficient (or a hazard ratio of > 1.0) for a variable means
of the fitted Cox model are essentially the same as for other that the hazard is higher. Conversely, a negative coefficient (or
regression models, i.e., similar to diagnostic assessments in a hazard ratio of < 1.0) implies a lower hazard for subjects with
the ordinary least squares model that check for model mis- higher values of that variable. In this paper, we should note that
specification, outliers, influential point, etc. Thus, three types the subject is rubbernecking delay, unlike the life of a patient
of assessments are applied for the Cox model: 1) testing the or machine. Thus, if the estimated hazard ratio is increased
assumption of PH; 2) identifying outliers/leverage points; and in comparison to another case, it implies that the rate of the
3) assessing the overall model fit. Table IV shows the fitted rubbernecking delay is decreased.
CHUNG AND RECKER: ANALYSIS OF TRAFFIC CONGESTION CAUSED BY RUBBERNECKING AT FREEWAY ACCIDENTS 1421

TABLE V Overall, it is anticipated that the results of this paper will


P ERCENTAGE C HANGE IN RUBBERNECKING
D ELAY FOR F ITTED C OX M ODEL be useful for the efficient operation of accident management
systems and the evaluation of its performance by quantifying
accident congestion in terms of total delay to evaluate the
benefit of accident management systems accrued from efficient
traffic operations. In particular, nonrecurrent congestion has
been estimated for the same direction of accident occurrence.
However, nonrecurrent congestion occurs not only in the acci-
dent direction but also in the opposite direction. Thus, traffic
congestion costs will be estimated within a reasonable value. In
addition, most studies on secondary accident identification were
conducted based on spatiotemporally predefined thresholds for
the negative impact on accidents. In addition, studies on sec-
ondary accidents due to rubbernecking were very limited. Since
For example, the hazard ratio of 5-min occupancy in the this study provides spatiotemporally congested regions with
opposite direction of the accident location is estimated as respect to each accident, the proposed method and its result
0.958. This would indicate that for every 1% increase in the can be used in the identification of secondary accidents caused
5-min occupancy in the opposite direction of the accident by rubbernecking. Finally, most microscopic traffic simulators
location, the rate of rubbernecking delay would increase by (e.g., CORSIM) use capacity reduction rates to capture the
4.2%. Conversely, the hazard ratio for accidents during night rubbernecking phenomenon due to traffic incidents; however,
time periods is estimated as 1.289, which would indicate that the reduction rates vary with incident types, occurrence time,
the rubbernecking delay for accidents during night time periods and so on. Thus, the results from this study will provide a basis
is about 28.9% less than for those during other time periods. for simulation modeling of the rubbernecking phenomenon. For
Table V shows the percentage change in rubbernecking delay instance, they can be used for the calibration of the capacity
due to one unit change (change from 0 to 1 in the case of the reduction rate in traffic simulation models.
dummy variable) of each variable for the model in Table V. Finally, although accident data used in this study did not in-
From the percentage changes in Table V, all of the re- clude accident duration time information and with the intuitive
sults were consistent with intuitive expectation. Particularly, sense that accident duration has a high impact on congestion
an increase in percent in the 5-min occupancy in the opposite caused by accidents and subsequent rubbernecking, this study
direction of accident, in the number of persons injured, and in applied the estimated duration time rather than the observed
1000 truck AADT resulted in greater rubbernecking delay. In time for the multivariate statistical analysis. However, based on
addition, accidents that are related to longer duration exhibited duration times observed in the field, a new model is recom-
greater rubbernecking delay. Moreover, rubbernecking delay mended for testing the significance of the variable of duration
was found to be greater with increasing maximum congested time, although it seems to be statistically significant to delay
time or length in the accident direction. Finally, accidents that caused by rubbernecking.
occurred during the night time period tended to have less
rubbernecking delay than those during the other time periods.
This result could be due to the fact that there is low traffic R EFERENCES
volume during the night time period.
[1] V. L. Knoop, H. J. Zuylen, and S. P. Hoogendoorn, Microscopic Traffic Be-
haviour Near Incidents Transportation and Traffic Theory 2009: Golden
Jubilee, W. H. K. Lam, S. C. Wong, and H. K. Lo, Eds. New York, NY,
V. C ONCLUSION AND F UTURE S TUDIES USA: Springer-Verlag, 2009, pp. 75–97.
[2] H. Zhang and A. Khattak, “What is the role of multiple secondary inci-
Although congestion due both to the direct effect of acci- dents in traffic operations?” J. Transp. Eng., vol. 136, no. 11, pp. 986–997,
dents and to its secondary problems caused by rubbernecking Nov. 2010.
[3] C. Sun, V. Chilukuri, T. Ryan, and M. Trueblood, Evaluation
is critical, virtually all research regarding delay caused by of Freeway Motorist Assist Program. Columbia, MO, USA: Univ.
incidents has been focused on quantifying congestion that is Missouri–Columbia, 2010.
only attributable to the direct effect in the lanes in the accident [4] A. Khattak, X. Wang, and H. Zhang, “Spatial analysis and modeling of
traffic incidents for proactive incident management and strategic plan-
direction. In this paper, nonrecurrent total delay caused by ning,” Transp. Res. Rec.: J. Transp. Res. Board, vol. 2178, pp. 128–137,
rubbernecking and its causal factors were statistically analyzed 2010.
by Cox’s PH approach. Based on the results, seven factors [5] C.-S. Chou, “Understanding the impact of incidents and incident manage-
ment programs on freeway mobility and safety,” Ph.D. dissertation, Dept.
were found to be statistically significant in contributing to the Civil Environ. Eng., Univ. Maryland, Baltimore, MD, USA, 2010.
delay encountered due to the principle of rubbernecking. These [6] R. R. Saddi, “Studying the impacts of primary incidents on freeways to
factors include the level of 5-min occupancy in the opposite identify secondary incidents,” M.S. thesis, Elect. Comput. Eng., Univ.
Nevada, Reno, NV, USA, 2009.
direction of accident, accident duration time, maximum con- [7] A. Khattak, X. Wang, and H. Zhang, “Are incident durations and sec-
gested time and length in the accident direction, number of ondary incidents interdependent?” Transp. Res. Rec.: J. Transp. Res.
persons injured due to the accident, and an accident time effect Board, vol. 2099, pp. 39–49, 2009.
[8] C.-S. Chou and E. Miller-Hooks, “Exploiting the capacity of managed
for the night period. In addition, the results were consistent with lanes in diverting traffic around an incident,” Transp. Res. Rec.: J. Transp.
intuitive expectation. Res. Board, vol. 2229, pp. 75–84, 2011.
1422 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 14, NO. 3, SEPTEMBER 2013

[9] V. Knoop, S. Hoogendoorn, and K. Adams, “Capacity reductions at in- [23] M. J. Lighthill and G. B. Whitham, “On kinematic waves. I. Flood move-
cidents sites on motorways,” Eur. J. Transp. Infrastruct. Res., vol. 9, ment in long rivers,” Proc. Roy. Soc. London. Ser. A, Math. Phys. Sci.,
pp. 363–379, Dec. 2009. vol. 229, pp. 281–316, May 10, 1955.
[10] V. Knoop, S. Hoogendoorn, and H. van Zuylen, “Capacity reduction at [24] S. Robinson and J. W. Polak, “Inductive loop detector data cleaning treat-
incidents: Empirical data collected from a helicopter,” Transp. Res. Rec.: ments and their effect on performance of urban link travel time models,”
J. Transp. Res. Board, vol. 2071, pp. 19–25, 2008. in Proc. 85th Annu. Meeting Transp. Res. Board, Washington, DC, USA,
[11] J. P. Masinick and B. L. Smith, An Analysis on the Impact of Rubberneck- 2006, [CD-ROM].
ing on Urban Freeway Traffic. Charlottesville, VA, USA: Univ. Virginia, [25] Y. Chung and B.-J. Yoon, “Analytical method to estimate accident dura-
Aug. 2004, No. UVACTS-15-0-62. tion using archived speed profile and its statistical analysis,” KSCE J. Civil
[12] Y. Chung and W. W. Recker, “A methodological approach for estimating Eng., vol. 16, no. 6, pp. 1064–1070, Sep. 2012.
temporal and spatial extent of delays caused by freeway accidents,” IEEE [26] M. A. Cleves, W. W. Gould, and R. G. Gutierrez, An Introduction to
Trans. Intell. Transp. Syst., vol. 13, no. 3, pp. 1454–1461, Sep. 2012. Survival Analysis Using Stata, 3rd ed. College Station, TX, USA: Stata
[13] Y. Chung, “Quantification of nonrecurrent congestion delay caused by Press, 2004.
freeway accidents and analysis of causal factors,” Transp. Res. Rec.: J. [27] D. R. Cox and E. J. Snell, “A general definition of residuals,” J. Roy. Stat.
Transp. Res. Board, vol. 2229, pp. 8–18, 2011. Soc. Ser. B: Methodol., vol. 30, no. 2, pp. 248–275, 1968.
[14] Managing Urban Traffic Congestion. Paris, France:OECD, 2007.
[15] R. Dowling, A. Skabardonis, M. Carroll, and Z. Wang, “Methodology
for measuring recurrent and nonrecurrent traffic congestion,” Transp. Res. Younshik Chung received the M.S. degree from
Rec.: J. Transp. Res. Board, vol. 1867, pp. 60–68, 2004. Ajou University, Suwon, Korea, in 1999 and the
[16] A. Skabardonis, P. Varaiya, and K. Petty, “Measuring recurrent and non- Ph.D. degree in civil and environmental engineering
recurrent traffic congestion,” Transp. Res. Rec.: J. Transp. Res. Board, from the University of California, Irvine, CA, USA,
vol. 1856, pp. 118–124, 2003. in 2007.
[17] Y. Chung, H. Kim, and M. Park, “Quantifying non-recurrent traffic con- Since 2007, he has been a Research Fellow with
gestion caused by freeway work zones using archived work zone and ITS The Korea Transport Institute, Goyang, Korea. His
traffic data,” Transportmetrica, vol. 8, no. 4, pp. 307–320, 2012. current research interests include traffic safety, in-
[18] Y. Chung, “Assessment of non-recurrent congestion caused by precipita- telligent transport systems, traffic control and oper-
tion using archived weather and traffic flow data,” Transp. Policy, vol. 19, ation, and sustainable transportation.
no. 1, pp. 167–173, Jan. 2012.
[19] Y. Chung, “Assessment of non-recurrent traffic congestion caused by
freeway work zones and its statistical analysis with unobserved hetero-
geneity,” Transp. Policy, vol. 18, no. 4, pp. 587–594, Aug. 2011. Wilfred W. Recker received the B.S., M.S., and
[20] C. Chou, E. Miller-Hooks, and I. Promisel, “Benefit-cost analysis of free- Ph.D. degrees in civil engineering from Carnegie
way service patrol programs: Methodology and case study,” Adv. Transp. Mellon University, Pittsburgh, PA, USA, in 1964,
Stud. Int. J. Sec. B, vol. 20, pp. 81–96, 2010. 1966, and 1968, respectively.
[21] K. Heaslip, A. Kondyli, D. Arguea, L. Elefteriadou, and F. Sullivan, He is a Professor with the Department of Civil
“Estimation of freeway work zone capacity through simulation and field and Environmental Engineering, University of Cal-
data,” Transp. Res. Rec.: J. Transp. Res. Board, vol. 2130, pp. 16–24, ifornia, Irvine, CA, USA. From 1983 to 2006, he
2009. was the Director of the Institute of Transportation
[22] D. Chen, J. Laval, Z. Zheng, and S. Ahn, “A behavioral car-following Studies, University of California. His research inter-
model that captures traffic oscillations,” Transp. Res. B: Methodol., ests include activity-based travel demand modeling
vol. 46, no. 6, pp. 744–761, Jul. 2012. methodologies and transportation system analysis.

View publication stats

You might also like