Accident Analysis and Prevention 38 (2006) 295–301

A method for identifying rear-end collision risks using

inductive loop detectors
Cheol Oh a,∗ , Seri Park b,1 , Stephen G. Ritchie b,2
a Center for Advanced Transportation Technology, Korea Transport Institute, 2311 Daewha-dong,
Ilsan-gu, Koyang-shi, Kyunggi-do 411-701, Republic of Korea
b Department of Civil and Environmental Engineering, Institute of Transportation Studies,

University of California, Irvine, CA 92697-3600, USA

Received 7 June 2005

Abstract
An innovative feature of this study is to firstly attempt to capture rear-end collision potentials from the analysis of inductive loop detector
data. Signals collected from loops are applied for monitoring individual vehicle information on freeways to estimate safe stopping distances in
car-following situations. An index to quantify the potential of rear-end collisions is derived, and further employed for developing criteria to evaluate
levels of rear-end collision risks. The proposed methodology based on loop detector data enables to identify collision potentials in real time. It is
believed that the index would be a valuable tool for operating agencies in developing various strategies and policies toward enhancements of traffic
safety.
© 2005 Elsevier Ltd. All rights reserved.

Keywords: Rear-end collision; Risk analysis; Loop detectors

1. Introduction observable situation in which two or more road users approach

A first step for tackling complex transportation problems is
to identify and evaluate the current status of dynamic traffic con-
ditions. Measuring traffic performance can be conducted from
two different perspectives such as ‘mobility’ and ‘safety’. A vari-
ety of studies have focused on assessing transportation systems
based on mobility by utilizing travel time, speed, and other fun-
damental traffic parameters. Such mobility-based evaluations
support advanced traffic control strategies including route guid-
ance, ramp metering, and signal control, etc. However, less
efforts have been devoted to develop methodologies for real-
time safety evaluation, although research subjects on safety are
often regarded as more important ones because of direct impacts
on human lives.
One of the promising methods for evaluating traffic safety
is the traffic conflict technique. A conflict can be defined as an
∗ Corresponding author. Tel.: +82 31 910 3097; fax: +82 31 910 3228.
E-mail addresses: cheolo@koti.re.kr (C. Oh), serio@uci.edu (S. Park),
sritchie@uci.edu (S.G. Ritchie).
1 Tel.: +1 949 824 5623; fax: +1 949 824 8385.
2 Tel.: +1 949 824 4214; fax: +1 949 824 8385.
each other in time and space to such an extent that there is a risk of
collision if their movements remain unchanged (Leur and Sayed,
2002). Various studies related to traffic conflict techniques can
be found in the literature. Existing studies are classified into
two groups. The first group deals with the traffic conflict for
off-line decision making and safety policy derivation at the plan-
ning level (Hyden, 1987; Fazio et al., 1993; Ha and Berg, 1995;
Tarrall and Dixon, 1998; Lee and Berg, 1998). Another approach
includes analyses of traffic conflicts for both identifying safety
indicators and assessing traffic safety at the microscopic level,
which focused on the analysis of time to collision (TTC) (Heijer
et al., 1997; Minderhoud and Bovy, 2001; Vogel, 2003; Kiefer et
al., 2005). TTC has been considered as an effective indicator to
evaluate traffic safety in the field of the traffic conflict. However,
we are not aware of any study analyzing rear-end collision risks
using the ubiquitous traffic surveillance system, inductive loop
detectors, in the context of the traffic conflict.
Two traffic conflicts can be observed on freeways. One is
a lane change conflict occurring when a vehicle attempts to
change lanes. The rest one is rear-end conflict resulting poten-
tially in rear-end crash. Rear-end collisions have been identified
in the literature as one of the main crash types on freeways.

0001-4575/$ – see front matter © 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.aap.2005.09.009
296 C. Oh et al. / Accident Analysis and Prevention 38 (2006) 295–301
For example, Golob et al. (2002) have analyzed freeway acci-
dent data in Southern California and found rear-end collisions
occupied about 31.2% of total accidents (sideswipe: 43.0%, hit
object: 11.7%, others: 14.1%). From this perspective, identifying
rear-end conflicts which can lead to rear-end collisions would be
the first step toward devising viable solutions to safety related
problems in transportation networks.
This study focuses on developing method to capture rear-end
collision potentials. Signals collected from loops are utilized in
a smart way for monitoring individual vehicle information on
freeways to estimate safe stopping distances in car-following
situations. A risk index to quantify the potential of rear-end col-
lisions is derived, and further employed for developing criteria
to evaluate levels of rear-end collision risks. A fuzzy c-means
algorithm is applied to categorize the index, which can be used in
identifying hazardous traffic conditions in a more systematical
way. The outcomes of this study that measure and quantify the
potential of rear-end collisions in real time could be an invaluable
tool for traffic operators in evaluating traffic systems in terms of
safety. Furthermore, the approach can support the development
of effective countermeasures to prevent rear-end traffic accidents
on freeways. It should be noted that the purpose of this study
is to develop a methodology to quantitatively evaluate relative
hazard traffic conditions.
This paper consists of four sections, including this introduc-
tion. The proposed methodology for evaluating freeway safety is
presented in Section 2. Section 3 describes the data used in this
study, and presents both results and their application. The final
section is dedicated to describing findings and further studies.
2. Proposed methodology
To develop a methodology for identifying collision poten-
tials on freeways, this study uses individual vehicle information
and traffic conflict techniques. Advanced inductive loop detector
technology that produces unique inductive vehicle signatures for
each vehicle passing over a loop detector station is employed to
monitor detailed individual vehicle information. Use of induc-
tive vehicle signatures also allows us to classify vehicle types
in real time based on the analysis of vehicle features extracted
from the inductive vehicle signatures. This vehicle classification
information is highly valuable for identifying unsafe car follow-
ing situations since the performance of large vehicles and their
effects on neighboring vehicles are obviously different from
relatively small vehicles. Large vehicles involved in accidents
usually show higher severity of injury due to the operating char-
acteristics of large vehicles, such as long vehicle length, heavy
weight, and longer stopping distance. In addition, large vehicles
limit the capability of following vehicles to identify traffic con-
ditions ahead, which could increase the possibility of rear-end
collisions.
A new risk index based on rear-end conflicts derived from
advanced inductive loop detector technology is utilized to eval-
uate traffic safety on a freeway in real time. The basis of the
proposed index is obtained from the computation of the safety
distance defined by the difference between a leading and follow-
ing vehicles’ safe stopping distances. The safe stopping distance
Fig. 1. Research framework.
of a leading vehicle should be always greater than that of a
following vehicle to prevent an accident. In order to use the
proposed risk index more effectively, a clustering technique
was further applied to categorize the index resulting in various
levels of collision risks. Fig. 1 shows the framework for identi-
fying rear-end collision risks on freeways based on the proposed
methodology. More detailed descriptions for each element are
presented in the following subsections.
2.1. Monitoring individual vehicle information from
inductive loop detectors
The proposed methodology utilizes inductive loop detectors
(ILDs) that are the most dominant traffic surveillance infrastruc-
ture in practice currently. Traditional loop detectors are able to
provide fundamental traffic parameters such as volume, occu-
pancy, and speed. Conventional ILDs are usually in bivalent in
nature, when the detector card output is either “0” or “1”. Here
“1” denotes the presence of a vehicle passing over the detec-
tor. The aggregated outputs are obtained over a pre-determined
aggregation interval, for example 20 or 30 s. With the advance in
detector technology, it is possible to obtain disaggregated infor-
mation for individual vehicles by using advanced high speed
scanning loop detector cards that capture more detailed induc-
tance changes over the loops, vehicle signature.
A variety of useful information including vehicle length, sym-
metry, and clearance between the bottom of vehicle and the
ground, can be extracted from the inductive vehicle signatures.
Those represent the unique physical characteristics of individual
vehicles. In addition, other information including arrival time,
duration, lane, and speed for individual vehicles can be collected.
The disaggregated information of each vehicle was utilized to
classify vehicle types, and to derive the safety index for eval-
uating real-time safety performance. Fig. 2 shows examples of
inductive vehicle signatures.
Maximum magnitude, electronic vehicle length, shape
parameter (SP), degree of symmetry (DOS), and number of
samples of high magnitudes (NHM) are examples of vehicle fea-
tures derived from inductive vehicle signatures. The scheme of
 C. Oh et al. / Accident Analysis and Prevention 38 (2006) 295–301 297

Fig. 2. Inductive vehicle signatures for different vehicle types.

processing vehicle signature and detailed descriptions on vehicle
features are presented in Table 1.
2.2. Vehicle classification
The proposed risk index that will be introduced in the next
section is based on safety distances of the lead and following
vehicles in each car-following situation. Vehicle performance is
one of the major factors in computing safety distance. There-
fore, identifying vehicle types enables to apply various vehicle
deceleration performances, resulting in more accurate safety dis-
tances. The vehicle classification capability also supports the
evaluation of real-time traffic safety in terms of indicating acci-
dent severity in unsafe car-following situations, because large
vehicle-involved accidents show more severe damage, as found
in many previous studies (Jackson, 1986; Golob et al., 1987;
Duncan et al., 1998; Chang and Mannering, 1999; Kockelman
and Kweon, 2002). In addition, vehicle classification informa-
tion can be also used in traffic flow modeling and simulation as
well as in highway maintenance.
Vehicle features obtained by processing raw signatures can be
utilized for identifying vehicle classes since the vehicle signature
is a unique function of each vehicle type. In this study, vehi-
cle classification consists of two parts: feature extraction from

Table 1
Vehicle signature and feature vector
Vehicle signature processing

Feature vector Feature description

Maximum magnitude Maximum absolute magnitude value (a)

Electronic vehicle length d
Shape parameter (SP) b/(b + c)
Degree of symmetry (DOS) Sum of the distances from median (e), to each point that is above “0.5” y-value
Number of high magnitude (NHM) Number of samples above “0.5” y-value after x, y normalization
298 C. Oh et al. / Accident Analysis and Prevention 38 (2006) 295–301
the inductive vehicle signatures and development of a heuristic
discriminant algorithm. Vehicle feature vectors obtained from
inductive vehicle signatures were applied as an inputs of heuris-
tic discriminant algorithm.
2.3. Derivation of a new risk index
Rear-end collisions have been identified in the literature as
one of the main crash types on freeways. For example, Golob
et al. (2002) have analyzed freeway accident data in South-
ern California and found rear-end collisions are over 30% of
total accidents. From this perspective, identifying rear-end con-
flicts potentially leading to rear-end collisions is very useful for
evaluating safety. Since a rear-end collision is primarily caused
by monitoring insufficient safety distance between the leading
vehicle and the following vehicle is not maintained, a method
to obtain safety distance from a freeway real-time monitoring
system is of keen interest for establishing safety performance
measures.
One of the nicest features of the monitoring system of this
study is to compute vehicles’ safe stopping distances. A pro-
posed index is based on the concept of ‘safe stopping distance’,
which can be defined as the minimum distance that the follow-
ing vehicle can safely reduce speed and avoid colliding with the
leading vehicle when the leading vehicle reduces speed with the
maximum deceleration rate and stops. Therefore, the stopping
distance of the leading vehicle should be larger than that of the
following vehicle for each car-following situation to avoid rear-
end collision. Vehicle classification information enables us to
apply different deceleration rates for each vehicle type (aL/F dec )
to estimate more reliable safe stopping distance. The stopping
distance index is derived as follows:
dL > dF
VL2 VF2
⇒ VL h + + l L > V F t R +
2aLdec 2aFdec
  2 2
 observation in fuzzy clustering is given with fractional member-
 0 (safe) if V h − V t + l + VL − VF >0
L F R L ship in multiple clusters. Therefore, the degree of membership
∴ SDI = 2aLdec 2aFdec
 can be used for quantifying reliability for uncertainty arising
1 (unsafe) otherwise
where SDI is the stopping distance index for each car-following
event, dL the safe stopping distance for leading vehicle, dF the
safe stopping distance for following vehicle, lL the length of
leading vehicle, VL the speed of leading vehicle, VF the speed of
following vehicle, tR the brake reaction time, h the time headway,
aLdec the deceleration on rate of leading vehicle and aFdec is the
deceleration on rate of following vehicle.
As a result, the number of stopping distance indices (SDIs)
less than zero, observed over a given time period, is used to derive
a rear-end collision risk index (RCRI). In general, the index rep-
resenting the risk level is described in the form of rates, which is
a function of accident or conflict counts and an exposure mea-
sure. It should be noted that the exposure measure needs to be
employed to establish the risk index as the base because provid-
ing total number of accidents or conflicts would be misleading.
This study employs the maximum possible number of car-
following situations over a certain time interval as an exposure.
The proposed RCRI, which is a surrogate measure for rear-end
collision, is defined as the ratio of the number of unsafe car-
following events showing a safety distance index less than zero
to the maximum possible car-following events on freeway detec-
tor stations. An unsafe car-following event is derived from the
computation of safe stopping distance presented above, and a
maximum car-following event can be derived by the minimum
headway that results in freeway capacity. The proposed RCRI
is described as follows:

number of rear-end conflicts SDIi
RCRI = = max i
exposure Ncar (T/3600)Nl
max the maximum
where RCRI is the rear-end collision index, Ncar
number of car-following events per hour (derived from freeway
capacity), Nl the number of freeway lanes, and T is the analysis
period (s).
2.4. Development of criteria for rear-end collision risks
For more comprehensive application of the RCRI, the evalua-
tion criteria of the level of rear-end collision risks are developed.
The proposed RCRI is categorized by the threshold values to be
used by users including operating agencies and drivers. The basic
background for categorization is that RCRIs within the same cat-
egory represent similar traffic conditions as much as possible in
terms of traffic safety. On the other hand, RCRIs in different
categories should also represent dissimilar traffic conditions in
terms of traffic safety. A solution that can satisfy both aspects
is obtained by the formulation of two maximization problems:
first to maximize dissimilarity between categories, and second
to maximize similarity within categories. Along with the above
concept a clustering algorithm is applied to determine the “opti-
mal” number of categories in a given data set. A fuzzy c-means
(FCM) algorithm was employed to cluster traffic conditions.
Unlike conventional crisp clustering, such as K-means where
each object of the data set is assigned to exactly one cluster, each
in evaluating real-time safety performance measures. Once the
clustering analysis is completed, threshold values categorizing
RCRIs are determined by averaging minimum and maximum
RCRIs of each adjacent category.
FCM was originally proposed by Bezdek (1981) based on
fuzzy factors. fuzzy clustering problems can be formulated as
the following minimization problem:

n 
c
min Jm (U, V ) = um 2
ij dij , dij = xj − vi 
j=1 i=1

c
s.t. uij = 1, 1 ≤ j ≤ n uij ≥ 0, 1 ≤ i ≤ c, 1 ≤ j ≤ n,
i=1

n
uij > 0, 1≤i≤c
j=1
 C. Oh et al. / Accident Analysis and Prevention 38 (2006) 295–301 299

where x is the sample, v the clustering center, n the number of

samples, c the number of clusters (1 < c < n), m the fuzzy factor,
dij the distance between xj and vi and U the matrix of c × n, and
V is the matrix of c × n.
The minimization problem for the fuzzy clustering can be
solved according to the following procedures.

• Step 1. Initialization, V (0)

Let k = 0; select ε
• Step 2. Calculation of U(k)
◦ If ∀j, r, drj (k) > 0, then
1
uij = c 2/m−1
(d
r=1 ij (k)/d rj (k))
◦ If there exist j, r, such that drj (k) = 0, then urj (k) = 1 and
uij (k) = 0 for i = r
• Step 3. Calculation of V(k+1) Fig. 3. I-405 detector testbed at Laguna Canyon.
n m
j=1 uij (k)xj
Vi (k + 1) = n m
j=1 uij (k) to 10:00 a.m. and from 15:00 to 20:00 p.m., were used. A total of
24 h of data, excluding time periods that were not available due
• Step 4. Stopping criteria to communication problems in the field, resulted in 136,817 car-
stop if ||V (k) − V (k+1) || < ε following events, which were used to derive RCRIs. Descriptive
Otherwise, go to step 2. statistics on RCRI are presented in Table 2. It is identified that
the average of RCRI of given data set is 0.353. This value rep-
3. Results and application resents that about 35% of the car-following events passing the
loop detector station would be identified as unsafe conditions.
This section presents the description of data used for the
derivation of RSIs. Vehicle classification results from the devel- 3.2. Vehicle classification
oped heuristic discriminant algorithm and clustering results by
FCM are also presented with the application of the proposed 2486 vehicles collected on the freeway were used to develop
methodology. the heuristic discriminant algorithm for vehicle classification.
Ground-truthed data based on the visual inspection were pre-
3.1. Data pared to investigate the performance of the algorithm. Fig. 4
shows the developed classification algorithm providing promis-
The data used in this study were collected from the traf- ing results with an 82.02% correct classification rate. More
fic detection and surveillance sub-testbed (TDS2 ) on the I-405 detailed classification results based on vehicle types are pre-
northbound freeway in Irvine, California. The 0.7 mile free- sented in Table 3.
way section contains two contiguous detection stations equipped
with double inductive loop detectors. Vehicle signature data are
extracted in real-time using high-speed scanning loop detector 3.3. Clustering results
cards and stored in dedicated computers in the field. In addition,
overhead vertical-mount video cameras were installed over each RCRIs obtained over 5 min aggregation intervals were used
lane of traffic and were connected to a ground-truthing video as inputs for the FCM clustering algorithm. One major issue
image processing system. In addition, both radar detectors and arising from clustering analysis is to find appropriate cluster size.
acoustic detectors were installed. Fig. 3 shows the configuration To determine the optimum number of groups, Wilk’s lambda (Λ)
of TDS2 on the I-405 northbound freeway. defined as the ratio of within-groups variance to total variance
Communications between the upstream and downstream sites is used. In general, the optimal cluster size is achieved when the
is by dedicated high-speed wireless Ethernet, with fiber optic marginal Wilk’s lambda ratio change is minimal (Johnson and
cable between the downstream mainline cabinets and the ramp
cabinet. Wireless Ethernet is also used to communicate from the
downstream site to a city of Irvine cabinet where data enters the Table 2
Statistics on RCRI
city network and is sent to the University of California, Irvine
(UCI) testbed labs and UCI data and web server, as well as the Minimum 0.087
internet. Maximum 0.664
Average 0.353
Inductive vehicle signature data from two time periods
Standard deviation 0.096
including peak hours during 4 days in October 2002, from 06:00
300 C. Oh et al. / Accident Analysis and Prevention 38 (2006) 295–301

Table 4
Rear-end collision risk criteria on freeway
Category Collision risk criteria

A ≤0.251
B >0.251 and ≤0.306
C >0.306 and ≤0.355
D >0.355 and ≤0.416
E >0.416 and ≤0.510
F >0.510

be stated in terms of RCRI averaged over the 5 min interval as

Fig. 4. Heuristic vehicle classification algorithm.
Wichern, 1992).
|W| |W|
Λ= =
|T | |B + W|
where W is the pooled within-groups variance, B the between
groups variance, and T is the total variance
As we can identify in the clustering results are shown in
Fig. 5, significant marginal improvement due to a decreasing
value of Wilk’s lambda is not observed after six groups, which
means the most appropriate number of clusters corresponding to
collision risk criteria for the data used is six. Based on the results,
real-time rear-end collision risk criteria for freeway safety can
shown in Table 4.
Category A is considered to be very safe traffic conditions
with low RCRI up to 0.251. Even if a traffic accident occurs
under this level, the possibility of subsequent accident occur-
rences by not maintaining safety distances would be expected
to be much lower than any other categories. On the other hand,
category F would describe traffic safety on freeway with RCRI
in excess of 0.510. Traffic conditions under this level can be
described as very unsafe leading to the highest possibility of
subsequent accident occurrences.
3.4. Application
To illustrate the application of the results in Table 3, RCRI
was computed over the 5 min aggregation interval using the
actual data. In an on-line situation, the RCRI is calculated for
each given time interval, and is assigned to an appropriate risk
category from Table 3. Fig. 6 illustrates the proposed procedure.

Fig. 5. FCM clustering results. Fig. 6. Rear-end collision risk analysis on freeway.

Table 3
Vehicle classification result
Group Vehicle type Total vehicle number Correct classified number Correct classification rate (%)

1 Motorcycle 4 4 100
2 Passenger car/mini van 1370 1126 82.19
3 Sport utility vehicle 474 327 68.99
4 Van/small pickup truck 504 449 89.09
5 Truck/bus 80 79 98.75
6 Trailer 54 54 100
2486 2039 82.02
 C. Oh et al. / Accident Analysis and Prevention 38 (2006) 295–301
In real-time implementation, a field computer to collect
inductive vehicle signatures will derive RSI and determine cor-
responding collision risks. The risk levels determined for each
given time period will be disseminated in real-time to the oper-
ating agency’s traffic management center, either for direct safety
evaluation, or as input to other software applications for devel-
oping countermeasures to prevent accident occurrence.
4. Conclusion
The most innovative feature of this study was to identify
hazardous traffic conditions based on a surrogate measure of
rear-end collision risks. This study proposed a new methodology
incorporating a traffic conflict technique and an advanced ILD-
based surveillance system. The proposed surveillance system
utilizing inductive vehicle signatures enables to obtain indi-
vidual vehicle information with comparatively lower costs by
using the existing loop infrastructure. Individual vehicle infor-
mation such as vehicle arrival time, speed, length, and some other
valuable features were used for deriving the proposed rear-end
collision risk index (RCRI) based on the safety distance in the
car-following situation, which is potentially related with free-
way rear-end traffic collisions.
In this study, fuzzy-clustering algorithm was employed to cat-
egorize RCRIs in order to effectively identify the collision risks.
Clustering results showed that six categories were most appro-
priate to establish collision risk criteria. It is believed that the
proposed methodology for identifying collision risks on free-
ways will be a valuable tool operating agencies in support of
countermeasure development to prevent accident occurrence.
Although the proposed methodology to evaluate traffic safety
on freeways is preliminary, the findings and insights are expected
to initiate an emerging research topic on traffic safety evaluation
in real time. Due to the use of point measures by ILDs in this
study, following assumptions could not be avoided. The leading
vehicle would continue to travel at constant speed (VL ) for a
time equal to time headway (h) after the vehicle passes the loop
detector station. In addition, the following vehicle would travel
at constant speed (VF ) for a time to equal the reaction time (tR )
after passing the loop detector stations.
Further studies need to be performed to fertilize this study.
First of all, the investigation of actual accident data with the
consideration of the proposed risk index should be performed to
validate if the index is truly correlated with accident occurrences.
The proposed risk index can also be applied as an indicator
for measuring traffic safety in developing a new traffic control
strategy to maintain safe traffic conditions. Also, the index can
be further processed in supporting advanced warning informa-
tion systems leading to careful driving, although more extensive
301
analyses are required. It is expected that the proposed method
would be an invaluable tool for evaluating traffic safety in real
time.
References
Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective Function Algo-
rithms. Plenun, New York.
Chang, L.-Y., Mannering, F., 1999. Analysis of injury severity and vehicle
occupancy in truck- and non-truck-involved accidents. Accid. Anal. Prev.
31, 579–592.
Duncan, C.S., Khattak, A.J., Council, F.M., 1998. Applying the ordered pro-
bit model to injury severity in truck–passenger car rear-end collisions.
In: Transportation Research Record No. 1635. TRB, National Research
Council, Washington, DC, pp. 63–71.
Fazio, J., Holden, J., Rouphail, N.M., 1993. Use of freeway conflict rates as
an alternative to crash rates in weaving section safety analyses. In: Trans-
portation Research Record No. 1401. TRB, National Research Council,
Washington, DC, pp. 61–69.
Golob, T.F., Recker, W., Leonard, J., 1987. An analysis of the severity and
incident duration of truck-involved freeway accidents. Accid. Anal. Prev.
19, 375–395.
Golob, T., et al., 2002. Freeway safety as a function of traffic flow: the FITS
tool for evaluating ATMS operations. TO 4117 Final Report to California
PATH, Institute of Transportation Studies, University of California, Irvine.
Ha, T.-J., Berg, W.D., 1995. Development of safety-based level-of-service
criteria for isolated signalized intersections. In: Transportation Research
Record No. 1484. TRB, National Research Council, Washington, DC, pp.
98–104.
Heijer, T., Wiersma, E., Polak, P.H., 1997. Proposed safety criteria for traffic
control on motorways. Safety Sci. 26 (1/2), 141–154.
Hyden, C., 1987. The Development of a Method for Traffic Safety Evaluation.
Lund Institute of Technology.
Jackson, L., 1986. Truck accident studies. In: Transportation Research
Record No. 1052. TRB, National Research Council, Washington, DC,
pp. 137–145.
Johnson, R.A., Wichern, D.W., 1992. Applied Multivariate Statistical Analy-
sis. Prentice-Hall, Englewood Cliffs, NJ.
Kiefer, R., LeBlanc, D., Flannagan, C.A., 2005. Developing an inverse time-
to-collision crash alert timing approach based on drivers’ last-second
braking and steering judgments. Accid. Anal. Prev. 37, 295–303.
Kockelman, K.M., Kweon, Y.-J., 2002. Driver injury severity: an application
of ordered probit models. Accid. Anal. Prev. 34, 313–321.
Lee, S.-B., Berg, W.D., 1998. Development of safety-based level-of-service
parameters for two-way stop-controlled intersections. In: Transportation
Research Record No. 1635. TRB, National Research Council, Washing-
ton, DC, pp. 127–132.
Leur, D.P., Sayed, T., 2002. Development of a road safety risk index. In:
Transportation Research Record No. 1784. TRB, National Research Coun-
cil, Washington, DC, pp. 33–42.
Minderhoud, M.M., Bovy, P.H.L., 2001. Bovy extended time-to-collision
measures for road traffic safety. Accid. Anal. Prev. 33, 89–97.
Tarrall, M.B., Dixon, K., 1998. Conflict analysis of double left-turn lanes
with protected-plus-permitted signal phases. In: Transportation Research
Record No. 1635. TRB, National Research Council, Washington, DC, pp.
105–112.
Vogel, K., 2003. A comparison of headway and time to collision as safety
indicators. Accid. Anal. Prev. 35, 427–433.