You are on page 1of 17

Transportation Research Part F 9 (2006) 155–171

www.elsevier.com/locate/trf

q
Intersection approach speed and accident probability
a,*
A.C.E. Spek , P.A. Wieringa b, W.H. Janssen c

a
Netherlands Forensic Institute, Department of Justice, The Netherlands
b
Delft University of Technology, The Netherlands
c
TNO Human Factors, The Netherlands

Received 1 September 2005; accepted 12 October 2005

Abstract

A statistical model is derived for gap acceptance at intersections, taking into account limitations of human perception.
The model assumes the logarithm of the odds for acceptance of a gap as a linear function of the logarithms of gap time and
the speed of the vehicle closing the gap. Fitting is performed against gap acceptance decisions as observed in real traffic at
12 unsignalized intersections by Brilon and Weinert (2001) [Ermittlung aktueller grenz- und folgezeitlücken für
auerortsknoten ohne lichtsignalanlagen. Straßenbau und Straßenverkehrstechnik, vol. 828.] and in a driving simulator
experiment by Hancock and Caird (1993) [Factors affecting older drivers left turn decisions. Technical Report, Transpor-
tation Research Board, National Research Council, Washington, DC]. The model fit was reasonable for the traffic obser-
vations (R2 = 0.76) and very good for the simulator data (R2 = 0.97). Parameter estimation yields that the speed has a
significant effect on gap acceptance behavior. Drivers crossing a priority stream tend to accept shorter time gaps as the
speed of the approaching vehicle increases. This effect is stronger for older drivers than for younger drivers. It is proposed
that this model may be extrapolated to quantify the relation between approach speed and accident probability. The val-
idation that this kind of extrapolation would require, has not yet been performed.
Ó 2005 Elsevier Ltd. All rights reserved.

Keywords: Accident probability; Accident reconstruction; Conflict rate; Crossing behavior; Gap acceptance; Intersection accidents;
Speeding

1. Introduction

The objective of this study is to present a model that allows for estimating the likeliness of an intersection
accident as a function of intersection approach speed. The motivation for this research is twofold. Firstly, the
Netherlands Forensic Institute considers the function essential for accident speed reconstruction. Currently, a
Monte Carlo approach is used, that approximates a probability density function (PDF) for the speed given the

q
The study was performed as a master graduation project at Delft University, sponsored by TNO Human Factors and The Netherlands
Forensic Institute.
*
Corresponding author. Tel.: +31708886491.
E-mail addresses: a.spek@nfi.minjus.nl, aart.spek@tiscali.nl (A.C.E. Spek).

1369-8478/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trf.2005.10.001
156 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171

evidence found on the scene, such as vehicle crush, tire marks and end positions (Moser, Spek, Steffan, &
Makkinga, 2003). The method necessarily presupposes that the two vehicles collide at the intersection. Proper
speed reconstruction should therefore include the likeliness for such a collision, as that likeliness may depend
on speed. Secondly, the function may be a base for cause attribution in cases where one driver violated right of
way regulations while the other driver exceeded the speed limit.
The study is confined to the situation where a free traffic stream, the major stream, is crossed. The word
free is used to indicate that the individual vehicles approach the location without altering speed, regardless
of the presence of other traffic at the location. This implies that the major stream has traffic priority over all
conflicting traffic streams, consequently called minor streams. Another necessary assumption is that the loca-
tion itself has no properties that cause the major stream drivers to reduce speed, such as a sharp bend or speed
reduction measures. Typically these circumstances arise at residential entrances and some rural unsignalized
intersections. An example of a rural intersection is shown in Fig. 1. In that figure, the encircled numbers 2
and 8 indicate the major streams, and numbers 4, 6 and 7 indicate the minor streams.
When attempting to cross, the minor stream driver has to select the first occurrence of a sufficient gap
between two successive major stream vehicles. Whether the gap is sufficient depends on its duration, or gap
time. Gap time cannot be perceived as such, at least not without having to wait until the gap has ended: a
waiting driver must predict the gap time. One probable mechanism is tau-based TTC-estimation (Time-
To-Collision), as proposed by Lee (1976). The measure tau is defined as the area that an approaching object
occupies at the retina, divided by its rate of change, and may be considered as a direct percept. Retinal area is
a function of distance, and the change of change of retinal area is a function of both distance and speed.
Assuming that the object is not too close and approaching along the line of sight with a constant speed,
tau is proportional to the time it would take for the vehicle to collide with the observer. The mechanism is
attractive from an evolutionary point of view as it allows for predicting the arrival of threats of unidentified
size. It may be argued that in modern traffic the observer is familiar with both the approaching vehicle and the
world it moves in, so that other mechanisms may be more appropriate. Whatever the true nature of gap time
estimation in traffic, it requires the integration of distance- and speed-related perceptions, if not by mental pro-
cessing then through a specific property of our ingeniously engineered eye. In any case, speed perception plays
an important role in gap time estimation. Olson (1996) suggests that the speed of an approaching vehicle may
not easily be perceived, as the vehicle will hardly move against its background and—until it is very near—its
angular size will change only slightly. In an overview of previous studies, Dewar, Olson, and Alexander (2002)
report a general tendency toward overestimation of lower speeds and underestimation of higher speeds. As a
result it is expected that cross decisions become riskier as the speed of approaching major traffic increases.
Such an effect has been found in previous investigations. Simulator experiments by Hancock, Caird, Shekhar,
and Vercruyssen (1991), Hancock and Caird (1993), Alexander, Barham, and Black (2002) and Staplin (1995)
showed that the willingness to accept a gap with certain duration increases with the speed of the approach-
ing major stream vehicle. Based on cross decisions observed at unsignalized intersections in Canada,

Fig. 1. Typical three-legged rural intersection layout (taken from Weinert (2001)).
A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 157

Abou-Henaidy, Teply, and Hunt (1994) expressed the likelihood of a gap being accepted as a logit model, in
which both space gap (distance between successive major stream vehicles) and speed (of the approaching
major stream vehicle) were parameters. It appeared that, in order for the likelihood of acceptance to remain
unchanged, an increase in approach speed of 1 m/s must be balanced by a 3 m increase in space gap. As this is
insufficient to maintain the average 6 s critical gap, the latter result implies that decisions to cross tend to
become riskier as the speed of the approaching major stream vehicle increases. The studies discussed above
present ample evidence for an effect of major traffic approach speed on gap acceptance behavior. A qualitative
extrapolation of this effect suggests that accident probability increases with approach speed. Some recent stud-
ies suggest that this is actually the case. Two case-control studies in the Adelaide area (Kloeden, Ponte, &
McLean, 2001; Kloeden, McLean, & Gloneck, 2002) showed an increased (reconstructed) individual speed,
as compared to mean speed at the location, to be an important factor in accident probability. Further analysis
of the data underlying the latter study showed that the phenomenon holds for the subset of intersection acci-
dents for which the major stream vehicle was free (in the previously defined sense) and the speed of the major
stream vehicle could be reconstructed. An analysis of accident statistics for rural roads in England (Taylor,
Baruya, & Kennedy, 2002) found the junction accident frequency per road link within homogeneous road
categories to be proportional to the fifth power of the average speed of vehicles on that road link. The latter
study focused on average speed per road link rather than on individual vehicle speed. The result is nevertheless
still meaningful as road category may influence the speed expectancy of the crossing drivers.
This research aims to derive the relation between intersection approach speed and accident probability by
extrapolating gap acceptance behavior of crossing drivers. As a first step, a decision model is derived that has
gap time as well as major stream approach speed and required cross time as its variables. Limitations of
human perception are accounted for using Stevens power law (Stevens & Stevens, 1975). The model has a
number of free parameters that are adjusted to conform against the rejection and acceptance of gaps as
observed in a driving simulator experiment (Hancock & Caird, 1993) and in real traffic (Brilon & Weinert,
2001). Therefore, following the Panel on Human Performance Modeling of the National Research Counsel
(Baron, Kruser, & Messick Huey, 1990), the model may be classified as a descriptive model. The panel pro-
poses that, once parameter values have been established, a descriptive model may be used to raise predictions
for new situations. Since the acceptance of a too short time gap, by definition, yields a traffic conflict, the gap
acceptance model can be extrapolated to predict the likelihood of traffic conflicts as a function of approach
speed. Given the occurrence of a traffic conflict, the driver of the major stream vehicle may be able to delay
its arrival at the intersection through braking, thereby avoiding the accident. However, both human reaction
speed and vehicle dynamics limit the effectiveness of braking. Taking into account these limitations, the model
is further extrapolated to predict accident probability as a function of approach speed. Extrapolation yields a
predictive model, which requires validation. Validation has not been performed yet, as the available data was
inadequate, and gaining new data was beyond the scope of this study. However, to demonstrate the models
ability, it is fitted to describe the speed dependency and gap time dependency of conflict occurrence as found in
the simulator experiment.

2. Modeling gap acceptance behavior

Decision making at unsignalized intersections is often modeled as a process in which time gaps between
successive major stream vehicles whose duration exceeds a certain value—the critical gap—are accepted,
while shorter gaps are rejected. A comprehensive description of such models, referred to as critical gap mod-
els, is given in Troutbeck and Brilon (1999, chap. 8). In these models, the critical gap value is considered dis-
persed amongst individual drivers whilst the distribution of values is considered unique for each intersection.
A presumption of the critical gap models is that drivers are sensitive to the time it takes for an approaching
vehicle to arrive at the intersection. However, as discussed, several investigations show that both approach
time perception and gap acceptance behavior are influenced by the speed of the approaching vehicle. There-
fore an alternative gap acceptance model is derived, assuming that humans construct gap time indirectly by
dividing perceived distance over perceived approach speed. Both perceived distance and perceived approach
speed are modeled as exponential functions of their physical values, thus following Stevens power law (Ste-
vens & Stevens, 1975). Stevens law assumes the relation between the physical value I and its perceived value
158 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171

IPerc to be IPerc = kIv. The exponent v, usually called the power exponent, depends, according to Stevens the-
ory, on the mode, or type of stimulus. For a number of modes, Stevens estimated power exponent values
from experimental data. The perceived length of a projected line, for example, was found to vary almost lin-
early with the real length, yielding an exponent of about 1. For values of v between 0 and 1 the perception may
be thought of as being conservative with respect to expectation, in the sense that departures from this expec-
tation are underestimated. An example of this is the area of a projected square, for which Stevens reported an
exponent of 0.7, implicating that the area of squares smaller than some expected level is likely to be overes-
timated while the area of a larger square is likely to be underestimated. Exponents higher than 1, on the other
hand, suggest oversensitivity to the stimulus. The constant k is determined by expectation and by the choice of
units for stimulus and sensation. The power law is applied to model perceived approach time in Eq. (1). The
model estimates the perceived approach time (Tappr,Perc) as a function of real approach time and approach
speed (Tappr and V, respectively). The distance (s) is eliminated from the model, as it is simply the product
of approach time and approach speed. Please note that applying Stevens power law to tau-based perception
(Lee, 1976), mapping retinal area and its rate of change back onto distance and speed, would yield a similar
model. Parameters in the model are the power exponents for distance and approach speed (vs and vV, respec-
tively) and the associated constants (ks and kV). A list of denotations used is added in Appendix A.
k s svs ks v
T appr;Perc ¼ ¼ ðT appr Þ s V vs vV ð1Þ
k V V vV kV
A rational driver or pedestrian can be assumed to cross whenever his or her perception of approach time safely
exceeds his or her prediction of the time it would take to cross (Tcr). Let the cross time prediction be modeled
by Stevens power law as well, taking into account a safety factor g, then Eq. (2) is derived as a condition for
the acceptance of a gap.
vT
 1
ks vs vs vV vT ð v
1 vV Þ  vs k T k V vs
T appr;Perc > gT cr;Perc ) ðT appr Þ V > gk T ðT cr Þ ) T appr V s T cr > g ð2Þ
kV ks
The condition of Eq. (2) is not yet sufficiently defined, as the variable Tappr for approach time is not constant
in time, i.e. Tappr = Tappr(t). For an approaching vehicle, Tappr will decrease continuously until by the time the
vehicle passes Tappr reaches zero. For a waiting driver however, time can be divided in gaps. A first gap (in
traffic engineering often called the lag) starts when the minor stream vehicle arrives at the yield line and ends
when the first major stream vehicle passes. Subsequent gaps start when the preceding gap ends, and end when
the next major stream vehicle arrives. It is assumed that a decision to either cross or not cross is taken shortly
before or after the beginning of each gap. Accordingly it is the time span of the gap in combination with the
speed of the approaching major stream vehicle that determines how likely it is whether the gap is accepted.
From this derives the decision model of Eq. (3).
Cross if T gap V 1a T b
cr c
1
>1 ð3Þ
In the above Eq. (3), the constants are taken together to raise the following three model parameters: a ¼ vvVs ,
vT k T k V v1s
b ¼ vs and c ¼ ðg ks Þ . The parameters a and b govern the effect of variation in approach speed and cross
time respectively. The c parameter scales the threshold for crossing. If a = 1 and b = 0, implying that the time
gap perception is independent of speed and that the time required for crossing is not taken into account, the
model converges to the critical gap model. a = 0 yields another special case, for which speed is not at all per-
ceived and gap distance becomes the condition for gap acceptance. By definition, the parameters a and b are
directly related to the Stevens exponents for the perception of vehicle approach speed, distance and the pre-
diction of cross duration. Relevant exponent values for these specific modes, however, could not be found in
literature. It is doubtful whether published values would be of use, as it can be expected that numerous exper-
imental conditions influence the value. For instance, an experiment by Scialfa, Guzy, Leibowitz, Garvey, and
Tyrrell (1991) raised an average exponent of 1.36 for speed perception, whereas the previously mentioned gen-
eral tendency reported by Dewar et al. (2002) corresponds to an exponent below unity. The exceptional result
can possibly be attributed to a very specific condition, i.e. that for each trial the subjects were to watch the
vehicle approach over a fixed distance (122 m). It is therefore proposed that the parameters be estimated
for specific conditions such as a specific maneuver on a specific intersection at a specific time of day, by fitting
A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 159

the model to cross behavior as observed under those conditions. Thereby it must be considered that the model
parameters a, b and c will be dispersed among individual drivers. A set of values for the variables Tgap, V and
Tcr therefore yields a probability that the gap is accepted, rather than a certain outcome. To account for this, a
statistical model is derived. The only proposition made is that the logarithms of the odds (the logit) for cross-
ing against not crossing is equal to the left hand side of Eq. (3), scaled by a new parameter k. The proposition
implies that there is a 50% probability of crossing in case that the left hand side equals one. The resulting sta-
tistical model, shown as Eq. (4), takes the form of a logistic (cumulative) distribution. It can be shown that this
is a symmetric S-shaped cumulative distribution of which the standard deviation is inversely related to the
value of the k-parameter.
1
P ðCrossjT gap ; V ; T cr Þ ¼ ð4Þ
1 þ ekðln T gap þð1aÞ ln V b ln T cr ln cÞ

3. Parameter estimation

The model of Eq. (4) was fitted to gap acceptance data from simulator experiments as well as from traffic
observations. The simulator environment may not have provided all visual cues available to drivers in the real
world and behavior of subjects in a simulator may differ from real world behavior. It is, however, interesting
to compare results, as this may provide a basis for future simulator based studies. A simulator study may be
preferable over an observational study, as it allows for investigating differences between groups of subjects or
the implications of new safety measures. Additionally, subjects in a simulated environment can safely be sub-
jected to extreme conditions.
Considering that an observed decision is in fact a realization, thus having either 0 or 1 probability of accep-
tance, parameter values were obtained by applying logistic regression to Eq. (4). This method is implemented
in commercially available statistical packages such as SPSS, Statistica and S+; for this research the latter pro-
gram was used. The choice for logistic regression was thoroughly pondered, as Brilon, Koenig, and Troutbeck
(1997) showed by simulation that the use of logistic regression for critical gap estimation might lead to bias
and the detection of nonexistent correlations. Specifically, logistic regression was shown to yield a nonexistent
correlation between critical gap time and major stream traffic volume. Another cause for concern was that by
adopting the logistic distribution the population variation is assumed one-dimensional (governed by the
k-parameter), whereas all parameters of the decision model (a, b and c) may be distributed independently.
Therefore, the estimation procedure was validated for this specific application using Monte Carlo simulation.
It was found that the estimated a-values show no bias.
The available data, both from experiments as from observations, did not include measured cross times.
Therefore the model term b Æ ln Tcr was not taken into account. The costs of this neglect are that b is not esti-
mated, that estimates for the parameters c and k will be biased and that the estimator variance is increased.
Assuming that approach speed and cross time are not correlated, no bias will be introduced to the a-estimates.
However, when extrapolating the model to predict conflicts and accidents, cross time may prove to be an indis-
pensable parameter. An indicative value for b of 0.67 was measured in a separate experiment at the Nether-
lands Forensic Institute (Post, 2002). Estimation results are listed in Appendix B. Discussion of the estimator
results is mainly confined to the a-parameter, as this parameter relates approach speed and gap acceptance
behavior in the model.

3.1. Simulator experiments

The experimental gap acceptance data used here originates from simulator experiments by the University of
Minnesota, summarized in a publication by Hancock and Caird (1993). In a first experiment, 40 young driv-
ers, aged 24.2 years on average, were seated in a fixed-base driving simulator. The subjects were instructed to
drive up to the next intersection and turn left if they felt it was safe to do so. In the simulated environment,
each subject did arrive at a rural non-regulated intersection 49 times, to be confronted with an oncoming
stream of traffic of uniform speed. Each of the 49 times a different combination of gap time (3, 4, 5, 6, 7, 8
or 9 s) and speed (10, 20, 30, 40, 50, 60 or 70 miles per hour) was presented in a random order, such that
160 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171

all possible combinations were presented once to every subject. Each subject was confronted only with one of
four types of vehicles, a compact car, a full size car, a delivery truck or a motorcycle, which was randomly
assigned. For a more complete description of the experiment (see Hancock et al., 1991). The experiment
was later repeated with 40 older subjects, aged 67.1 years on average (Hancock & Caird, 1993). Fig. 2 shows
for every condition the percentage of subjects that considered a condition to be safe for crossing. (The 1993
publication by Hancock & Caird (1993) featured eight such plots, one for each combination of age group and
vehicle type. Fig. 2 presents their averaged percentages, as measured manually from these eight plots.)
The model fitted the data very well; more than 97% of the variation was accounted for. Overall estimates
for c and k were 19.4 (SE = 3.5) and 3.6 (SE = 0.15), respectively. Estimated a-values are presented in Fig. 3.
In addition to the lumped result, results are shown for the two age groups and per vehicle type. A lumped a-
value significantly below 1 was estimated (a = 0.602, SE = 0.027), implicating that the incorporation of per-
ceived speed or speed perception itself is not sufficient for pure gap time estimation. As a result, minor stream
drivers seeking an opportunity to turn left against free driving oncoming traffic may be expected to accept gaps
with a shorter duration as the speed of the vehicle closing the gap increases. This effect does not vary signif-
icantly with the type of the oncoming vehicle, but is stronger for older drivers (a = 0.497, SE = 0.046) than for
younger drivers (a = 0.671, SE = 0,022). A combined effect of age and approach speed on left turn gap judg-
ment has also been reported by Staplin (1995). Although the difference may in part be simulator dependent,
for instance because of younger subjects adapting better to the simulated environment than older drivers, we
consider it likely that a similar difference exists in the real world. The result does not necessarily imply that
older drivers behave more hazardously than younger drivers as degraded speed perception may be compen-

100%

90%

80%

70%
% Acceptance

60%

50%

40%

30%

20% 70

10%
50
0%
9 30
d
ee
Sp

7
Time 10 (mph)
gap 5

3(s)

Fig. 2. Gap acceptance data from the simulator study by Hancock and Caird (1993).
A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 161

0.8
α ± 2 SE

0.6

0.4
All Young Old Bike Compact Full Truck

Fig. 3. a-Values estimated from the experimental data reported by Hancock and Caird (1993).

sated by cautiousness. The number of dangerous decisions, (reported by Hancock & Caird (1993) as colli-
sions, although we would rather use the term conflicts) was actually lower for older drivers (33) than for
younger drivers (34). Apparently, older drivers compensate any degraded speed perception or speed integrat-
ing skills by adopting a higher threshold for perceived gap time, as may be deduced from the b-estimates: 35.1
(SE = 10.6) for older subjects against 12.4 (SE = 1.9) for younger subjects. This can be interpreted as a form
of target risk homeostasis (Wilde, 1994). If the number of conflicts with major stream traffic is indeed kept
constant in the longer term by increasing the threshold, the fraction of conflicts where the other vehicle is
speeding will rise. As the probability of a conflict turning into a collision increases with speed (this will be
shown in chapter 5), the number of collisions is likely to rise with age. This mechanism may explain the
increased involvement rates of older drivers as the priority violating party in intersection accidents (Caird
& Hancock, 2002).

3.2. Traffic observations

Traffic observation data were acquired from the Ruhr University in Bochum, Germany, obtained within a
project (Brilon & Weinert, 2001) sponsored by the Bundesanstalt für das Straßenwesen (BASt). The data were
derived from video observations at 10 three-legged and 2 four-legged unsignalized intersections in rural areas
in Germany. A typical example of the layout of a three-legged intersection was shown in Fig. 1. All of these
intersections featured two major streams of vehicles, indicated by the numbers 2 and 8, which had priority
over all conflicting minor streams numbered 4, 6 and 7. On the three-legged intersections, three types of
maneuvers are distinguished: stream 4 crossing streams 2 and 7 and then merging into stream 8, stream 6
merging into stream 2 and stream 7 crossing stream 2. For convenience, these types are referred to as left
turn, merge and oncoming. (NB: The latter maneuver is comparable to the maneuver that was studied in
the aforementioned experiments.) Each of the four-legged intersection were basically dealt with as two sepa-
rate three-legged intersections.
Intersections were observed for a 3- or 4-h period. For each passage of a major stream vehicle, the time and
speed of that passage were logged. The time of arrival at the stop- or yield line and the time of departure from
that line were logged for each minor stream vehicle. For detailed information about the data collection pro-
cedure, refer to the original publication (Weinert, 2001). From these observation data, a database of almost
30,000 decisions was derived. Each decision is associated with the corresponding outcome (wait or cross), gap
162 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171

time, speed, location and maneuver type. As only the passages of major vehicles in stream 2 were taken into
account, the results for the left turn maneuver may be subject to bias. Therefore, although results will be pre-
sented for all maneuvers, conclusions will be based on results for the merge and oncoming maneuvers only.
The model fits the observations reasonably, although the fraction of variance explained was considerably
less than for the simulator data (R2 = 0.76 for all observations). Presumably, behavior in the real world is
influenced by factors that are not incorporated in the model. For instance, as suggested by one of the review-
ers, drivers may reject a shorter but otherwise acceptable gap if he sees a larger gap coming along down the
traffic stream. This latter effect did not play a role in the simulator experiment, as subjects were confronted
with gaps of equal size. Over the entire database of traffic observations, an a-exponent relatively close to unity
is estimated (a = 0.856, SE = 0.036). From this result, it would seem that approach speed has a minor effect on
gap acceptance. When different maneuvers are distinguished, however, the relevance becomes more explicit, as
can be seen in Fig. 4. The oncoming and left turn maneuvers raise a-estimates considerably less than unity.
For the oncoming maneuver, a value of a = 0.565 (SE = 0.058) was estimated, which is consistent with the
results for the experimental data. The merge maneuver, in contrast, yields an estimate larger than unity
(a = 1.320, SE = 0.058). For the left turn maneuver, a value of a = 0.669 (SE = 0.070) was estimated, a result
that does not differ significantly from that of the oncoming maneuver. However, as discussed, the result for
the left turn maneuver is likely to be influenced by the traffic stream in far lane (indicated by the number 8 in
Fig. 1). In this respect it is worth noting that the fit for left turn maneuvers was worse than for the oncoming
maneuvers (R2 = 0.63 versus R2 = 0.84).
The estimated a-value for the merge maneuver corresponds to the results from a study by Wang, Chin, and
Lim (2003). The difference between the merge maneuver on the one hand and the left turn and oncoming
maneuvers on the other hand may be understood as follows. For the latter, some fixed minimum time gap
is needed to cross the major stream; there is no logical reason for a minor speed driver to adapt this time
gap criterion to the speed of the approaching vehicle. The a-value therefore relates mainly to limitations in
human speed estimation. In contrast, the merge maneuver implies not only entering the stream but also adap-
tation to the speed of the vehicle that closes the accepted gap. Thus, to merge before a faster vehicle, a larger
gap is needed. In addition, as approach speed is more critical when merging, drivers may increase a by putting
more effort in estimating the speed, for instance by watching for a longer period or by taking more glances.
The study by Scialfa et al. (1991) showed that a high sensitivity for approach speed is possible, given that suf-
ficient cues and an abundance of time are available.

1.6

1.4

1.2
α ± 2 SE

0.8

0.6

0.4
All Left turn Oncoming Left turn & Merge
Oncoming

Fig. 4. a-Values estimated from observation data gathered by Brilon and Weinert (2001), per maneuver type.
A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 163

2.5

1.5
α ± 2 SE

0.5

-0.5
All 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Location
Oncoming Merge

Fig. 5. a-Values estimated from observation data gathered by Brilon and Weinert (2001), per location.

As Fig. 5 shows, estimated values for a differ between locations. Variation may be expected as cues for
speed perception may differ between intersections. For instance, vehicles approaching the intersection along
a straight line toward the observer will not move relatively to the background, as do vehicles approaching
on a road bending toward the intersection. In the latter situation, better speed perception and (thus) higher
values for a may be expected. This explanation would imply that the respective a-estimates for the different
classes of maneuvers share a source of variance and, thus, that these values correlate positively. A significant
correlation, however, could not be established. Neither could the variation be explained in terms of the aver-
age speed on the major road and the number of vehicles per hour. It is noteworthy that the a-values for the
merge maneuver on locations number 11 and 12 are below 1, whereas merge maneuvers on all other locations
yield values higher than 1. Apparently, drivers on location number 11 base their decision to start the merge
maneuver merely on gap distance, as a is close to zero. This specific location presented a special case as a
nearly 100 m long bus stop was reportedly used as an acceleration lane by a substantial fraction of the drivers.
Location number 12, for which the estimated a-value—although not significantly—was just below one, fea-
tured a 40 m bus stop.

4. First extrapolation: Conflict probability

4.1. Modeling

A traffic conflict arises when a minor stream driver accepts a gap that is shorter than the time required for
the cross maneuver. The incidence of traffic conflicts can thus be regarded as an extrapolation of gap accep-
tance behavior. The probability of a traffic conflict per cross maneuver given a certain speed can be modeled
accordingly, as in Eq. (5). In that equation, UV(Tgap), denotes the PDF for gap time given the approach speed
V, and W (Tcr) is the PDF for cross time. The combined distribution U(V, Tgap) may be estimated from obser-
vations at the location. Any dependence between cross time on the one hand and gap time and approach speed
on the other hand is ignored. In other words, it is assumed that drivers will not increase their (intended) cross-
ing speed in order to be able to accept a shorter gap. Future validation may prove this to be an oversimpli-
fication. Measuring this correlation is not trivial, though, since cross time can only be observed for accepted
gaps. Within this selection, faster drivers will be overrepresented in the shorter gap times, whether the depen-
dence exists or not.
164 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171
Z 1Z 1
UV ðT gap ÞWðT cr Þ
P ðConflictjV Þ ¼ dT cr dT gap ð5Þ
0 T gap 1 þ ekðln T gap þð1aÞ ln V b ln T cr ln cÞ
Provided that a is between 0 and 1, a characteristic of Eq. (5) is that the probability of conflicts continuously
increases, in an S-shaped manner, from zero at zero speed to a certain limit at infinite speed. The limit corre-
sponds to the risk taken by a driver who crosses at the beginning of a random gap, without considering the
duration of that gap. This is intuitively right, as it is impossible to predict the arrival of a vehicle that ap-
proaches with an infinite speed. An expression for the conflict probability given a certain gap time, Eq. (6),
is derived in the same way. Here UT gap ðV Þ denotes the PDF for approach speed given the gap time. The result-
ing function has a limit of zero for gap times approaching zero as well as for very large gaps. The reason is that
very short gaps are not accepted, and therefore safe, while the large gaps offer ample time to safely complete
the cross maneuver.
Z 1Z 1
UT gap ðV ÞWðT cr Þ
P ðConflictjT gap Þ ¼ kðln T gap þð1aÞ ln V b ln T cr ln cÞ
dT cr dV ð6Þ
0 T gap 1 þ e

A straightforward approach for the validation of the extrapolated model would be to evaluate the equation
with parameters estimated for a specific location, and compare the result with conflicts observed at the same
location. Conflict counts (referred to as collisions by Hancock & Caird (1993)) were available for the exper-
iments. However as discussed, these data did not include cross times. To show that the model has at least the
ability to appropriately describe conflict rate as a function of speed, Eqs. (5) and (6) were fitted to the conflict
counts as reported by Hancock and Caird (1993). The a- and b-parameters were constrained to their estimates
of 0.60 and 0.67, respectively. A log-normal distribution was chosen for cross time, of which the parameters l
and r were fitting parameters. Considering that the aforementioned estimation (without the term b Æ ln Tcr)
found a c-value of 19.4, the c-parameter was constrained to 19.4 Æ e0.67l. The model was slightly modified
by assuming that an unknown fraction p of the crossing decisions were trivial, i.e. acceptances regardless
of speed and gap time. This was necessary to explain a substantial number of conflicts that arose at low speed.
These trivial decisions are attributed to failure of subjects to understand the experiment, as in normal traffic
the concern is to reject one gap for a later gap rather than to reject making a left turn for not turning at all.
The resulting fit is shown in Figs. 6 and 7. Fitted parameter values were p = 0.043, k = 11.5, l = 1.26 and
r = 0.23. Thus 4.3% of the decision may be regarded as trivial. Without these trivial decisions, the conflict

7%

6%

5%
Conflict rate

4%

3%

2%

1%

0%
0 10 20 30 40 50 60 70 80 90 100
Speed (mph)
Observations Model

Fig. 6. Model fit: conflict rate against approach speed.


A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 165

6%

5%

4%
Conflict rate

3%

2%

1%

0%
2 3 4 5 6 7 8 9
Gap time (s)
Observations Model

Fig. 7. Model fit: conflict rate against gap time.

rate would reach zero as speed and gap time approach zero. The k-value of 11.5 is higher than the value that
was estimated from the gap acceptance data without cross times (3.6), suggesting less variance in the model.
This is to be expected, since trivial decisions and variation in cross time may partly explain the variance of the
model. The values estimated for l and r cannot be compared to known values, as the cross times for the exper-
iment were not reported. However, a log-normal distribution with parameters l = 1.26 and r = 0.23, having
5th and 95th percentiles of 2.4 and 5.2 s, respectively, is not out of the ordinary for one-lane crossing times.
As can be seen from Figs. 6 and 7, the model describes the conflict rates that arose in the experiments reason-
ably well. The fit is not good for the larger gap times (P6 s in Fig. 7), for which the rates predicted by the model
are well below the presented rates. A possible explanation is that the first crossing opportunity for each condition
(the lag) may have been shorter than the reported gap time, depending on the time of arrival of the subject at the
crossing. A subject accepting such a lag in fact accepts the risks associated with the shorter gap time. The problem
would not occur with observation data, as long as the gap time for the first crossing opportunity is measured from
the minor vehicles arrival at the crossing. Another possibility, however, is that the assumption is wrong that the
decision whether or not to cross is made shortly before or after the start of the gap. For instance, drivers may have
adopted a strategy to let the vehicle approach up to a certain distance, to allow for a more accurate speed esti-
mation. A future validation study may discriminate between these possible causes.

5. Second extrapolation: Accident probability

Not every traffic conflict results in an accident. A major stream driver, when confronted with a vehicle
crossing his path, may extend the available time by braking. The effective braking time commences only after
a certain reaction time and braking deceleration is limited physically, and therefore the relative time gain
decreases as speed increases. As a result the fraction of conflicts which do not result in an accident (the conflict
recovery rate) will decrease for higher initial speeds. The remaining accident probability can be calculated
using the right hand side of Eq. (5), taking the extended available time as the lower integration limit for
the inner integral (instead of the gap time Tgap). Considering that the speed remains constant during a certain
brake reaction time (Treac) after which a constant deceleration (a) is present, Eq. (7) yields the extended avail-
able time. In that equation, the brake time is calculated as the distance available for braking divided by the
average speed over that distance. Special cases arise when the gap time is shorter than the reaction time
and when the vehicle comes to a stop before reaching the conflict area.
166 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171
8
>
> T gap if T gap  T reac 6 0
< V
T av ¼ 1 if T gap  T reac P 2a ð7Þ
>
> ðT gap T reac Þ
: T reac þ p2V
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
V
if 0 < T gap  T reac < 2a
Vþ V 2aV ðT gap T reac Þ

Let deceleration level a and brake reaction time Treac be independently distributed conform PDFs X(a) and
P(Treac) respectively, then Eq. (8) yields the accident probability given the major vehicles speed.
Z 1Z 1
UV ðT gap ÞWðT cr ÞXðaÞPðT reac Þ
P ðAcc.jV Þ ¼ dT cr dðT gap ; a; T reac Þ ð8Þ
0 T av 1 þ ekðln T gap þð1aÞ ln V b ln T cr ln cÞ
It is recognized that a direct validation of this latter extrapolation will be difficult, as accidents occur relatively
infrequently. Nevertheless, Eq. (8) is evaluated to predict the number of accidents per cross maneuver, as a
function of approach speed, for the experimental situation. The deceleration level for emergency braking (a)
will vary for different locations, vehicles, tires and weather conditions (Hegmon, 1987). For specific situations,
for instance when analyzing a certain accident, a value may be measured using an accelerometer. Measured
deceleration levels are also published in literature (Cliff & Bowler, 1998; Ebert, 1989; Goudie, Bowler, Brown,
Heinrichs, & Siegmund, 2000; Wallingford, Greenless, & Christoffersen, 1990). A Gaussian distribution was
adopted for X(a), with mean 8.1 m/s2 and standard deviation 0.32 m/s2, as proposed by Goudie et al. (2000)
for dry road conditions. Some research is published on the behavior of major stream drivers being confronted
with sudden intrusions of pedestrians and minor stream vehicles (Araki & Matsuura, 1990; Lechner & Mala-
terre, 1991). The study by Lechner and Malaterre (1991) showed that the majority of drivers in such situations
choose to brake whereas a minority tries to swerve around without braking. Obviously the presented model
does not describe the latter behavior. Values for the reaction time (Treac) vary between individuals and condi-
tions. Some approximate values for the reaction time for braking after a surprise-intrusion are also reported in
literature; a recent summary was found in Green (2000). In accordance with indicative guidelines (Green, 2000),
a log normal distribution with parameters r = l = 0.5 was chosen for the reaction time distribution P(Treac).
The predicted accident rate for the experimental situation, taken into account the previously estimated fraction
of trivial decisions, is shown against approach speed in Fig. 8. The figure also shows the predicted conflict rates
as calculated in the previous chapter. As was expected, the accident rate increases continuously with approach
speed. This implies that the PDF for the speed of major stream traffic given involvement in a collision with cross-

7%
Modelled conflict / collision rate

6%

5%

4%

3%

2%

1%

0%
0 10 20 30 40 50 60 70 80 90 100
Speed (mph)
Conflicts Collisions

Fig. 8. Conflict and accident rate against approach speed.


A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 167

100%

80%
Conflict recovery

60%

40%

20%

0%
0 10 20 30 40 50 60 70 80 90 100
Speed (mph)
Fig. 9. The fraction of conflicts not resulting in an accident.

ing traffic, is shifted to higher speeds when compared with the unconditional PDF (that is, the PDF for the
speeds of all vehicles in that stream, whether involved in a collision or not). Note that the percentage of conflict
recovery, i.e. the fraction of conflicts that do not result in accidents as a result of emergency braking by the ma-
jor stream vehicles driver, decreases with increasing approach speed. This is shown in Fig. 9. Note that without
the trivial decisions, the limit for the conflict recovery percentage would be 100 as the speed approaches zero.

6. Conclusions and recommendations

A model is proposed for gap acceptance at intersections (4). Merely as a means to derive the model,
assumptions were made on human perception and decision making. The model shows a very good fit for
the simulator data and a reasonable fit for the traffic observations, and can be considered meaningful in
the sense that yields significant parameter values. This result does by no means prove the underlying assump-
tion right; studying human perception and cognition was not a goal of this study. The model does show that a
drivers decision to cross or merge before an approaching major stream vehicle is not only governed by the
time it takes the latter vehicle to arrive at the intersection, but also by the speed of the approaching vehicle.
Drivers who need to cross a stream tend to accept smaller time gaps as the approach speed increases. That
effect appears to be stronger for older drivers than for younger drivers, suggesting that the older driver is more
prone to collide with speeding vehicles. A reversed effect rules merging into a traffic stream: for higher
approach speeds, gaps need to be longer to be accepted. The latter difference may be explained by the fact that
for merging into a stream, adaptation to the speed of the approaching vehicle is necessary.
As an implication of the speed dependency of gap acceptance behavior, the probability that a crossing vehi-
cle collides with a crossing major stream vehicle can be expected to increase with the latter vehicles approach
speed. It is proposed that this probability may be quantified as a function of approach speed by extrapolating
the gap acceptance model. Therefore it is deemed necessary that a measure for crossing time is available for
each minor stream vehicle, which was not the case with the available data. Using the regression techniques
described in this article, parameters of the model can be estimated for specific situation, such as location,
maneuver and time of day of a certain accident. The necessary data can be derived from video observations
if recorded under the same conditions.
The model is not yet sufficiently validated to allow for this kind of extrapolation. The ideal validation study
would imply parameter estimation and extrapolation for a number of intersections, and comparing model
168 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171

predictions to accident counts. As accidents are relatively infrequent events, a validation against conflict
counts may be more practical. However, this leaves the second extrapolation invalidated. The latter extrapo-
lation relies on assumptions regarding driver behavior in the face of a looming collision, which may best be
studied in a driving simulator. A plausible explanation for the model to fail in validation would be, that driv-
ers adapt their crossing speed to the (perceived) gap time. The model may be adapted for this kind of behavior,
by introducing a parameter that describes a correlation between cross time and gap time.
For the intended use in forensic accident reconstruction and accident cause attribution, the observation
technique as well as the extraction of data from video may be far too expensive. Therefore it is recommended
that the parameter variation between locations and conditions is further studied. This may allow for a cate-
gorization of intersections, minimizing the need for case-specific studies. An additional merit of categorization
is that it enables the adoption of the case-control design, as used by Kloeden et al. (2001) and Kloeden et al.
(2002), for a validation study. It is further suggested that the model is extended to account for situations where
more than one lane is crossed or where one lane is crossed and another one is merged, such as the left turn
maneuver in this study. The present restriction to simple maneuvers is a major disadvantage for forensic
practice.

Acknowledgements

Some of the key statistic methods used in this research were recommended by Dr. André Hoogstrate of the
Netherlands Forensic Institute. Dr. A. Weinert and Prof. Dr. Ing. W. Brilon, kindly permitted by the German
Bundesanstalt für das Strassenverkehrswesen, provided the raw data from the traffic observations and made
some very useful suggestions in discussing the preliminary results. We are obliged to the reviewers, whose con-
tributions greatly improved the article.

Appendix A. List of annotations

a model parameter: sensitivity to approach velocity


b model parameter: sensitivity to cross time
c model parameter: threshold for crossing
g safety factor
k model parameter: (inverse to) population variance
l lognormal distribution parameter
r lognormal distribution parameter
U(V, Tgap) combined distribution of V and Tgap
vi exponent in Stevens power law, index for mode
W(Tcr) distribution of Tcr
a, X(a) brake deceleration, distribution of a
ki constant in Stevens power law, index for mode
P(Acc.jV) probability of an accident given V
P(ConflictjTgap) probability of a conflict given Tgap
P(ConflictjV) probability of a conflict given V
P(CrossjTgap, V, Tcr) probability of crossing given Tgap, V and Tcr
s distance of the approaching vehicle
Tappr time for approaching vehicle to arrive
Tappr,Perc perception of Tappr
Tav available time, if approaching vehicle brakes
Tcr time needed for the cross maneuver
Tcr,Perc perception of Tcr
Tgap gap time
Treac, P(Treac) reaction time for emergency braking, distribution
V speed of the approaching vehicle
A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 169

Appendix B. List of estimation results

B.1. Simulator experiments


Condition # R2 a SE c SE k SE
All 3920 0.974 0.602 0.027 19.347 3.462 3.611 0.153
Young 1960 0.980 0.671 0.022 12.435 1.886 4.695 0.192
Old 1960 0.953 0.497 0.046 35.063 10.615 3.824 0.238
Bike 980 0.896 0.579 0.060 19.062 7.320 3.631 0.332
Compact 980 0.936 0.622 0.040 15.713 4.071 3.617 0.235
Full 980 0.915 0.606 0.055 20.152 7.696 4.465 0.399
Truck 980 0.931 0.610 0.042 22.468 6.477 3.221 0.214

Fitted model (using S+ for non-linear regression (Insightful Corporation, 2001))


1
P ðCrossjT gap ; V Þ ¼ ðB:1Þ
1 þ ekðln T gap þð1aÞ ln V ln cÞ

B.2. Traffic observations


Condition # R2 a SE c SE k SE
All 29562 0.762 0.867 0.036 0.985 0.005 4.915 0.071
Left turn 9759 0.628 0.669 0.070 0.999 0.010 3.762 0.086
Oncoming 9761 0.854 0.566 0.058 0.998 0.008 7.226 0.233
Left turn and oncoming 19520 0.749 0.616 0.046 0.984 0.007 4.696 0.082
Merge 10042 0.792 1.328 0.058 0.998 0.009 5.585 0.147
Not a lag, merge 3965 0.822 1.305 0.090 1.180 0.020 6.124 0.269
Not a lag, oncoming 3513 0.862 0.569 0.096 1.103 0.016 7.852 0.437
Lag, merge 6077 0.761 1.277 0.081 0.890 0.011 5.360 0.193
Lag, oncoming 6248 0.832 0.507 0.078 0.930 0.011 6.867 0.289
No queue, merge 3086 0.770 1.164 0.124 0.999 0.018 4.914 0.231
No queue, oncoming 4110 0.852 0.425 0.099 0.993 0.013 7.129 0.354
Queue, merge 6956 0.800 1.378 0.066 0.996 0.010 5.952 0.191
Queue, oncoming 5651 0.855 0.647 0.072 1.002 0.011 7.293 0.311
No follow-up, merge 4038 0.809 1.351 0.097 1.105 0.017 5.668 0.235
No follow-up, oncoming 3706 0.862 0.405 0.106 1.062 0.015 7.230 0.379
Follow-up, merge 6004 0.780 1.236 0.073 0.933 0.011 5.628 0.197
Follow-up, oncoming 6055 0.845 0.623 0.072 0.963 0.010 7.336 0.306
Did not accept the lag, merge 5471 0.806 1.221 0.080 1.300 0.023 5.993 0.230
Did not accept the lag, oncoming 4933 0.842 0.516 0.085 1.206 0.019 7.342 0.342
Motorcycle, merge 91 0.818 2.867 0.999 1.023 0.107 5.518 1.560
Motorcycle, oncoming 118 0.908 1.228 0.521 0.891 0.101 7.402 2.446
Location 1, Merge 149 0.743 1.137 0.819 1.000 0.081 4.515 0.883
Location 1, Oncoming 248 0.835 1.095 0.523 1.000 0.052 6.567 1.212
Location 2, Merge 1113 0.834 1.539 0.161 1.000 0.027 6.660 0.589
Location 2, Oncoming 1126 0.859 0.414 0.157 1.000 0.026 7.291 0.726
Location 3, Merge 1330 0.820 1.317 0.160 1.000 0.023 6.117 0.470
Location 3, Oncoming 686 0.883 0.897 0.172 1.000 0.031 8.151 1.066
Location 4, Merge 504 0.767 1.350 0.216 1.000 0.038 5.598 0.691
(continued on next page)
170 A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171

Appendix B (continued)
Condition # R2 a SE c SE k SE
Location 4, Oncoming 329 0.913 0.121 0.432 1.000 0.060 8.432 1.686
Location 5, Merge 268 0.886 1.454 0.399 1.000 0.067 6.390 1.174
Location 5, Oncoming 233 0.966 0.727 0.417 1.000 0.045 18.765 8.095
Location 6, Merge 252 0.823 2.165 0.586 1.000 0.039 8.228 1.607
Location 6, Oncoming 246 0.821 0.707 0.406 1.000 0.052 6.671 1.214
Location 7, Merge 1391 0.794 1.331 0.169 1.000 0.024 6.122 0.481
Location 7, Oncoming 1482 0.844 0.593 0.149 1.000 0.021 7.895 0.727
Location 8, Merge 947 0.742 1.239 0.137 1.000 0.022 6.317 0.492
Location 8, Oncoming 165 0.778 0.079 0.432 1.000 0.045 8.493 1.856
Location 9, Merge 532 0.819 2.054 0.353 1.000 0.038 5.914 0.671
Location 9, Oncoming 431 0.861 0.574 0.269 1.000 0.035 8.339 1.302
Location 10, Merge 1242 0.774 1.552 0.180 0.999 0.022 6.373 0.494
Location 10, Oncoming 1394 0.852 0.741 0.140 1.000 0.020 8.283 0.717
Location 11, Merge 993 0.730 0.120 0.386 1.000 0.044 3.619 0.292
Location 11, Oncoming 1937 0.825 0.393 0.171 1.000 0.020 6.366 0.453
Location 12, Merge 817 0.773 0.940 0.183 1.000 0.035 4.788 0.395
Location 12, Oncoming 1075 0.836 0.558 0.161 1.000 0.027 6.171 0.524
Location 13, Merge 64 0.927 2.306 1.322 1.000 0.132 12.322 9.044
Location 13, Oncoming 94 0.791 0.734 0.461 1.000 0.081 6.034 1.694
Location 14, Merge 440 0.728 1.571 0.301 1.000 0.048 4.846 0.623
Location 14, Oncoming 315 0.884 0.572 0.430 1.000 0.068 7.643 1.720

Fitted model (S+ was used for logistic regression (Insightful Corporation, 2001))
 
P ðCrossjT gap ; V Þ
ln ¼ a0 þ a1 ln T gap þ a2 ln V ðB:2Þ
1  P ðCrossjT gap ; V Þ
With
s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2  2
a2 oa oa
a¼1 SEðaÞ ¼ SEða1 Þ þ SEða2 Þ
a1 oa1 oa2
s
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2  2
a oc oc ðB:3Þ
a 0
c¼e 1 SEðcÞ ¼ SEða0 Þ þ SEða1 Þ
oa0 oa1
k ¼ a1 SEðkÞ ¼ SEða1 Þ
For the in-traffic observations, velocities (V) were normalized to the mean velocity at the location, and gap
times (Tgap) were normalized to the critical gap time for the specific maneuver at each location. This to enable
fitting against data of multiple locations. As a result, estimates for the gamma-parameter approach unity (1).
The critical gapa0time Tcrit was estimated by fitting the observations to the model in Eq. (B.4), and then calcu-

lating T crit ¼ e a1 .
 
P ðCrossjT gap Þ
ln ¼ a0 þ a1 ln T gap ðB:4Þ
1  P ðCrossjT gap Þ

References

Abou-Henaidy, M., Teply, S., & Hunt, J. D. (1994). Gap acceptance investigations in Canada. In Second international symposium on
highway capacity (vol. 1, pp. 1–20).
A.C.E. Spek et al. / Transportation Research Part F 9 (2006) 155–171 171

Alexander, J., Barham, P., & Black, I. (2002). Factors influencing the probability of an accident at a junction: results from an interactive
driving simulator. Accident Analysis and Prevention, 34, 779–792.
Araki, K., & Matsuura, J. (1990). Drivers response and behavior on being confronted with a pedestrian suddenly darting across the road.
Technical Report 900144, Society of Automotive Engineers, Warrendale, PA.
Baron, S., Kruser, D. S., & Messick Huey, B. (Eds.). (1990). Quantitative modeling of human performance in complex, dynamic systems.
Washington, DC: National Academy Press.
Brilon, W., Koenig, R., & Troutbeck, R. J. (1997). Useful estimation procedures for critical gaps. In Third international symposium on
intersections without traffic signals (pp. 71–87).
Brilon, W., & Weinert, A. (2001). Ermittlung aktueller grenz- und folgezeitlücken für auerortsknoten ohne lichtsignalanlagen. Straßenbau
und Straßenverkehrstechnik (vol. 828).
Caird, J. K., & Hancock, P. A. (2002). Left-turn and gap acceptance crashes. In R. E. Dewar & P. Olson (Eds.), Human factors in traffic
safety (pp. 613–651). Tuscon, AZ: Lawyers & Judges Publishing Co. [Chap. 19].
Cliff, W. E., & Bowler, J. J. (1998). The measured rolling resistance of vehicles for accident reconstruction. Technical Report 980368,
Society of Automotive Engineers, Warrendale, PA.
Dewar, R. E., Olson, P. L., & Alexander, G. J. (2002). Perception and information processing. In R. Dewar & P. Olson (Eds.), Human
factors in traffic safety (pp. 13–42). Tuscon, AZ: Lawyers & Judges Publishing Co.
Ebert, N. E. (1989). Sae tire braking traction survey: A comparison of public highways and test surfaces. Technical Report 890638, Society
of Automotive Engineers, Warrendale, PA.
Goudie, D. W., Bowler, J. J., Brown, C. A., Heinrichs, B. E., & Siegmund, G. P. (2000). Tire friction during locked wheel braking.
Technical Report 2000-01-1314, Society of Automotive Engineers, Warrendale, PA.
Green, M. (2000). How long does it take to stop? methodological analysis of driver perception-brake times. Transportation Human
Factors, 2, 195–216.
Hancock, P. A., & Caird, J. K. (1993). Factors affecting older drivers left turn decisions. Technical Report, Transportation Research
Board, National Research Council, Washington, DC.
Hancock, P.A., Caird, J.K., Shekhar, S., & Vercruyssen, M., 1991. Factors influencing drivers left turn decisions. In Human factors society
35th annual meeting, Santa Monica, CA (pp. 1139–1143).
Hegmon, R. R. (1987). Tire-pavement interaction. Technical Report 8702241, Society of Automotive Engineers, Warrendale, PA.
Insightful Corporation. (2001). S-Plus 6 for Windows guide to statistics, Seattle, WA: vol. 1. Insightful Corporation.
Kloeden, C. N., McLean, A. J., & Gloneck, G. (2002). Reanalysis of traveling speed and the risk of crash involvement in adelaide south
australia. Technical Report CR207, Australian Transport Safety Bureau, Canberra.
Kloeden, C. N., Ponte, G., & McLean, A. J. (2001). Traveling speed and the risk of crash involvement on rural roads. Technical Report
CR204, Australian Transport Safety Bureau, Canberra.
Lechner, D., & Malaterre, G., 1991. Emergency maneuver experimentation using a driving simulator. Technical Report 910016, Society of
Automotive Engineers, Warrendale, PA.
Lee, D. N. (1976). A theory of visual control of braking based on information about time-to-collision. Perception, 5, 437–459.
Moser, A., Spek, A. C. E., Steffan, H., & Makkinga, W. J. (2003). Application of the Monte Carlo methods for stability analysis within the
accident reconstruction software PC-crash. Technical Report 2003-01-0488, Society of Automotive Engineers, Warrendale, PA.
Olson, P. (1996). Forensic aspects of driver perception and response. Tuscon, AZ: Lawyers & Judges Publishing Co.
Post, T. L. (2002). Het inschatten van de tijd die nodig is om een afstand af te leggen: hoe nauwkeurig zijn voetgangers en bestaan er
sexeverschillen? Internal report, Netherlands Forensic Institute (in Dutch).
Scialfa, C. T., Guzy, L. T., Leibowitz, H. W., Garvey, P. M., & Tyrrell, R. A. (1991). Age differences in estimating vehicle velocity.
Psychology and Aging, 6, 60–66.
Staplin, L. (1995). Simulator and field measures of driver age differences in left-turn gap judgements. Transportation Research Record,
1485, 49–55.
Stevens, S., & Stevens, G. (1975). Psychophysics, Introduction to its perceptional, neural and social prospects. New York: Wiley.
Taylor, M. C., Baruya, A., & Kennedy, J.V. (2002). The relationship between speed and accidents on rural single carriageway roads.
Technical Report 511, TRL.
Troutbeck, R. J. & Brilon, W. (1999). Unsignalized intersection theory. In N. Gartner, C. Messer, & A.K. Raiti (Eds.), Revised monograph
on traffic flow theory, US Department of Transportation, Federal Highway Administration [chap. 8].
Wallingford, J. G., Greenless, W., & Christoffersen, S. (1990). Tire-roadway friction coefficients on concrete and asphalt surfaces
applicable for accident reconstruction. Technical Report 9000103, Society of Automotive Engineers, Warrendale, PA.
Wang, Z., Chin, H. C., & Lim, K. B. (2003). Collection and evaluation of projected time gap using CAMDAS. Road & Transport
Research, 12.
Weinert, A. (2001). Grenz- und Folgezeitlücken an Außerortsknoten ohne Lichtsignalanlagen. Ph.D. thesis, Lehrstuhl für Verkehrswesen,
Ruhr Universität, Bochum, Germany.
Wilde, G. J. S. (1994). Target risk: Dealing with the danger of death, disease and damage in everyday decisions. Toronto, Ontario, CA: PDE
Publications. Available from http://psyc.queensu.ca/target/.

You might also like