Professional Documents
Culture Documents
A consistent design allows drivers to perform safely the task of driving, Several research efforts have been undertaken to measure geo-
allowing attention or capacity to be dedicated to obstacle avoidance and metric consistency. These studies have used various methods and
navigation. A measure of the consistency of a design is the amount of parameters to model or rate geometric features, or combinations of
visual information needed by a driver to maintain an acceptable path on features, by the difficulty they could pose to a driver. The logic under-
the roadway. Vision occlusion is a technique that measures driver visual lying these efforts is that the more difficult a feature or feature com-
demand on a roadway. It allows a more direct evaluation of the effects of bination, the greater the visual information processing requirement
various geometric elements on the driver. Studies of the effects of varia- and, in turn, the less desirable the feature.
tions of curve radius, deflection angle, spacing, and sequences revealed A recent study sponsored by FHWA used a vision occlusion
several relationships between roadway geometry and visual demand. method to evaluate visual demand in both closed-course condi-
Curve radius and its reciprocal were found to be significantly related to tions (called a test track study) and on actual rural two-lane road-
visual demand in both on-road and test track studies. Small changes in ways (called an on-road study) under controlled conditions (1).
visual demand were also found between types of curve pairs (S and bro- This paper presents the findings from the study.
ken back) with differing spacing between the curves. Visual demand was
found to be a promising measure of effectiveness for use in studies of
design consistency. OBJECTIVE
of available volunteers. The primary criteria used for selection were the second curve identical in direction to the first curve) were tested
gender and age, with equal numbers selected in each of the age cate- for their effects on visual demand. Broken-back curves have been
gories used in the study (18 to 24, 35 to 54, and 55 and older). All cited as violating expectancy because drivers do not expect to be
drivers were required to have a valid U.S. driver’s license. The driv- confronted with two closely spaced turns in the same direction; this
ers used in the study encompassed a wide range of education levels; is contrasted with the more typical pattern of the S-curve (2). The
11 had a high school diploma or a diploma plus some college, 8 had spacing between the paired curves was varied in a controlled man-
bachelor degrees, and 5 had advanced degrees. Additional informa- ner in the hope of detecting any carryover effect, as portrayed in the
tion obtained from the drivers was the approximate number of miles work by Messer et al. (3). Each curve in the sequence had a radius
driven annually [42 percent reported 16 095 to 24 140.2 km (10,001 of 145 m (475.7 ft) and a deflection angle of 45° with a separation
to 15,000 mi)] and where they drove most (79 percent reported a mix distance of either 61 or 274 m (200 or 899 ft).
of all locations).
Although it was desirable for analysis that all drivers drive all test
conditions, three drivers were unable to complete the testing program. Equipment
One driver resigned from the study after completing approximately
one-third of the study (data not used), and two other drivers were The LCD visor was worn by the driver (see Figure 2). This visor pro-
unable to complete the final third of the study because of scheduling vides clear vision if an AC voltage is applied across it but is other-
conflicts (data included in analysis). wise opaque. When it is opaque, diffused light is still transmitted
through the visor, although objects or features are not visible. This
diffused light greatly reduces the amount of light adaptation for the
Test Courses driver that would otherwise be necessary when the visor is closed
and opened. The visor was controlled by a laptop computer with
Testing was performed at Texas A&M University’s Riverside Cam- associated analog-digital conversion module. The laptop created a
pus Test Facility. This facility, a former military air field, provides data file that recorded the times at which vision requests took place
the opportunity to display a wide variety of large-scale driving sce- and recorded downrange distances.
narios. The various runways available range in width from approxi- Distance and velocity were sensed by means of a Datron electronic
mately 90 m (295.3 ft) to more than 150 m (492 ft) and are more than (no contact) fifth wheel suspended from the back of the 1991 Ford
2000 m (6561.7 ft) long. Taurus station wagon. Path excursions were denoted by an input by
Because it was not possible to construct a single test course that the experimenter to the data file (in the “event” column). A backup
contained all of the features to be examined, three different courses camcorder recording from behind the rearview mirror was available
were designed and delineated. Figure 1 illustrates the course 2 lay- for confirmation of the data or an excursion, if necessary. The video-
out. Lanes 3.6 m (11.8 ft) wide were delineated through use of tapes included an imprint of the driver number, the run and course
removable yellow markings. The markings were 10 cm wide by 5 cm numbers, and the running distance and time. The major data point, in
high (4 in. by 2 in.) and were placed at 6-m (19.7-ft) intervals on both any event, was whether a lane excursion occurred.
sides of the marked lane. Some extraneous pavement markings were
present on the test track, although subjects did not generally appear
to be adversely affected by the markings. Driver Demand Occlusion
A limited number of curve sequences were also tested. Separation
When testing, the driver could request a glimpse of the road for a
distance and sequence type (i.e., an S-curve has the second curve
set interval of time by pressing a floor-mounted button with his or
opposite in direction to the first curve, and a broken-back curve has
her left foot. Drivers were instructed to request only as much vision
as necessary to stay on the course. The vehicle was equipped with
cruise control, and for the purposes of this study the speed was set
to 72 km/ h (45 mph). This speed was selected primarily because of
previous successful studies conducted at similar speeds at TTI and
in the literature and the desire to compare results obtained in the cur-
rent study with those obtained earlier (4, 5). This speed also permit-
ted the layout of a challenging but drivable course in the available
test area.
During the testing, vision requests and the location of the vehicle
were recorded on a personal computer. The length (fixed at 0.5 s as
in past studies, the approximate duration of an eye fixation between
saccades) and location of each vision request were related to the test
FIGURE 1 Closed-loop test course 2. course. A visual demand observation was calculated from this infor-
Wooldridge et al. Paper No. 00 - 0596 73
mental task (i.e., driving while using the occlusion visor). The use of
equipment and vehicle was demonstrated by the experimenter; the
demonstration included informing the subjects of emergency over-
rides for the visor and the availability of a trainer brake for use by the
researchers in the event of an emergency. The drivers then signed
informed consent statements.
After the experimenters drove each subject to a practice area, the
subjects moved to the driver’s position in the test vehicle. A researcher
sat in the back seat with the recording equipment and an assistant sat
in the front passenger seat. Wearing the occlusion visor and practic-
ing with the foot switch and emergency overrides, the drivers drove
around a course apart from the test setups to familiarize themselves
with the equipment and the vehicle.
After the drivers became relatively comfortable with the use of the
equipment (judged by their willingness to engage in casual conver-
sation while driving), they were directed to the starting point for the
test. The drivers were first asked to drive naturally through the set of
(a) experimental features, without instruction or use of the occlusion
device, to assess comfortable speeds through the features and ensure
a limited degree of familiarity with the test facility.
The drivers then drove continuously through each of the features,
with the occlusion system in place. The drivers controlled the occlu-
sion visor while driving on cruise control (set at 72 km/h) through the
test sections of the course, pushing the foot switch to obtain 0.5-s
glimpses of the roadway. The drivers had to release and then press the
switch for each glimpse; they could not obtain a sustained view by
holding down the switch. In sections of the course where unrestricted
vision was necessary for safety, the researchers cleared the visor.
They drove each test feature at least six times, three in each direction
where feasible.
On-Road Study
The same equipment and test vehicle used in the test track study were
used in a study conducted on local rural roadways. Because these runs
were on public highways, certain extra precautions were taken to pre-
(b) vent a mishap with a driver who was periodically “blinded.” In addi-
tion to the precautions considered on the test track (i.e., override
FIGURE 2 LCD visor in (a) clear and (b) opaque settings. switches available to the subject, the experimenter, and the assistant;
battery backup; and the trainer brake available to the assistant), a
safety driver was ahead of the instrumented car in constant cellular
mation and was computed as the ratio of the glimpse length divided telephone contact with the on-board safety person. Any advisory from
by the time elapsed from the last glimpse until the time of the present this car concerning opposing traffic or overtaking traffic, encroaching
request, or traffic, or other hazards resulted in terminating the run.
tglance length
VDi =
trequest − tlast request Drivers
where VDi is visual demand and t is time, in seconds. A sample of 6 volunteer drivers was selected from the 24 TTI employ-
This calculation provides a measure of the percentage of time a ees who had participated in the test track study. These subjects were
driver is observing the roadway at any point along the roadway. It not selected randomly, but rather were those drivers judged best
can readily be seen that a given observation’s value increases as the qualified to drive in potentially hazardous on-road tests, based on
time between successive glances decreases, and vice versa. The more their confidence and skill as observed during closed-course maneu-
information the driver needs to carry out his or her control function vers. The sample set included equal numbers of men and women and
in driving, the more often the scene ahead must be sampled, leading included representation from each of the three age groups used in
to the inference that the mental workload increases. the test-track study.
The researchers began the testing program by giving the drivers Two rural two-lane highways within 30 min driving time from
instructions that provided information on the nature of the experi- Bryan–College Station were used. Construction plans detailing the
74 Paper No. 00 - 0596 Transportation Research Record 1737
alignments of these highways were obtained from the Texas Depart- • A test track study examining single curves,
ment of Transportation. Horizontal curves with parameters similar • A test track study examining paired curves, and
to those created at the Riverside campus were found and tested, • An on-road study examining single curves.
although exact matches to the test-track curves were not found. The
facilities varied in cross section from the test track, including bor- The studies examining individual curves sought to determine the rela-
row ditches and other restricting components, and they did include tionship between driver visual demand and geometric characteristics
vertical curves, although the curves tested did not exhibit sight dis- of roadway curves; the study examining paired curves sought to
tance restrictions because of the relatively modest vertical curvature examine issues related to visual demand carryover and expectancy.
present at the test curves. These elements were wholly lacking at the
Riverside campus test facility.
It was not possible to locate curves in the field that were exact dupli- Test Track: Single Curves
cates of those tested at the Riverside campus test facility; instead,
researchers concentrated on locating curves that covered a substan- In this examination of the effects of individual curves, researchers var-
tial range of the same individual curve experiment space. Figure 3 ied the basic geometric parameters of the curves to assess their effects
provides a comparison of the two experiments. on visual demand. Radius and deflection angle were systematically
varied, and the resulting effects on visual demand were analyzed.
As discussed previously, visual demand was measured to examine
Procedure driver workload. Essentially a measure of the proportion of the time
that a driver was looking at the road, the value visual demand pro-
An initial run with each driver was made through the course. This run vided a way to construct a profile of the visual demand for the driver
was accomplished without closing the occlusion device and without as the roadway was driven. Several methods for examining the data
speed control. The driver then drove with the occlusion device not were explored, including averaging visual demand over the length of
occluded back to the starting point. He or she next drove through the the curve, averaging visual demand over a portion of each curve (i.e.,
features of interest, with cruise control set and the occlusion system the first half, etc.), and averaging visual demand over a fixed length in
working in the driver-controlled mode of operation. each curve [i.e., the first 30 m (98.4 ft)]. Each of these methods was
With the richness of cues in the actual driving environment, only a considered for analysis; each had its advantages and disadvantages.
limited number of passes through the features were possible before Averaging the visual demand over the length of the curve provides
the driver was familiar with each feature and thus visual demand lev- a measure of workload that is intuitively appealing and that provides
els were flattened out. Four test runs were completed with each driver a value representative of the entire curve (denoted VDL). A problem
through each of the on-road test curves. with this measure, however, is that visual demand reaches a peak
near the start of curves (i.e., near the point of curvature) and then
declines to a much lower value for the remainder of the curve. If two
RESULTS curves of unequal length are compared, the lengths of the curves are
confounded with the magnitude of the estimate of VDL. In other
In this study, several different independent and dependent variables words, the length is directly proportional to the product of the radius
were analyzed in an attempt to provide a better understanding of and the deflection angle, thus averaging in a larger portion of the
the effects of roadway geometry on visual demand. Visual demand curve with a lower visual demand “tail” for the longer curve. This
[defined in Equation 1] served as a means of examining driver work- method does provide an overall impression, however, of how visual
load. As evaluated in the study, visual demand represents a measure demand varies for different combinations of roadway geometry, and
of one attribute of workload: input load. This measure examined the it was judged acceptable for comparisons between identical curves
amount of visual information that the driver believed was needed to in the paired curve study.
drive various types of curves. Another measure explored was averaging visual demand over
The results of the study have been separated into three sections, the first 30 m of the curve (denoted VD30). It provides a fixed com-
corresponding to the three major studies: parison between curves that is not confounded with the measure-
ment used. This measure gives a better understanding of the effect for driver variability (blocking factor) and learning effect (repeated
of the roadway geometry in or near the peak visual demand, although measures). Similar to the finding for VDL, the learning effect was
the absolute peak associated with a particular run might or might not found to be in run 1 only, and no statistically significant difference
be included in the 30 m. In comparisons of the 30-m segments imme- was found between runs 2 and 6. In light of the run effect and based
diately before and immediately after the point of curvature, or begin- on the assumption that the first run is more likely to represent real-
ning of the curve, approximately 80 percent of the tests indicated life situations, it was decided to rerun the ANOVA with run 1 data
that the peak occurred in the section immediately after the point of only. Although the researchers generally accepted the premise that
curvature with no systematic differences. the data from the first run more accurately represent real-life situa-
The dependent variables considered in this analysis included aver- tions, an alternative view was that runs 2 through 6 might better rep-
age VDL and VD30. Independent variables considered included the resent real-life situations because of increased familiarity with the
reciprocal of radius (1/R), deflection angle (DA), interaction between experimental apparatus. Because of this alternative interpretation,
1/R and DA, and run number. results from runs 2 through 6 are also presented. Average VD30 for
Of the 24 drivers who participated, 21 drivers drove six individual run 1 was approximately 5.5 percent higher than that for runs 2
test curves, each six times (i.e., six runs per driver per curve). Two through 6. Deflection angle was not a significant factor ( p = 0.293)
drivers drove five of the six curves, and one drove only one curve. during run 1 but was significant ( p < 0.0001) during runs 2 through
The analysis was performed without the latter driver. Thus, 23 driv- 6. Also, the R2 value for run 1 is somewhat higher than that for runs
ers provided a total of 810 measurements, where each measurement 2 through 6 (70.5 percent versus 53.0 percent, mainly because of the
represents the visual demand for a given driver in a given curve for larger sample size for runs 2 through 6).
a given run.
Initial analyses were completed by using VDL as the dependent
variable. Using a repeated-measures analysis of variance (ANOVA), Test Track: Paired Curves
the inverse of radius ( p < 0.0001), deflection angle ( p = 0.046), and
The 24 drivers tested pairs of curves for which the distance (61 m or
the interaction of the inverse of radius and deflection angle ( p =
274 m) between otherwise identical curves and the direction of the
0.0017) as independent variables was found to be significantly
second curve relative to the first curve (S-curve or broken-back curve)
related to VDL. The run effect was also statistically significant ( p =
were varied. The curves tested had radii of 145 m and deflection
0.016), an indication that the repeated driving of the curves had an
angles of 45°.
effect on VDL.
Visual demand was calculated over the entire second curve of each
In light of the run effect and based on the assumption that the first
curve pair for six runs per driver. The resulting data were analyzed
run is more likely to represent real-life situations, it was decided to
by means of repeated-measures ANOVA to estimate the effect of
rerun the ANOVA by using run 1 data only. On the basis of this
direction and type of curve on visual demand. To account for driver
reduced data set, the DA ( p = 0.076) and the interaction R × DA ( p =
variability, a blocking factor was included in the models, as was done
0.53) were no longer significant. The final model including 1/R only
previously. Data from driver 35 were excluded because that driver
was significant ( p < 0.0001) with an R2 value of 71.
tested only one type of curve pair. Thus, a total of 524 observations
Although the researchers generally accepted the premise that the
were available.
data from the first run more accurately represent real-life situations,
The effect of direction of curve and separation of curves and their
an alternative view was that runs 2 through 6 may better represent
interaction on visual demand was estimated as well as that of the
real-life situations because of increased familiarity with the experi-
six repeated runs. The run effect was significant ( p = 0.0321), with
mental apparatus. Because of this alternative interpretation, results
run 1 resulting in slightly higher VDL (by approximately 5 percent)
from runs 2 through 6 are presented. The following ANOVA results
than runs 4 through 6; no significant difference was found between
were found based on the data from the last five runs. There was no
runs 1, 2, and 3 or between runs 2 through 6, respectively. Thus,
longer a run effect (p = 0.91) on VDL; that is, the learning curve after
although there was not as clear a learning pattern over time as that
run 1 stabilized. Furthermore, deflection angle was no longer signifi-
found previously, the results were analyzed separately for run 1 and
cant ( p = 0.103), although the interaction term, (1/R) × DA, remained
runs 2 through 6, as was done in the analysis of VDL for single curves.
significant (p = 0.002). A model with 1/R but without the interaction
The run 1 ANOVA showed the following: neither direction of
(1/R) × DA term was rerun. The R2 value dropped slightly, from 53.8
curve ( p = 0.298) nor separation of curves ( p = 0.828) had a signifi-
to 52.6, with a highly significant radius effect ( p < 0.0001) and driver
cant effect on VDL. However, their interaction was significant ( p =
effect ( p = 0.0015).
0.049), showing that S-curves (least squares mean VDL = 0.492)
Analyses also were completed by using VD30 as the dependent
resulted in significantly higher VDL (by approximately 8.8 percent)
variable. A repeated-measures ANOVA was performed with the
than broken-back curves (least squares mean VDL = 0.452), each with
independent variables radius, deflection angle, and the interaction of
a 61-m separation. No other comparisons were statistically significant.
radius and deflection angle. Drivers were included in the ANOVA as
The model’s R2 value was 75.3.
a blocking factor to account for variability among the 23 drivers. The
Only curve separation had a significant effect ( p = 0.002) on VDL
six runs were used as repeated measures to estimate the effect of learn-
when analyzing the results of runs 2 through 6. However, the relative
ing with consecutive driving of the curves. Radius was the single increase in VDL from 61-m separation (least squares mean VDL =
most significant factor affecting visual demand ( p < 0.0001). The 0.440) to 274-m separation (least squares mean VDL = 0.459) was
interaction between radius and deflection angle was not significant only 4.3 percent. The model’s R2 value was 64.4.
( p = 0.055) and was subsequently dropped from the model.
A significant learning effect ( p = 0.031) was observed, with run
1 VD30 significantly higher (by approximately 6 percent) than VD30 On Road: Single Curves
for runs 2 through 6. The R2 for this model was 56.5.
The VD30 data were analyzed by means of regression analysis Six subjects selected from among the 24 who took part in the test
using 1/R, deflection angle, and their interaction, while accounting track experiment also participated in testing curves on two local rural
76 Paper No. 00 - 0596 Transportation Research Record 1737
roadways. A regression analysis using 1/R, DA, and their interaction interaction between radius and deflection, these contributions
was performed within a repeated measures approach, including a were generally quite low and provided little practical change in the
blocking factor for driver on the five test curves. Run was found to predictive model over a model based solely on radius.
be a significant factor ( p ≤ 0.0001). As before, run 1 VDL was signif- The models developed had relatively high R2 values, indicating that
icantly higher than runs 2 through 4 VDL, by approximately 10 per- the models explained much of the variation observed. These R2 val-
cent. The regressions were performed separately for run 1 and runs ues were calculated after the variation among drivers was accounted
2 through 4. In the final models, neither the interaction between for by the model. It was found that visual demand was significantly
radius and deflection angle nor the deflection angle was significant. higher on curves with sharper radii.
In addition, the variability among drivers was not significant ( p = An on-road testing program provided researchers with the oppor-
0.76) when compared with the overall error. Thus the blocking fac- tunity to confirm whether the relationships discerned in the test track
tor (driver) was excluded from the model for run 1. The only signif- environment could be observed in the real world of actual highways.
icant factor for run 1 and runs 2 through 4 was 1/R. The model developed is shown in Figure 5, where it may be compared
with that developed in the test track effort. As with the test track
results, the only significant relationship between visual demand and
SUMMARY OF FINDINGS curve geometry was based on radius.
consistent with driver expectations (and be associated with lower ment on visual demand would also enhance the possible application
workload) and that more closely spaced curves would impose a of driver workload measures.
greater workload. The VDL changes observed were relatively small,
however, and further research would need to be conducted to confirm
or extend these results. ACKNOWLEDGMENTS