Sunteți pe pagina 1din 9

Accident Analysis and Prevention 45 (2012) 296–304

Contents lists available at ScienceDirect

Accident Analysis and Prevention


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

Modeling the impact of spatial relationships on horizontal curve safety


Daniel J. Findley a,∗ , Joseph E. Hummer b , William Rasdorf c , Charles V. Zegeer d , Tyler J. Fowler e
a
Institute for Transportation Research and Education, North Carolina State University, Centennial Campus Box 8601, Raleigh, NC 27695-8601, United States
b
North Carolina State University, 423 Mann Hall, North Carolina State University, Raleigh, NC 27695-7908, United States
c
North Carolina State University, 200 Mann Hall, North Carolina State University, Raleigh, NC 27695-7908, United States
d
Engineering and Planning, UNC Highway Safety Research Center, 730 Martin Luther King, Jr. Blvd., Chapel Hill, NC 27599-3430, United States
e
Institute for Transportation Research and Education (ITRE), North Carolina State University, Centennial Campus Box 8601, Raleigh, NC 27695-8601, United States

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

Article history: The curved segments of roadways are more hazardous because of the additional centripetalforces exerted
Received 24 March 2011 on a vehicle, driver expectations, and other factors. The safety of a curve is dependent on various factors,
Received in revised form 20 July 2011 most notably by geometric factors, but the location of a curve in relation to other curves is also thought to
Accepted 24 July 2011
influence the safety of those curves because of a driver’s expectation to encounter additional curves. The
link between an individual curve’s geometric characteristics and its safety performance has been estab-
Keywords:
lished, but spatial considerations are typically not included in a safety analysis. The spatial considerations
Horizontal curve
included in this research consisted of four components: distance to adjacent curves, direction of turn of
Safety
Spatial
the adjacent curves, and radius and length of the adjacent curves. The primary objective of this paper
Collision is to quantify the spatial relationship between adjacent horizontal curves and horizontal curve safety
Radius using a crash modification factor. Doing so enables a safety professional to more accurately estimate
safety to allocate funding to reduce or prevent future collisions and more efficiently design new roadway
sections to minimize crash risk where there will be a series of curves along a route. The most important
finding from this research is the statistical significance of spatial considerations for the prediction of hor-
izontal curve safety. The distances to adjacent curves were found to be a reliable predictor of observed
collisions. This research recommends a model which utilizes spatial considerations for horizontal curve
safety prediction in addition to current Highway Safety Manual prediction capabilities using individual
curve geometric features.
© 2011 Elsevier Ltd. All rights reserved.

1. Introduction 2001; AASHTO, 2010). This paper seeks to explore and quantify the
impact of spatial relationships on horizontal curve safety.
A road essentially consists of two types of segments: straight The goal of standardizing safety prediction of roadways in the
segments and curved segments. Curves, particularly those with United States led to the development of the Highway Safety Manual
curves along the horizontal plane, are more hazardous to drivers (HSM) which can predict the safety of a curve based on its phys-
because of the additional centripetalforces exerted on a vehicle ical characteristics (AASHTO, 2010). However, safety analysis has
passing through the curve and the more difficult driving tasks traditionally considered each curve as an isolated highway feature,
required of the driver (Hummer et al., 2010a). Horizontal curves are without considering other nearby roadway and roadside features.
a primary area of focus for transportation agencies to improve the This presents a prime opportunity to explore spatial considera-
safety of their roadway network. The safety of a curve is dependent tions with a consistent and widely accepted analysis strategy for
on various factors, most notably geometric factors of the individual individual curve characteristics. Another reason for assessing the
curve, but the location of a curve in relation to other curves is also inclusion of spatial considerations is to incorporate the concept of
thought to influence safety (Glennon, 1987). The link between an maintaining a driver’s expectations which is a practice of highway
individual curve’s geometric characteristics and its safety perfor- designers. Typically, a road which violates a driver’s expectations
mance has been established (Zegeer et al., 1991; Strathman et al., will be more hazardous than a road which does not. In relation to
curves, a driver who experiences a set of curves will expect to expe-
rience additional curves on the route, while an isolated curve can
∗ Corresponding author. Tel.: +1 919 515 8564; fax: +1 919 515 8897.
be hazardous because of the disruption to a driver’s expectations.
E-mail addresses: Daniel Findley@ncsu.edu (D.J. Findley), hummer@ncsu.edu
In a similar manner, driver alertness can also be impacted by hor-
(J.E. Hummer), rasdorf@ncsu.edu (W. Rasdorf), Charlie Zegeer@unc.edu izontal curves; where multiple, closely spaced horizontal curves
(C.V. Zegeer), tjfowle2@ncsu.edu (T.J. Fowler). demands more effort from a driver and can keep a driver more

0001-4575/$ – see front matter © 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.aap.2011.07.018
D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304 297

alert on the roadway. The opposite can be true for long tangent percent of vehicles exceeding the design speed decreased dras-
sections between curves; where a driver’s alertness can wane and tically) and increased the margin of safety (through a greater
the approaching curve might become a safety hazard for the driver. difference between the required side friction and provided side
Longer tangent sections between adjacent curves can also lead to friction), while decreasing actual reported collisions.
higher speeds which can lead to more severe collisions. The Highway Safety Manual (HSM) provides a model to pre-
The inclusion of spatial relationships in a safety analysis can be dict the safety performance of two-lane rural highways based on
an important consideration for a more accurate and comprehensive the following characteristics: traffic volume, lane width, shoul-
examination of actual safety performance. Thus, to improve perfor- der width, length, radius, superelevation, grade, driveway density,
mance this paper investigates a curve’s spatial relationship relative roadside hazard rating, spiral transition, passing lanes, roadway
to adjacent curves, including the distance to adjacent curves, the lighting, centerline rumble strips, two-way left turn lanes, and
direction of turn of the adjacent curves, the radius of the adjacent automated speed enforcement (AASHTO, 2010). The HSM consists
curves, and the length of the adjacent curves. Vertical curvature of safety performance functions (SPFs) (which determine the pre-
and roadway grades are also important features which can influ- dicted crash frequency of a tangent roadway segment under base
ence the safety of a horizontal curve or a series of curves. However, conditions) and crash modification factors (CMFs) (which adjust
vertical alignment was not considered in this study, due to a lack the base conditions assumed in the SPF to model site specific condi-
of available vertical alignment data. The objective of this paper is tions). The two-lane highway SPF, shown in Eq. (1), was developed
to quantify the spatial relationship of nearby horizontal curves on from 619 rural two-lane highway segments in Minnesota and 712
horizontal curve safety for rural, two-lane roads so that a safety segments in Washington (Harwood et al., 2000). The HSM CMF
professional can accurately estimate safety and effectively allo- for horizontal curve alignment, shown in Eq. (2), was developed
cate funding. A curve within a series of curves might be expected from curve data in the state of Washington using crash data for
to experience fewer collisions than an isolated curve with simi- 10,900 curves from 1982 to 1986 (Zegeer et al., 1992). The horizon-
lar geometric characteristics, no matter how severe its geometric tal alignment CMF represents the ratio of expected curve collisions
characteristics. to expected collisions on a similar two-lane tangent section. For
instance, a curve with a length of 500 ft, a radius of 500 ft, and no
spiral transitions would have a horizontal alignment CMF of 2.09,
2. Literature review
meaning that this curved section is expected to have 2.09 times
the amount of collisions of a comparable tangent section. In addi-
Although much attention has been focused on the safety of a
tion to the horizontal alignment CMF, this study accounted for the
curve through the examination of an individual curve’s attributes,
other factors presented in the two-lane highway methodology in
spatial considerations in relation to the nearby roadway features
the HSM.
have not been widely considered in the literature. Therefore, this
analysis will not focus on the characteristics of an individual curve SPF = AADT × L × 365 × 10−6 × e−0.312 (1)
and their impact on safety, but rather, on the impact of spatial
relationships with other curves on the safety of a particular curve. where SPF = predicted total crash frequency per year for roadway
segment base conditions, AADT = average annual daily traffic vol-
ume (vehicles per day), L = length of roadway segment (miles)
2.1. Safety considerations  80.2 
(1055 × Lc ) + R
− (0.0012 × S)
Numerous horizontal curve countermeasures have been a focus CMF = (2)
(1.55 × Lc )
of transportation professionals and researchers because of the fre-
quent and severe nature of collisions at curves. One impetus for where CMF = crash modification factor for the effect of horizontal
further work is a study of safety management conducted in Norway alignment on collisions, R = radius (feet), Lc = length of horizontal
that found horizontal curve treatments provide more benefits than curve (miles), including spiral transitions if present, S = spiral tran-
they cost with a benefit-cost ratio of 2.37 (Elvik, 2007). A number sition presence: 1 if present, 0 if not present, 0.5 if present on one
of studies have focused on the relationship between an individual end only.
horizontal curve’s characteristics and the safety performance of the
curve, including design attributes (Zegeer et al., 1991; Strathman 2.2. Spatial considerations
et al., 2001), and signage and markings (Retting and Farmer,
1998; Eccles and Hummer, 2000; Hammond and Wegmann, 2001; Glennon et al. (1985) studied 3304 curved segments and 253
Chrysler et al., 2005; Lyles and Taylor, 2006; Charlton, 2007). Other tangent segments on rural, two-lane roads in Florida, Illinois, Ohio,
efforts have created guidelines for signing and marking on hor- and Texas. The researchers examined 0.6-mile (1 km) long road-
izontal curves (Nielsen et al., 1999; Herrstedt and Greibe, 2001; way segments which included horizontal curves (with an average
Bonneson, 2007) and strategies for improving the safety of hori- length of 0.17 miles) and the adjacent tangent lengths which are
zontal curves (Torbic et al., 2004; McGee and Hanscom, 2006). thought to be influenced by the curve. Glennon studied only iso-
Wong and Nicholson (1992) studied driver behavior at hor- lated curves by setting a minimum tangent length between curves
izontal curves through vehicle movement tracking using video of 656 ft (200 m). The analysis found that curves experience three
recordings on three curves in New Zealand that were realigned times the rate of collisions that tangent segments do.
due to safety concerns. The video tracking allowed vehicle speeds In a synthesis of research, Glennon (1987) identified four cat-
and paths to be determined. Each of the curves was lengthened (to egories of elements relating to the safety of horizontal curves:
approximately twice the length of the original curve) and recon- horizontal alignment, cross section, vertical alignment, and other.
structed with a larger radius (to approximately three times the The horizontal alignment elements describe the longitudinal fea-
radius of the original curve). The study found that there is a consid- tures of horizontal curves, such as radius, length, and sight distance.
erable difference among drivers of the speed and path they take on The cross section elements define the features of a roadway which
horizontal curves. Therefore, the design radius and radius driven by are perpendicular to the direction of travel, such as the lane width,
the vehicle can vary considerably which can alter the forces expe- shoulder width, and cross-slope of the roadway. The vertical align-
rienced by the individual vehicle. In general, this study found that ment elements specify the relationship between horizontal curves
the realignment of a curve increased vehicle speeds (although the and vertical elements (grades and curves). Among the remaining
298 D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304

elements, the distance to adjacent highway curves, intersections, Another research effort focused on developing a two-lane rural
and bridges were the most relevant as spatial considerations for roads safety evaluation process that incorporated spatial features
horizontal curves. Although spatial considerations were identified in the safety criteria (Lamm et al., 2002). The criteria were used
as relevant to horizontal curve safety, no research had been con- to classify the safety of roadways sections as good, fair, or poor.
ducted prior to the synthesis report by Glennon which was able to Two spatial criteria were included in the process: (1) a design con-
quantify the effect of distance of nearby curves on collision risk. sistency criteria which desires differences of less than 10 km/h
Aram (2010) considered spatial elements as influential in curve between the 85th percentile speed and the design speed along
safety based on the research by Glennon. His individual curve the entire segment of analysis and (2) an operating speed consis-
study focused on radius, degree of curve, deflection angle, curve tency criteria which desires differences in 85th percentile speed
length, lane width, roadside clearance, signage, markings, pave- of less than 10 km/h between successive roadway elements (curve
ment condition, and operating speeds. Aram found that the most to curve or tangent to curve). The safety evaluation criteria found
significant variables for horizontal curve collisions were degree of similar conclusions in approximately 75% of analyses of collisions
curve, length of curve, superelevation, length of spiral transition, at 236 roadway sections with a total length of 490 km, showing
shoulder width, and traffic volume. The author also noted that the that the safety of a two-lane rural road could be estimated with
collision difference between curves and tangents was apparent at measures other than collision history.
radius values less than 3000 ft and was particularly significant at The Manual of Uniform Traffic Control Devices (MUTCD) has a
radius values less than 600 ft. spatial guidance component related to horizontal curves through
Zegeer et al. (1991) reported that the effect of distance between the application of horizontal alignment warning signs (FHWA,
curves on collisions was a gap in the existing curve safety knowl- 2009). The MUTCD recommends warning signs and advisory speed
edge. The researchers identified previous studies using surrogate warning signs in advance of curves when the difference between
measures which had differing conclusions about the safety impacts the speed limit and advisory speed is 5 mph or more and requires
of the distance between curves (Terhune and Parker, 1986; Datta the signs when the difference is 10 mph or greater. Chevrons or one
et al., 1983). Despite the contradictory surrogate studies, Zegeer large directional arrow are optional at a 5 mph difference, recom-
et al. noted that long tangents sections would make drivers less mended at a 10 mph difference, and required at a 15 mph difference
prepared to safely traverse a horizontal curve than a curve of sim- between the speed limit and advisory speed. The guidance provided
ilar characteristics in a series of curves, due to driver expectations. by the MUTCD promotes a spatial examination of curves and their
Furthermore, the researchers discussed the potential for traffic con- surroundings by comparing the speed limit of the roadway to the
trol countermeasures if curves at the end of long tangent sections recommended advisory speed of the curve.
truly are more hazardous than other curves. Zegeer et al. attempted Guo et al. (2010) incorporated spatial considerations into the
to include the distance to adjacent curves in the model but did not safety modeling of signalized intersections for 170 intersections
find these spatial factors to be significant in their analysis. However, along 25 corridors in central Florida. The spatial consideration
when the authors used a categorical variable (distance less than or utilized in their equations was the distance between the intersec-
greater than 0.3 miles) to describe the distances, the variable was tions. A negative binomial model was recommended among models
found to be marginally significant (p-value = 0.06). The variable was tested which included Poisson and negative binomial models and
not included in the final model, but the authors noted that there variations of those models. The model with spatial components was
appeared to be some evidence that longer tangents might result in found to be superior to other models by providing a better repre-
more collisions. sentation of intersection safety. According to the authors, exclusion
Milton and Mannering (1998) developed a predictive model of the spatial factors would lead to biased inference in the model
for estimating collisions based the geometrics of highways and and incorrect conclusions.
traffic-related characteristics with 31,306 observations in Wash-
ington. Among other geometric and traffic-related characteristics, 2.3. Summary
the authors included two components related to the spatial char-
acteristics of horizontal curves: (1) the tangent length (total length The majority of previous literature has relied on an individual
between two adjacent horizontal curves) and (2) a tangent-curve curve’s attributes, not spatial considerations in relation to nearby
indicator variables (which is input as a 1 if the curve radius less roadway features that could influence safety. However, spatial con-
than 2850 ft and the tangent length between adjacent curves is siderations have been identified as a feature which could impact
more than half a mile, or is input as 0 otherwise). The study found horizontal curve safety. This paper presents a procedure that uti-
that collision frequency increases as the tangent length increases, lizes the HSM methodology as a foundation for the analysis of curve
due to driver expectations of encountering additional curves and safety and incorporates spatial characteristics of nearby curves to
higher speeds on the longer sections. The tangent-curve indicator assess their influence on the safety of a curve.
variable also showed that curves with longer tangent lengths expe-
rience higher collision frequencies, particularly in the set of curves 3. Methodology
with longer tangent lengths (over half a mile) and smaller radii (less
than 2850 ft). This study consisted of three major components: horizontal
A roadway design’s conformance to the expectations of drivers curve data collection, HSM safety prediction, and an analysis of
can be explained through the concept of design consistency spatial considerations created by adjacent curves. Horizontal curve
(Fitzpatrick et al., 2000). A typical way to measure design con- data collection was performed to obtain the missing curve char-
sistency is through the measurement of operating speeds on a acteristics that constitute the basis for further analysis. The HSM
roadway. Changes in operating speeds can indicate a potential vio- safety prediction utilized a contemporary safety model to estimate
lation of a driver’s expectations and, therefore, might be a site with the safety of the curves. The spatial relationship analysis combined
increased crash risk and therefore, for potential countermeasures data from both of the previous steps to develop a more accurate
to improve the roadway. Fitzpatrick et al. studied operating speeds and comprehensive safety prediction model. The overall effort was
on horizontal curves and developed a model for predicting the to compare HSM predicted collisions to reported collisions with
85th percentile operating speed based on the radius of the curve respect to spatial considerations.
for inclusion in the design consistency module of the Interactive This study included two two-lane roads (NC42 and NC96) which
Highway Safety Design Model (IHSDM). are predominately rural and reside in central and eastern North
D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304 299

Carolina. The analysis included the entire length of each route miles) to determine the predicted crash frequency using default
except where a higher order route (i.e., US route) ran concurrently base conditions.
with the NC route. The entire route of NC42 is 223 miles long; the The resulting collision frequency from the SPF is multiplied
analysis sections included 168 miles of the route containing 246 by appropriate CMFs to adjust the collision prediction based on
curves and traversing east and west through 11 counties. The entire the specific roadway characteristics of the segment (lane width,
route of NC96 is 107 miles long; the analysis sections included shoulder width, radius, super elevation, grade, driveway density,
95 miles of the route containing 174 curves and traversing north roadside hazard rating, spiral transition, passing lanes, roadway
and south through 5 counties. The collision data were supplied by lighting, centerline rumble strips, two-way left-turn lanes, and
the North Carolina Department of Transportation and included the automated speed enforcement). A CMF less than one means that
milepost location of the collision, collision severity, and the county roadway feature creates a safer road and therefore, the predicted
location of the collision. Five years of collision data were stud- collision frequency is reduced, while a CMF greater than one means
ied (January 1, 2005 to December 31, 2009) with 4505 reported that crash frequency will increase as a result of the treatment.
collisions, including all types of collisions. Research by Hummer et al. (2010b) found that the most sensi-
tive inputs for horizontal curves are length, traffic volume, lane
3.1. Horizontal curve data collection width, and radius. The sensitivity of these elements was based on
a comparison of the number of predicted collisions when individ-
Curvature Extension is a GIS based application which uses line ual curve data was used for each curve versus an average value for
work to derive horizontal curve geometric data, including the each element being applied to each curve. The analysis found a +4%
length and radius of a curve. Curvature Extension determines the change in predicted collisions for length, +6% for radius, −7% for
radius of a curve by creating a circular arc utilizing the chord AADT, and −7% for lane width when the average value for the ele-
length, chord angle, and length of the curve along the route. ment was utilized in the prediction model instead of the actual field
The curve length is calculated based on the end points selected measured value for each individual curve. For instance, this sensi-
by the user. Curvature Extension can be added as a toolbar in tivity analysis found that if average input values for radii were used
ArcGIS software and was developed by the Florida Department in the analysis, the number of predicted collisions when utilizing
of Transportation FDOT, 2010). A study of three comparable GIS individual curve data was 6% greater than the number of predicted
applications, including Curvature Extension, found that Curvature collisions had an average radii been applied to all curves.
Extension outperformed the other two programs under theoretical Therefore, site specific values were used in the HSM model
and real-world conditions (Rasdorf et al., in press). The theoreti- for the sensitive inputs which were available without a field visit
cal assessment found that Curvature Extension was able to report (radius, length, and AADT) and two-lane rural road average values
the exact radius of various precisely drawn curves regardless of were used for all other inputs (shoulder width, lane width, superel-
the number of GIS points tested in the analysis. The real-world evation variance, grade, driveway density, and roadside hazard
assessment found that Curvature Extension was capable of return- rating). Among the four most sensitive parameters, an average
ing the radius of the curve within 50% of the field measured value value was needed for lane width due to a lack of data, which is
on 80% of the curves included in the study, the radius of the curve a reasonable methodology because of a relatively small variance of
within 25% of the field measured value on 45% of the curves, and 0.7 ft2 with a mean of 10.5 ft. The radius and length values were
the radius of the curve within 10% of the field measured value on determined from the horizontal curve data collection through the
24% of the curves. These accuracies are reasonable because many execution of Curvature Extension on GIS line work data and the
two-lane rural curves were not designed using current standards AADT values were collected from traffic volume maps. The average
with consistent radii throughout the curve and representing those values were collected from field measurements at 51 horizontal
situations with a single radius value inherently adds variability to curve sites dispersed throughout central and eastern North Car-
the results. olina. Average values included lane width of 10.5 ft, inside shoulder
For this study, the horizontal curve data collection procedure width of 6 ft, outside shoulder width of 8 ft, no superelevation vari-
utilized the methodology proposed by Rasdorf et al. (in press) by ance, level grade, driveway density of 9.6 driveways per mile, and
executing Curvature Extension on GIS line work data to obtain the a roadside hazard rating of 4.
required horizontal curve spatial data. Data collection consisted of The HSM predictive model was calibrated using local conditions
two processes to obtain the curve data: (1) a manual, visual curve such as climate, driver population, and crash reporting thresholds
identification of the beginning and ending points of the curve by (which are likely different in North Carolina than in Washington
the user through the delineation of tangent sections and the ter- where the SPF was developed). The calibration factor adjusts the
mini of the curves and (2) a geometric characterization (radius and predicted collision frequency for elements not covered by CMFs to
arc length) completed by the program. Horizontal curve spatial data convert the national prediction to a local prediction. A calibration
are a critical component of the HSM safety prediction and the spa- factor of 1.33 is recommended for two-lane road curves in North
tial relationship analyses. Thus, obtaining quality spatial data in an Carolina (Hummer et al., 2010b). This factor accounts for all of the
automated fashion is critical and is the focus of the work by Rasdorf local conditions enumerated above.
et al. (in press).
3.3. Spatial relationship analysis
3.2. HSM safety prediction
The spatial relationship analysis consisted of three primary
The HSM safety model was used to predict the safety of a curve analyses: an investigation of descriptive statistics, a correlation
based on its characteristics (AASHTO, 2010). The HSM methodology analysis, and a regression analysis. The investigation of descrip-
is primarily applied through three steps: selection and application tive statistics and correlation analysis were principally useful for
of a safety performance function, application of crash modifica- comprehensively understanding the data and to provide a basis
tion factors, and calibration of predicted collisions. Each step is for the selection of variables to include in the regression analysis.
repeated for every roadway segment. The SPF used for this anal- Statistical testing was conducted using Stata software (Stata, 2010).
ysis is appropriate for tangent two-lane road sections and is shown The selection of the most appropriate model was an iterative
in Eq. (1) (for tangent sections only). The SPF relies on the AADT process which involved a close examination of the correlation coef-
volume (in vehicles per day) and on the length of the segment (in ficients to determine which variables were significantly correlated
300 D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304

to the observed collisions. Those significantly correlated variables Curve #2


were then subjected to regression analysis to determine the most
Y
appropriate regression model for the inclusion of spatial consid-
X
erations. The significance of individual variables in the model and
the appropriateness of Poisson, through the overdispersion param-
eter, were considered for choosing the recommended model. The Curve #3
(Proximal to Curve #2)
recommended model was also validated with a set of 25 randomly
selected curves in central North Carolina with similar characteris-
tics as the data set used to develop the model.
In North Carolina collisions are reported to the nearest 0.1 mile Curve #1
location along a roadway at a location known as a milepost. How- (Distal to Curve #2)
ever, collisions are not actually clustered at these incremental
Fig. 2. Spatial curve example layout.
locations. Instead, they are dispersed between mileposts. Curves,
on the other hand, are identified to a higher level of precision
than were collisions. Therefore, the ending and beginning locations Important considerations in the selection of potential variables
of a curve do not necessarily correspond to tenth mile mileposts. to include in a regression analysis include the required effort of data
This can create difficulty when performing the collision data anal- collection to populate the model and the availability of accurate
ysis. Some collisions might have actually occurred on the curve, and repeatable data collection procedures. The spatial characteris-
but for reporting purposes were listed at a milepost outside the tics included in this study consisted of distance to adjacent curves,
curve boundary because the location was closer to the milepost direction of turn of the adjacent curves, radius of the adjacent
outside of the curve, and vice versa. To overcome this potential curves, and length of the adjacent curves. The distance to adja-
inconsistency, collisions were distributed based on the interpolated cent curves consists of two distances, one for each of the adjacent
distance between the terminal points (beginning and end) of the curves (upstream and downstream of the curve of interest). Adja-
curve and the edge of the nearest milepost’s influence area (mile- cent curves are identified by their proximity to the curve of interest,
post ±0.05 miles). The mid-point was used for the analysis because as either the proximal curve (i.e., the adjacent curve which is clos-
the influence area around each milepost is effectively 0.1 miles est to the subject curve of interest) or distal curve (i.e., the adjacent
(0.05 miles on each side of the milepost or the mid-point between curve which is furthest away from the subject curve of interest).
each consecutive milepost). This approach to finding the number Fig. 2 shows an example consisting of three curves with straight
of collisions that occur within the limits of the curve is consistent segments between the curves. If Curve #2 was the curve of inter-
with the methodology employed by Zegeer et al. (1991) which was est, Curve #1 would be the distal curve, or farthest curve from Curve
based on analyses of various influence areas. #2, since the distance X is greater than Y. Similarly, Curve #3 would
An example of curve collision distribution by interpolating mile- be the proximal curve, or closest curve to Curve #2.
posts is shown in Fig. 1. shows mileposts, the curve locations, and The relative direction of turn of adjacent curves was another
the number of reported collisions attributed to each milepost. The spatial characteristic studied. The relative direction of turn of adja-
purpose of this analysis is to assign to the curve the appropriate cent curves defines whether the adjacent curves turn in the same
number of reported collisions attributed to each milepost by the direction or the opposite direction as the curve of interest. In the
police officer. Examining the collisions near the beginning of the example shown in Fig. 2, Curve #1 and Curve #3 each turn in the
curve, ten reported collisions were attributed to milepost 5.7. The opposite direction from Curve #2. The radius and length of each
beginning of the curve (milepost 5.68) is within the 0.05 mile influ- adjacent curve were also considered in the analysis.
ence area of milepost 5.7 which means that the actual number of A further examination of the characteristics of curves in a series
collisions assigned to the curve at this milepost should be adjusted was also conducted as part of the analysis. The definition of a curve
downward to 7 collisions (10 collisions times 0.07 miles/0.1 miles). as an isolated, serial, or leading curve depends on the distance
This calculation is consistent with the concept that the curve requirement. Distance thresholds of 1/8, 1/4, 1/2 and 1 mile incre-
occupies 70% of the influence area surrounding milepost 5.7 and ments were tested. An isolated curve is a curve in which neither
therefore, the curve is assigned 70% of the 10 total collisions at adjacent curve is within the threshold distance, a serial curve is one
the milepost. This concept assumes that the reported collisions in which both adjacent curves are within the threshold distance,
are evenly distributed throughout the influence area of the mile- and a leading curve is a curve in which only one of the adjacent
post and not clustered exactly at the milepost. For the collisions curves is within the threshold distance. Considering Fig. 2 as an
at the end of the curve (milepost 5.87), five reported collisions example, if the threshold was set at 1/4 mile and X was 0.3 miles
were attributed to milepost 5.9. However, only one of the collisions and Y was 0.2 miles, Curve #2 would be considered a leading curve.
attributed to milepost 5.9 should be assigned to the curve (5 col- The objective of the regression analysis was to develop an equa-
lisions times 0.02 miles/0.1 miles). Therefore, nine total reported tion which incorporates actual reported collisions as the dependent
collisions would be assigned to this curve (seven collisions from variable which can be explained by the number of HSM predicted
milepost 5.7, one collision from milepost 5.8, and one collision from collisions and quantifiable curve spatial characteristics. An equa-
milepost 5.9). tion incorporating collisions as a function of HSM parameters and

Beginning of Curve End of Curve


(Milepost 5.68) (Milepost 5.87)

Milepost 5.6 Milepost 5.7 Milepost 5.8 Milepost 5.9


(2 Collisions) (10 Collisions) (1 Collision) (5 Collisions)

Fig. 1. Collision interpolation milepost example.


D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304 301

spatial considerations, as a CMF, will provide a useful addition to within 1 mile. The average distance from the curve to the distal
the current safety methodology provided by the HSM without the curve is 0.66 miles, with 3% of the distal adjacent curves within
development of a complete new model. Inclusion of HSM collision 0.1 miles, 27% within 0.25 miles, 54% within 0.5 miles, and 83%
predictions in the model allows the user to utilize a widely accepted within 1 mile.
method for estimating the safety of a curve based on its geomet-
ric characteristics. The basis of the HSM safety model prediction 4.1. Model selection
is total collisions (not individual injury type or property damage
collisions). A correlation analysis established which variables were most
correlated with the desired dependent variable, observed col-
4. Results lisions. Among the spatial elements, correlations between the
observed collisions and distances to nearby curves were found to be
A close examination of the descriptive statistics reveals insight significantly positive, but curve directions were not significant. The
when selecting appropriate variables to include in a regression statistically significant positive values for the distance to nearby
analysis, for explaining results, for considering further analysis, and curves demonstrate that in general, as the distance between curves
for assessing the applicability to other locations. The following is increase, the number of expected collisions increases.
a summary of the most useful results of the descriptive statistics Numerous regression equations were tested and examined
investigation. A total of 420 curves were included in the analysis based on overall model statistics, the individual variables, parsi-
(246 on NC42 and 174 on NC96). The following points describe the mony, and the significance of the variables. The dependent variable
results. in the model was observed collisions and the independent variables
were HSM predicted collisions and the spatial factors shown in Eq.
• At 43% of the curves, the adjacent curves both turn in the opposite (3). Therefore, the HSM procedure should be conducted to deter-
direction (i.e., if the curve turns to the right, both of the adja- mine the predicted collisions for inclusion in the model to provide
cent curves would turn to the left). Only 8% of adjacent curves a more accurate prediction of observed collisions. To include HSM
both turn in the same direction as the curve of interest (i.e., if predicted collisions in an unmodified manner (i.e., coefficient = 1),
the curve turns to the right, both of the adjacent curves also turn the HSM predicted collisions term was input as an offset to the
to the right) leaving 51% of the adjacent curves turning in both model.
directions.
• For the definition of a curve as an isolated, serial, or leading curve Coli = HSMi × exp[B0 + (B1 × D) + (B2 × P) + (B3 × D × P)] + ui
a threshold of 1/4 of a mile provides a roughly equal split of the (3)
curves into the three categories (32% were isolated, 27% were
serial, and 41% were leading curves). Larger thresholds result in a where Col = observed collisions (collisions/year), HSM = HSM pre-
lower percentage of isolated and leading curves, while resulting dicted collisions (collisions/year), D = distance to distal adjacent
in more serial curves. curve (miles), P = distance to proximal adjacent curve (miles),
• The average curve length was approximately 750 ft, with 40% of Bi ’s = coefficients.
the curves being shorter than 500 ft long, 79% shorter than 1000 ft Observed collisions typically follow a Poisson or negative bino-
long, and 91% shorter than 1500 ft long. mial distribution; in this analysis, the data showed overdispersion
• The average radius was approximately 2500 ft, with 19% of the (with an alpha of 1.45 at <0.001 significance, that is indicative of an
curves having a radius less than 1000 ft, 54% having a radius less overdispersed distribution) which led to the selection of a negative
than 2000 ft, and 75% having a radius less than 3000 ft. binomial distribution instead of a Poisson distribution. The Pois-
• The average number of HSM predicted collisions was 0.36 colli- son distribution assumes that the mean and variance are equal,
sions per curve per year (for a total of 152 predicted collisions which was not accurate for this data. The distance to adjacent
per year on the full sample of 420 curves). Twenty six percent of curves and an interaction term combining the distances to the dis-
the curves have predicted collisions of less than 0.1 collisions per tal and proximal adjacent curves was used in the model. Each of
year, 60% have less than 0.25 predicted collisions per year, 81% the terms, except the distance to the proximal curve, were statis-
have less than 0.5 predicted collisions per year, and 94% have less tically significant (at the 0.05 level). Table 1 shows the variables
than 1 collision per year. for the final model resulting from the analysis. The recommended
• The average number of observed collisions was 0.50 collisions model has a pseudo R-squared value of 0.0082 which is lower than
per curve per year (for a total of 210 collisions per year), 37% of expected for many applications. However, with safety applications
the curves have observed collisions of less than 0.1 collisions per and the inherent randomness of collisions, lower R-squared val-
year, 59% have less than 0.25 predicted collisions per year, 74% ues are expected to result from the analysis. The model that was
have less than 0.5 predicted collisions per year, and 89% have less used to explain the relationship between curvature and safety was
than 1 collision per year. developed by Zegeer et al. (1991) and was utilized in the HSM
• The average distance from the curve to the proximal curve is has a pseudo R-squared of 0.351. Although it is not appropriate
0.26 miles, with 29% of the proximal adjacent curves within to compare R-squared values across models with different depen-
0.1 miles, 68% within 0.25 miles, 86% within 0.5 miles, and 97% dent variables and a different number of variables, it would be

Table 1
Variables of the model.

Variable (All distances and lengths in miles) B Standard error z

Distance to distal adjacent curve D 0.367** 0.368 2.37


Distance to proximal adjacent curve P 0.122 0.155 0.33
Distance to distal adjacent curve*
Distance to proximal adjacent curve D×P −0.255* 0.144 −1.77
(Constant) 0.116 0.158 0.74
*
Significant at <0.10.
**
Significant at <0.05.
302 D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304

Table 2
Model predicted collisions (collisions per year) by adjacent curve distance.

Distance to distal curve (miles)

0.3 0.6 0.9 1.2 1.5 1.8 2.1 2.4 2.7 3.0

Distance to proximal curve (miles) 0.3 0.08 0.11 0.13 0.16 0.19 0.22 0.25 0.27 0.30 0.33
0.6 – 0.10 0.12 0.14 0.16 0.19 0.21 0.23 0.25 0.27
0.9 – – 0.11 0.13 0.14 0.15 0.17 0.18 0.19 0.21
1.2 – – – 0.11 0.11 0.12 0.13 0.13 0.14 0.14
1.5 – – – – 0.09 0.09 0.09 0.08 0.08 0.08

Note: Assumptions include length of 500 , radius of 500 , AADT of 5000 vehicles per day, lane width of 10.5 ft, inside shoulder width of 6 ft, outside shoulder width of 8 ft, no
superelevation variance, level grade, driveway density of 9.6 driveways per mile, and a roadside hazard rating of 4.

Fig. 3. Validation of model with difference in collisions from reported collisions.

expected that the model by Zegeer et al. would have a higher HSM prediction follow a similar track. However, the model predic-
R-squared because of the more direct relationship between an indi- tion shows an improvement over the HSM prediction in six of the
vidual curve’s attributes and safety then spatial effects and safety. eight categories of reported collisions.
The importance of spatial factors is a key determining factor in the
adequacy of the model as are the significance of the parameters in 5. Conclusions
the model. The recommended model is shown in Eq. (4). This model
does not include any of the parameter used in the HSM methodol- The most significant finding from this research is the impor-
ogy because those attributes are captured through the inclusion of tance and significance of spatial considerations for the prediction
the HSM procedures for estimating collisions, which is multiplied of horizontal curve safety. This analysis found that some spatial
by the spatial features of the recommended model. relationships significantly impact the safety of a horizontal curve.
Specifically, the distance to adjacent curves was found to be signif-
Col = HSM[0.116 + (0.367 × D) + (0.122 × P) + (−0.255 × D × P)]
icant in the estimation of observed collisions. Table 2 shows that
(4) the more closely spaced curves have fewer predicted collisions than
those curves which are more distant to each other. These findings
where Col = observed collisions (collisions/year), HSM = HSM pre- are consistent with the concept of driver expectations used by high-
dicted collisions (collisions/year), D = distance to distal adjacent way designers which affirms that a road which violates a driver’s
curve (miles), P = distance to proximal adjacent curve (miles). expectations will likely have more crashes than a road which does
The relationship between distal and proximal curve distances not violate those expectations. With regards to curves, a curve that
is shown in Table 2 using assumed values for horizontal curve fea- is a part of a series of curves is expected to be (and turns out to be)
tures. The computations required to generate the model prediction safer than a curve which is isolated from other curves.
of collisions per year were based on a 500-ft radius, 500-ft length, The isolated or serial nature of a curve can be represented by
and an ADT 5000 vehicles per day. Using these values, the resulting the longer distance between curves, the distance to the distal curve
HSM prediction was 0.32 collisions per year. A comparable tangent component of the model. Once a driver enters a series of curves, the
section of roadway with all the same characteristics except with- driver expects to experience additional curves, while a driver on a
out the horizontal curvature would be expected to experience 0.15 long, straight segment of roadway can be surprised by a curve. How-
collisions per year. The table shows that collisions predicted by the ever, drivers may expect some straight segment length between
model increase as the distance to the distal curve increases and curves, represented by the shorter distance to the adjacent curves,
decreases as the distance to the proximal curve increases. the distance to the proximal curve. The straight segment between
curves allows a driver the time needed to perceive and react to the
4.2. Model validation upcoming curve. The negative coefficient of the interaction term
between the distal curve distance and the proximal curve distance
Fig. 3 shows the results from the validation of the model using represent the desired balance between curves which are too close
25 randomly selected curves in North Carolina, which were not and those that are too far from each other (Table 1).
used in the development of the model. Average differences between An important consideration for any collision is the determina-
the model prediction and observed collisions, as well as the HSM tion of its location. A key added difficulty for curve collisions is how
prediction and observed collisions, were computed. As expected, to identify whether or not the collision even occurred on a curve.
because the HSM methodology is part of the recommended model, A collision is not typically contained at one finite location; rather,
the average differences returned by the model prediction and the a collision typically has a point of origin and a final stopping point
D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304 303

when the momentum of the vehicles has ended. This characteristic 7. Future research
of collisions adds to the complexity of defining whether a collision
occurred on, started on, or was caused by a horizontal curve. This analysis focused on the safety effects of adjacent curves
Another consideration for identifying curve collisions involves on a curve of interest. However, a more comprehensive, network
locating the collisions in relation to horizontal curves. In this anal- based approach that includes the 3rd, 4th, 5th, etc. curves in addi-
ysis, the reported collisions were located by a milepost number to tion to adjacent curves could provide additional accuracy and lead
the nearest 0.1-mile, while the curves were located using GIS tech- to better safety improvement decisions. In addition, direction of
niques to a higher level of precision. Therefore, an interpolation travel could be incorporated as an additional variable. Thus, the
technique was used to distribute the collisions to each curve for model could include the distance to both the proximal and distal
the nearest milepost outside and inside the curve boundaries. If one adjacent curves as well as a factor to account for travel direction
did not use the interpolation and curve collisions included only the which could be used to estimate the risk of a driver in a particular
reported collisions bounded by the 0.1-mile curve limits, the num- direction. Directional analysis could be useful for selecting coun-
ber of curve collisions would be underestimated. This could mean termeasures, such as advanced warning signs, which are specific to
that a hazardous curve is overlooked and is not investigated for one direction. Another element of interest for future research is the
improvements. Similarly, collisions reported as being bounded by spatial correlation among adjacent curves which would consider
the curve limits plus an additional fraction of a milepost distance to the directional information along with the adjacent curve that was
reach the next full milepost (to incorporate the next milepost out- traversed immediately prior to the collision. For the spatial consid-
side of the curve) could overestimate the total reported collisions erations in this paper, the average distance from the curve to the
assigned to the curve. proximal curve was approximately one quarter of a mile and the
Analysts could consider an additional distance beyond a curve average distance from the curve to the distal curve was 2/3 of a mile.
as a curve influence area. The influence area of a curve could be Further data collection at additional curves that are spaced more
expected to extend beyond its physical boundaries for a couple of closely together and those that are spaced further apart would add
reasons, including: (1) a collision could be caused by the instability more significance to the results.
of vehicular forces created by the transition from the end of the Another possible area for future research is the HSM calibration
curve to the straight segment, while the collision would occur out- method which could use further refinement in site selection. The
side the curve itself or (2) the momentum of a vehicle that leaves current recommendations do not specify a sampling technique to
the roadway or collides with another vehicle on the curve could reach the minimum number of sites (30–50 sites) or the minimum
carry the vehicle outside the boundary of the curve. However, this number of collisions (100 per year) (AASHTO, 2010). Site selec-
approach would be inconsistent with HSM procedures for defining tion could play a major role in the magnitude and variability over
the curve length and its parameters. time of the calibration factor, particularly if high crash locations are
purposefully included in the study.
6. Recommendations Furthermore, the CMF for superelevation could also be fur-
ther investigated to determine if the variance between actual and
The recommended equation to account for the spatial rela- required superelevation fully incorporates the impact of superel-
tionship of horizontal curves (while including the geometric evation on horizontal curve safety. The HSM incorporates the
considerations of the curve from the Highway Safety Manual variance in superelevation, but with advancements in data col-
model) is Eq. (4). The distance to the distal and proximal adjacent lection methods for superelevation, a CMF could become more
curves and an interaction term for the distances to adjacent curves sophisticated to more accurately model the superelevation effects
are the spatial variables of the model. Eq. (4) is especially useful and the changes of superelevation through a curve. The transi-
because the implementation of the HSM safety model is incom- tion from a normal crown cross-slope (each lane sloped away from
plete without the spatial consideration. Practitioners can input the the centerline at 2%) to a fully superelevated roadway (both lanes
HSM results along with the variables listed in Eq. (4) to estimate sloped in one direction to the design superelevation) can be difficult
the total number of collisions on the curve. When using Eq. (4) to design and if done incorrectly can cause safety concerns on the
distances and lengths should be non-zero and should be applied to curve. Further research could examine the effect of any deviation
curves with similar features as the curves analyzed for this research. from the recommended transition to full superelevation.
Although this study did not include any curves with no tangent dis-
tance between adjacent curves, the recommended model appears
Acknowledgements
to provide reasonable results for small or zero tangent values.
An example calculation using Eq. (4) is shown below. The same
This paper was based on a study funded by the North Car-
HSM assumptions as presented in Table 2 that yielded an estimate
olina Department of Transportation entitled “Procedure for Curve
of 0.32 predicted collisions per year for an example curve were
Warning Signing, Delineation, and Advisory Speeds for Horizontal
used. The distance to the distal curve is 1.5 miles and the distance
Curves.” Brian Mayhew served as the NCDOT Chair of the Steer-
to the proximal curve is 0.3 miles. The assumed curve radius for the
ing and Implementation Committee. NCDOT has not evaluated or
curve of interest is 1000 ft. Accounting for the spatial considerations
adopted the findings of this paper or accepted or rejected any of its
of the adjacent curves, the resulting prediction is 0.19 collisions per
conclusions.
year, showing that the recommended model expects this example
curve to experience more collisions than the HSM prediction for
the curve. References

Col = HSM[0.116 + (0.367 × D) + (0.122 × P) + (−0.255 × D × P)] AASHTO, 2010. Highway Safety Manual. American Association of State Highway
Transportation Officials, Washington, D.C.
Aram, A., 2010. Effective safety factors on horizontal curves of two-lane highways.
Col = 0.32 collisions per year [0.116 + (0.367 × 1.5 miles) Journal of Applied Sciences 10 (22), 2814–2822 (Elsevier).
Bonneson, J., et al., 2007. Horizontal Curve Signing Handbook. TTI Technical Report.
+ (0.122 × 0.3 miles) + (−0.255 × 1.5 miles × 0.3 miles)] Texas Transportation Institute, College Station, Texas.
Charlton, S.G., 2007. The role of attention in horizontal curves: a comparison of
advance warning, delineation, and road marking treatments. Accident Analysis
Col = 0.19 predicted collisions per year and Prevention 39 (5), 873–885 (Elsevier).
304 D.J. Findley et al. / Accident Analysis and Prevention 45 (2012) 296–304

Chrysler, S.T., et al., 2005. Simplifying Delineator Applications for Horizontal Curves. curves. NCDOT Report, No. FHWA/NC/2009-07. North Carolina Department of
Transportation Research Record, vol. 1918. Transportation Research Board. Transportation. Raleigh, NC.
National Research Council, Washington D.C, pp. 68–75. Lamm, R., Psarianos, B., Cafiso, S., 2002. Safety Evaluation Process for Two-Lane
Datta, T.K., Perkins, D.D., Taylor, J.I., Thompson, H.T., 1983. Accident Surrogates for Rural Roads: A 10-Year Review. Transportation Research Record, vol. 1796.
Use in Analyzing Highway Safety Standards. Publication FHWA-RD-82-103-105. Transportation Research Board. National Research Council, Washington D.C, pp.
Federal Highway Administration, U.S. Department of Transportation, Washing- 51–59.
ton D.C. Lyles, R.L., Taylor, W., 2006. Communicating changes in horizontal alignment, NCHRP
Eccles, K.A., Hummer, J.E., 2000. Safety effects of fluorescent yellow warning signs Report 559. Transportation Research Board. National Research Council, Wash-
at hazardous sites in daylight curves. In: Proceedings of the Transportation ington D.C.
Research Board 80th Annual Meeting. Transportation Research Board. Paper McGee, H.W., Hanscom, F.R., 2006. Low-Cost Treatments for Horizontal Curve Safety.
Number 01-2236, Washington D.C. Publication FHWA-SA-07-002. Federal Highway Administration. U.S. Depart-
Elvik, R., 2007. Road safety management by objectives: a critical analysis of the ment of Transportation, Washington, D.C.
Norwegian approach. Accident Analysis and Prevention 40 (3), 1115–1122 (Else- Milton, J.C., Mannering, F.L., 1998. The relationship among highway geometrics
vier). traffic-related elements and motor-vehicle accident frequencies. Transportation
Fitzpatrick, K. et al., 2000. Speed prediction for two-lane rural highways. Report 25, 395–413.
FHWA/RD-99/171. Federal Highway Administration. United States Department Nielsen, M.A., Greibe, P., Herrsted, L., 1999. Signing and Marking of Substandard
of Transportation. Washington D.C. Horizontal Curves on Rural Roads (Main Report). Road Directorate, Ministry
FHWA, 2009. Manual on uniform control devices. Federal Highway Administration. of Transport, Report 163. Traffic Management and Road Safety, Copenhagen,
United States Department of Transportation, Washington D.C., p. 110 (Table 2C- Denmark.
5. Horizontal Alignment Sign Selection). Rasdorf, W., Findley, D.J., Zegeer, C.V., Sundstrom, C.A., Hummer, J.E. Evaluation of
Florida Department of Transportation (FDOT), 2010. Geographic Information Sys- GIS applications for horizontal curve data collection. Journal of Computing in
tem (GIS): Curvature Extension for ArcMap 9.x. Transportation Statistics Office, Civil Engineering, American Society of Civil Engineers, in press.
FDOT. Tallahassee, FL. URL: http://www.dot.state.fl.us/planning/statistics/gis/ Retting, R.A., Farmer, C.M., 1998. Use of pavement markings to reduce excessive
(accessed February 14, 2011). traffic speeds on hazardous curves. ITE Journal: Institute of Transportation Engi-
Glennon, J.C., Neuman, T.R., Leisch, J.E., 1985. Safety and operational considerations neers 68 (9), 30–36.
for design of rural highway curves. Report FHWA/RD-86/035. Federal Highway Stata, 2010. Stata/SE 11.0 for Windows. College Station, TX.
Administration. United States Department of Transportation. Washington D.C. Strathman, J.G., et al., 2001. Analysis of Design Attributes and Crashes on the Oregon
Glennon, J.C., 1987. Effect of alignment on highway safety. State of the Art Report 6. Highway System. Publication FHWA-OR-RD-02-01. Federal Highway Adminis-
Transportation Research Board. National Research Council, Washington D.C. tration. U.S. Department of Transportation, Washington D.C.
Guo, F., Wang, X., Abdel-Aty, M.A., 2010. Modeling signalized intersection safety Terhune, K.W., Parker, M.R., 1986. An Evaluation of Accident Surrogates for
with corridor-level spatial correlations. Accident Analysis and Prevention 42 Safety Analysis of Rural Highways. Publication FHWA-RD-86-127. Federal
(1), 84–92 (Elsevier). Highway Administration. U.S. Department of Transportation, Washington
Hammond, J., Wegmann, F., 2001. Daytime effects of raised pavement markers on D.C.
horizontal curves. ITE Journal: Institute of Transportation Engineers (August), Torbic, D.J. et al., 2004. Guidance for Implementation of the AASHTO Strategic High-
38–41. way Safety Plan, Volume 7: A Guide for Reducing Collisions on Horizontal Curves.
Harwood, D.W., Council, F.M., Hauer, E., Hughes, W.E., Vogt, A., 2000. Prediction NCHRP Report 500. Transportation Research Board. National Research Council,
of the Expected Safety Performance of Rural Two-Lane Highways. Publication Washington D.C.
FHWA-RD-99-207. Federal Highway Administration. U.S. Department of Trans- Wong, Y.D., Nicholson, A., 1992. Driver behaviour at horizontal curves: risk com-
portation, Washington D.C. pensation and the margin of safety. Accident Analysis and Prevention 24 (4),
Herrstedt, L., Greibe, P., 2001. Safer signing and marking of horizontal curves on 425–436 (Elsevier).
rural roads. Traffic Engineering and Control Magazine 42, 82–87 (Hemming Zegeer, C.V., et al., 1991. Cost-Effective Geometric Improvements for Safety
Information Systems. London, England). Upgrading of Horizontal Curves, Publication FHWA-RD-90-074. Federal
Hummer, J.E., Rasdorf, W., Findley, D.J., Zegeer, C.V., Sundstrom, C.A., 2010a. Curve Highway Administration, U.S. Department of Transportation, Washington
crashes: road and collision characteristics and countermeasures. Journal of D.C.
Transportation Safety and Security 2 (3), 203–220 (Southeast Transportation Zegeer, C.V., Stewart, J.R., Council, F.M., Reinfurt, D.W., Hamilton, E., 1992.
Research Center). Safety Effects of Geometric Improvements on Horizontal Curves, Transporta-
Hummer, J.E., Rasdorf, W., Findley, D.J., Zegeer, C.V., Sundstrom, C.A., 2010b. Proce- tion Research Record. Volume 1356. Transportation Research Board. National
dure for curve warning signing, delineation, and advisory speeds for horizontal Research Council, Washington D.C, pp. 11–19.

S-ar putea să vă placă și