U.S. Department of Transportation
Federal Highway Administration
1200 New Jersey Avenue, SE
Washington, DC 20590

Skip to content
Facebook iconYouTube iconTwitter iconFlickr iconLinkedInInstagram


FHWA Home / Safety / Pedestrian & Bicycle / Synthesis of Methods for Estimating Pedestrian and Bicyclist Exposure to Risk at Areawide Levels and on Specific Transportation Facilities

Synthesis of Methods for Estimating Pedestrian and Bicyclist Exposure to Risk at Areawide Levels and on Specific Transportation Facilities

 < Previous Table of Content Next >

Chapter 4. Exposure Analysis on Specific Transportation Facilities

This chapter summarizes examples of exposure analyses conducted on specific transportation facilities. In some cases, exposure estimates are calculated for specific facilities, but also aggregated to various areawide geographies. For the purposes of this report, the examples of exposure analysis have been included in either Chapter 3 or 4 according to the most granular scale at which exposure estimates are produced.

Summary of Practice

Increasingly, public agencies are conducting regular counts and intercept surveys as part of routine pedestrian and bicyclist monitoring programs to monitor changes in behavior and volume.  Over the years, there has been an attempt to standardize the collection and reporting of this pedestrian and bicyclist count data, first through Alta/ITE’s National Bicycle and Pedestrian Documentation Project, FWHA’s Traffic Monitoring Guide (TMG), NCHRP 797 report, and most recently through the FHWA’s Exploring Pedestrian Counting Procedures report.

These count and direct measurement methods enable communities and regions to quantify exposure at a granular scale: points (i.e., intersection or mid-block crossings) and segments.  However, pedestrian and bicyclist counts are most often conducted as a sampling at a limited number of locations within a city, so directly measured counts are not available for network or areawide exposure estimation.  The limited number of counts could, nevertheless, be used as model inputs for estimating exposure at an aggregate scale for an entire area.  Therefore, counts of pedestrian and bicyclist volumes can play the following roles in the development of exposure estimates:

Table 2 provides a summary review of existing and proposed methods of directly measuring and estimating exposure on specific transportation facilities. Table 2 provides the following information for each example:

Table 2. Examples of Exposure Analysis on Specific Transportation Facilities
Reference Development Scales Coverage Data Sources Methods Unit of Exposure
Cameron 1982 Point, Segment Sydney, Australia Manual counts -- Product of Pedestrian and Vehicle Volumes
Tobey, Shunamen, and Knoblauch 1983 Point, Segment Five Standard Metropolitan; Statistical Areas Manual counts & vehicle ADT -- Number of Pedestrian-Vehicle (P x V) Interactions
Qin and Ivan 2001 Point Rural areas in Connecticut Manual counts, population & land use data Direct Demand Model Weekly Crossing Pedestrian Volume
Transport Research Laboratory 2001 Point,
National & Local Levels within Europe Regional or national travel surveys & direct counts -- Million Kilometers of Travel
Raford and Ragland 2004 Point Oakland, CA Manual counts & census data Network Analysis Model Average Annual Pedestrian Volume
Greene-Roesel, Diogenes, and Ragland 2007 Point,
-- Direct counts, surveys, & census data -- Number of Pedestrians; Trips;  Distance Traveled; Time Spent Traveling
Molino et al. 2009
Molino et al. 2012
Washington, D.C. Manual counts & crossing distances Direct Demand Model 100 Million Miles Traveled
Papadimitriou, Yannis, and Golias 2012 Segment Athens CBD, Greece Manual field surveys & counts Discrete Choice Model Vehicles Volume; Encountered While Crossing; Product of Vehicle; Volume and Pedestrian Crossing Time
Schneider et al. 2012 Point San Francisco, CA Manual & automated counts Direct Demand Model 10 Million Crossings
Schneider, Grembek, and Braughton 2013 Point UC Berkeley Campus Boundary Manual & automated counts -- 10 Million Crossings
Strauss, Miranda-Moreno, and Morency 2013 Point Montreal, QC, Canada Manual counts Direct Demand Model Million Cyclists or Pedestrians per Unit of Time
Strauss, Miranda-Moreno, and Morency 2014 Point Montreal, QC, Canada Manual counts Direct Demand Model Million Cyclists or Pedestrians per Unit of Time
Hankey and Lindsey 2016 Point,
Minneapolis, MN Manual counts, census & land use data Direct Demand Model Bicycle & Pedestrian Volumes
Liggett et al. 2016 Point, Segment Los Angeles County, CA Manual counts -- Average Number of Riders
Radwan et al. 2016 Point, Segment State of Florida Direct counts, population, distance crossed, vehicle ADT Direct Demand Model Million Pedestrian Miles Crossed per Entering Vehicle; 100 Million Vehicle Miles; Million Pedestrian Miles Crossed per Entering Vehicle
Wang, Lindsey, and Hankey 2016 Point,
Minneapolis, MN Manual counts Direct Demand Model Bicyclist Volumes

Note: -- indicates that the item did not pertain to the particular study.

Development Scales

Regarding direct measurement of exposure on specific transportation facilities, it is important to know at what spatial scale the measurements were initially developed for a particular application.  The development scale determines the ultimate application scale at which exposure can be applied or estimated for locations without direct measurements.  Table 2 offers a summary of the development scales for several nonmotorized exposure studies and applications by using the Highway Capacity Manual (HCM 2010) roadway system elements as levels of scale: points (i.e., intersections), segments, facilities, corridors, areas, and systems.  More information on these types of roadway system elements can be found in Chapter 2 of this report.

For the examples listed in Table 2, the exposure development scales varied based upon the type of data source and collection method, modelling efforts, and the intended application of the resulting exposure measure.  However, all of the examples utilized direct measurement of nonmotorized volumes at either the point or segment scales as their unit of exposure on specific transportation facilities (Papadimitriou et al. 2012; Qin and Ivan 2001; Raford and Ragland 2004; Schneider et al. 2012, 2013, Strauss et al. 2013, 2014).  A majority of the examples used exposure measurements at both the point and segment scales since their respective applications accommodated the scales either separately (Liggett et al. 2016; Radwan et al. 2016; Wang et al. 2016) or were aggregated (Cameron 1982; Greene-Roesel et al. 2007; Hankey and Lindsey 2016; Lyons et al. 2014; Molino et al. 2009, 2012; Rasmussen et al. 2013; Tobey et al. 1983; Transport Research Laboratory 2001).

There is a close connection between the exposure development scale and the data sources used.  Direct measurement of nonmotorized exposure is typically represented as pedestrian and bicyclist traffic volumes recorded through direct counts (manual or automated) at either points or segments.  Alternatively, a variety of the examples incorporated some form of exposure estimation via a direct demand or choice model that use additional data sources other than just counts (Hankey and Lindsey 2016; Lyons et al. 2014; Molino et al. 2009, 2012; Papadimitriou et al. 2012; Qin and Ivan 2001; Radwan et al. 2016; Raford and Ragland 2004; Rasmussen et al. 2013; Schneider et al. 2012; Strauss et al. 2013, 2014; Wang et al. 2016).  Keep in mind that the models still required a representative, random or non-random sample of nonmotorized count data for the scale at which exposure was estimated, i.e., intersections (i.e., points) and segments.

Units of Exposure

Greene-Roesel et al. (2007) states that “there is no single best definition of pedestrian exposure”; the same can be said about bicyclist exposure.  In epidemiology, a general definition of exposure is contact with or proximity to a potentially harmful agent or event (Last et al. 1995).  Table 2 shows that, within pedestrian and bicyclist exposure can be defined and measured in a variety of ways.

As illustrated in the previous section on development scale, the components of a safety study or risk assessment are closely interrelated and in some cases mutually exclusive.  Being that the unit of exposure is an important component, it is directly dependent upon other aspects of the study, such as: development scale, target population, estimation methods, geographic area, available resources, purpose, etc.  Therefore, the unit of exposure that best fits the needs and purposes of the study should be chosen (Greene-Roesel et al. 2007).

Tables 3 through 7 were adapted from Greene-Roesel et al. (2007) to help practitioners choose the appropriate unit of exposure for nonmotorized travel modes.  These tables help to summarize the appropriate uses and the pros and cons of each generalized form of exposure related to nonmotorized travel: volume, distance, and time.  Note that the population and trips exposure measures are best used for aggregate level estimations, such as neighborhoods, cities, regions, etc., which was discussed in Chapter 3.

Table 3. Exposure Based on Volumes/Counts
(adapted from Greene-Roesel et al. 2007)
Appropriate Uses
  • Estimating pedestrian and bicyclist volume and risk in a specific location.
  • Assessing changes in pedestrian volume or characteristics due to countermeasure implementation at that site.
How Data are Gathered
  • Manual or automated counts of bicycles and pedestrians.
  • Typically a daily count, and sometimes an annualized estimate
  • Counts are simpler to collect than other measures such as time or distance walked or biked.
  • Automated methods for counting number of bicycles and pedestrians are improving.
  • Does not account for the amount of time spent walking or biking nor the distance.
  • Not easily adapted to assess exposure over wide areas (for example, a city).
Common Measures
  • Number of bicycles and/or pedestrians per time period.
  • Number of crossings.
  • Average daily, weekly, or annual volume per point or segment.
  • Product of bicycle or pedestrian and vehicle volumes (interactions).
Table 4. Exposure Based on Distance
(adapted from Greene-Roesel et al. 2007)
Appropriate Uses
  • Estimating exposure at the micro or macro level.
  • Estimating whether risk increases in a linear manner with distance traveled.
  • Assessing how crossing distance affects risk.
How Data are Gathered
  • For individual level exposure, through travel surveys.
  • For aggregate level exposure, measurement of the length of the area of interest, combined with a manual or automatic count of the number of pedestrians.
  • Can be used to measure exposure at the micro and macro levels
  • More detailed than pedestrian volumes or population data
  • Can be used to compare risk between different travel modes
  • Common measure of vehicle exposure
  • Does not take into account the speed of travel and thus cannot be reliably used to compare risk between different modes (e.g. walking and driving)
  • Assumes risk is equal over the distance walked
  • Must typically assume that each bicycle or pedestrian travels the same distance in a crossing, along a sidewalk, street, bike lane, etc.
Common Measures
  • Miles traveled (total or average) per pedestrian.
  • Miles crossed (total or average).
Table 5. Exposure Based on Time
(adapted from Greene-Roesel et al. 2007)
Appropriate Uses
  • Estimating total pedestrian and bicyclist time exposure for specific locations.
  • Comparing risks between different modes of travel (e.g. walking vs. riding in a car).
  • Estimating whether risk increases in a linear manner with walking time.
  • Comparing risk between intersections with different crossing distances and between bicycles or individuals with different travel speeds.
How Data are Gathered
  • The number of bicycles or persons passing through an area multiplied by the time traveled.
  • Time spent on walking activities reported on surveys.
  • Accounts for different walking speeds
  • Allows for accurate comparison between different modes of travel.
  • Can be used to measure exposure at the micro and macro levels
  • More detailed than pedestrian volumes or population data
  • Time based measures assume risk is equal over the entire distance of a crossing. Only a small portion of time spent walking on roadways represents real exposure to vehicle traffic. This portion would include time spent crossing roads, walking on the road surface, or possibly walking along the roadside where there are no curved sidewalks (Chu 2003).
  • Time spent on walking can be overestimated in surveys, because people perceive that they spend more time walking than they actually do (Chu 2003).
  • Walking may be under-reported in surveys, because people may forget walk trips or may purposely choosing not to report. Both of these reasons are related to the fact that walking trips are relatively short. These very short trips may not register in the memory of respondents or the respondents may think that these short trips are unimportant (Chu 2003).
Common Measures
  • Amount time traveling (total or average).
  • Amount time crossing an intersection (total or average).
Table 6. Exposure Based on Trips
(adapted from Greene-Roesel et al. 2007)
Appropriate Uses
  • Assessing pedestrian and bicyclist behavior in large areas, such as cities, states, or countries.
  • Examining changes in pedestrian and bicyclist behavior over time.
  • Making comparisons between jurisdictions.
  • Assessing common characteristics of walking trips, such as purpose, route, etc.
How Data are Gathered
  • Data are gathered through use of travel surveys.
  • Appropriate for use in large areas.
  • Best metric to assess relationship of walking with trip purpose.
  • Trips can be assessed as a function of person, household and location attributes.
  • As with most surveys, a large number of respondents are needed to adequately represent the underlying population.
  • Less meaningful at the level of detail needed to assess risk at specific locations.
  • Pedestrian trips are often underreported in surveys (Schwartz and Porter 2000).
Common Measures
  • Number of trips (total or average) possibly by purpose.
Table 7. Exposure Based on Population
(adapted from Greene-Roesel et al. 2007)
Appropriate Uses
  • Used as an alternative to exposure data when cost constraints make collecting exposure data impractical.
  • Used to compare jurisdictions over time because population data are available for many geographies and time periods.
How Data are Gathered
  • Population data for most cities is available on an annual basis through the ACS.
  • Easy and low-cost to obtain; available for most geographies and time periods.
  • Adjusts for differences in the underlying resident population of an area – for example, sparsely populated suburbs versus densely populated inner-city areas.
  • Provides a crude adjustment for amount of vehicle traffic on the streets, since areas where more people live also tend to be areas where more people drive.
  • May be the only way to represent exposure if direct measurements cannot be taken.
  • Does not accurately represent pedestrian and bicyclist exposure.
  • Does not account for the number of people who travel as bicyclists or pedestrians in the area.
  • Does not provide information about amount of time or distance that members of the population were exposed to traffic.
Common Measures
  • Number of people in a given area: neighborhood, city, county, state or country.
  • Number of people in a particular demographic group: by age, sex, race, immigrant status or socioeconomic status.


The applications of nonmotorized exposure presented in Table 2 are centered on producing relative risk metrics (e.g., crash rates) that can help to make sense of pedestrian and bicyclist crash data.  Many cities and states have pedestrian and bicyclist crash data but no way of understanding whether or not relative risk has changed.  Numerous examples in Table 2 focused on developing crash rates for segments, intersections, or both for particular study areas like a city or county (Liggett et al. 2016; Radwan et al. 2016; Raford and Ragland 2004; Schneider et al. 2012, 2013, Strauss et al. 2013, 2014; Wang et al. 2016).  Others used nonmotorized exposure to assess risk for other purposes, such as: trends over time for an area (Lyons et al. 2014; Rasmussen et al. 2013), before-and-after studies for new countermeasures (Molino et al. 2009, 2012), crash rates by time of day (Schneider et al. 2013), and cost-benefit analysis of safety measures (Transport Research Laboratory 2001).

Summary of Data Sources and Methods: Direct Measurement

The most direct form of measurement or nonmotorized travel exposure are pedestrian and bicyclist traffic counts.  To help communities develop nonmotorized count programs there has been an attempt to standardize the data collection and reporting efforts, first through the Alta and Institute of Transportation Engineers (ITE) National Bicycle and Pedestrian Documentation Project (NBPD), FWHA’s Traffic Monitoring Guide (TMG), NCHRP 797 report, and most recently through the FHWA’s Exploring Pedestrian Counting Procedures report.  These references are summarized below. This direct measurement approach (calculating exposure from systematic traffic monitoring programs) is the current state-of-the-practice for motor vehicle exposure estimation.

The NBPD is a count and survey effort designed to provide a consistent model of data collection and ongoing data repository for use by local agencies and organizations.  Since 2003, the NBPD objectives have been the following (Alta Planning & Design and Institute of Transportation Engineers 2009a):

The NBPD data collection effort is split into counts and surveys as follows:

The official national count and survey days are Tuesday through Thursday and the following Saturday of the second week in September.  A description of the facility, location, and methodology should accompany any data to allow analysis of the relationship between biking and/or walking rates and demographic and/or geographic factors.  A variety of materials are offered for download on the NBPD’s website.

In 2006, Caltrans selected the NBPD as the data collection methodology for the Seamless Travel research study in San Diego County (Jones et al. 2010).  Also in 2006, the FHWA Volpe National Transportation Systems Center chose the NBPD as an evaluation methodology for the Nonmotorized Transportation Pilot Project (Lyons et al. 2014).  The 2013 TMG references NBPD guidance for short-duration counts, counter positioning, count duration determination, and months/seasons of year for data collection (Federal Highway Administration 2013).

The 2013 TMG includes chapter 4 that is dedicated to nonmotorized traffic monitoring, which encompasses bicycles, pedestrians, and other nonmotorized road and trail users (Federal Highway Administration 2013).  The chapter emphasizes the challenges of collecting nonmotorized data as compared to established motorized data collection efforts, such as: the scale of data collection, higher use of lower functional class road and streets, greater amount of error from short duration counts, and quickly-evolving technologies and their unestablished error rates (Federal Highway Administration 2013).  Several examples of data collection equipment are provided to help the users understand the strengths and limitations of each type of technology and what is appropriate for their situation and budget to ensure proper counts.  The chapter provides guidance on what to consider when developing a nonmotorized traffic data collection program that includes both continuous and short-duration count locations.  Topics such as data management, site selection, variability, count duration, and factoring (correction and expansion) are all summarily discussed.

The NCHRP 797 Report Guidebook on Pedestrian and Bicycle Volume Data Collection published in 2014 is the most comprehensive resource on pedestrian and bicycle volume data collection that is currently available (Ryus et al. 2014).  The objectives of the guidebook are to help practitioners understand the value of nonmotorized data, develop a collection plan, identify the most appropriate method, and how to account for error introduced by counting technology (Ryus et al. 2014).  The guidebook does not cover trip sampling techniques, presence detection, and trip generation estimation, but instead it specifically focuses on methods for counting the actual number of pedestrians and bicyclists crossing a screenline or intersection.  The NCHRP 797 report complements the TMG Chapter 4 by expanding on the accuracy of different counting technologies accompanied by real-world examples of counting applications.  One of the several example applications provided is safety analysis where a measure of exposure is necessary to make sense of crash data and assess risk relative to volume(s), i.e., exposure.  No explicit definition of exposure is provided; however, the report explains that “exposure relates to the frequency of a bicyclist or pedestrian being present in a conflict zone with the potential to be involved in a crash” (Ryus et al. 2014).

In May 2016, FHWA published the Exploring Pedestrian Counting Procedures report (Nordback et al. 2016).  This report expands on the existing guidance and best practices listed above to recommend strategies for accurate, timely, and feasible measurement of pedestrian travel (Nordback et al. 2016).  The authors offer five key recommendations based on review of the current practice and webinar and interview responses (Nordback et al. 2016):

  1. Expand the use of multi-day/multi-week counts to reduce estimation error rates, and rotate counts around the network;
  2. Validate equipment at installation and regularly thereafter;
  3. Tailor quality checks appropriate for low volume versus high volume locations;
  4. Compute bias compensation factors (e.g., occlusion adjustment factors) to account for limitations related to equipment and locations; and
  5. Conduct both short-duration and continuous counts to fully consider temporal and spatial aspects of pedestrian traffic patterns.

The report goes into great detail in describing the operation and management of pedestrian count data collection equipment and the subsequent data management in terms of count duration, validation, calibration, purchasing strategies, quality assurance/control, standardization, factoring, accessibility, and analysis.  This is a key resource for practitioners interested in creating a pedestrian count program.

El Esawey et al. (2015) is a noteworthy example of direct measurement, in which various short-duration bicyclist counts conducted in Vancouver, British Columbia, in various months and years have been assembled to provide one of the most compete estimates of bicyclist traffic on a street network in a major North American city. In particular, the authors indicate that bicyclist risk exposure is a key application of their bicyclist count database. The paper by El Esawey et al. illustrates the challenges that public agencies face in creating a facility-specific bicyclist and pedestrian count database for an entire citywide street network that is based solely on direct measurement. Because of these challenges, some exposure analyses have tried to supplement the limited number of direct pedestrian and bicyclist counts with model estimates, which is the topic of the next section.

Summary of Data Sources and Methods: Estimation

In addition to direct measurement methodologies, exposure analysis at the facility or segment-level can be performed using estimation or modeling-based methodologies (see Figure 5).

Figure 5. This figure shows a hierarchy for estimation methods. The first level of the hierarchy includes these four estimation methods: 1) direct demand models; 2) regional travel demand models; 3) special-focused models; and, 4) simulation-based traffic models. Under special-focused models, there are four sub-categories: 1) GIS-based models; 2) trip generation and flow models; 3) network analysis models; and, 4) discrete choice models.

Figure 5. Estimation Methods That Have or Could Be Used For Exposure Analysis

Direct (facility) demand models have played a major role in the area of bicycle/pedestrian safety whereas the other modeling types have been used infrequently or not at all. An overview of all these potential models is provided here for completeness. Given the focus of the pedestrian and bicyclist safety analysis, example studies provided in Table 2 are discussed in more details.

Direct Demand Models

Direct demand models are among the most widely used tools in the literature for pedestrian and bicyclist volume estimation modeling (Pulugurtha et al. 2006; Pulugurtha and Repaka 2008; Schneider et al. 2009; Griswold et al. 2011; Strauss and Miranda-Moreno 2013; Tabeshian and Kattan 2014; Fragnant and Kosckelman 2016; Schmiedeskamp and Zhao 2016). These models have been primarily used to develop facility-demand estimations for the local level of community, project, and facility planning. The FHWA has released a Non-Motorized Travel Analysis Toolkit, which includes various applications to support non-motorized transportation planning and modeling. This Toolkit includes several direct demand models to estimate pedestrian and bicycle volumes (FHWA 2016). Direct demand models have also served as the primary tools to measure bicyclist and pedestrian exposure for safety analysis. As discussed in Chapter 3, aggregate demand models are also similar to direct demand models though the analysis is performed at a larger level (e.g., regional level) using aggregate characteristics.

Direct demand models are generally based on different versions of regression modeling to explain “demand levels as recorded in counts as a function of measured characteristics of the adjacent environment” (Kuzmyak et al. 2014). For example, researchers from the University of Minnesota (Lindsey et al. 2012; Hankey et al. 2012) have developed several count-based pedestrian and bicyclist models to estimate nonmotorized traffic in Minneapolis, Minnesota using ordinary least squares and negative binomial regression. While neighborhood design and urban form were found to be more significant for estimating bicycle traffic, road classification, proximity to amenities and activity centers were identified as significant independent variables for pedestrian traffic. As part of Seamless Travel Project, Jones et al. (2010) developed pedestrian and bicyclist demand models to estimate volumes at intersections during 7 to 9 A.M. periods in San Diego County. These models were constructed using 80 manual, five automatic machine count locations, and GIS data on land use, demographics, etc. Fehr & Peers (2010) also developed similar pedestrian and bicycle models to estimate volumes during 5 to 6 p.m. peak periods for Santa Monica.

Direct demand models are appealing due to their simplicity in development and application, and since they are generally based on available data. However, they are limited in terms of capturing the behavioral structure and also not transferable due to relatively limited sample size and characteristics that the models are built on (see also Kuzmyak et al. 2014 for a detailed discussion on the advantages and limitations of such models).  Schmiedeskamp and Zhao (2016) explained such models as following “a similar approach of first proposing a set of explanatory variables, fitting some form of regression model, and then interpreting and justifying the results according to the guiding theory”. Several different types and forms of explanatory variables have been used in the development of these models. The variables included but not limited to transportation system variables, built environment variables, socio-economic characteristics, weather and typology. The model variables showed some differences based on the mode (i.e. pedestrian model versus bicycle model) as well as the statistical model type.  Several of the studies reviewed indicated the influential effects of density, accessibility and proximity on pedestrian models. In addition to similar neighborhood forms, bicycle models were found to be influenced by some other specific infrastructure and system characteristics, such as presence of bicycle lanes and traffic volume.  The direct demand models have also been used for predicting volumes at locations where the count data are not available, extending the study to an areawide level in the application process. While several similar works have been developed over the last decades, the studies differ such that researchers work on improving the estimation results by developing models that are more meaningful, practically more applicable, and statistically more robust.

The examples discussed in this section are pedestrian and bicyclist safety analysis examples that include exposure. Schneider et al. (2012) developed and applied a pedestrian intersection volume model for San Francisco. A sample of counts at 50 intersections were collected and adjusted to produce annual pedestrian crossing estimates at each sampled intersection. Next, the authors developed a log-linear regression model to identify the relationship between annual pedestrian volume estimate and various different explanatory variables including land use, transportation system, local environment, and socioeconomic characteristics near each sampled intersection.  The model was then used to evaluate pedestrian crossing risk at each intersection based on the exposure measure of the number of pedestrian crashes per 10 million crossings. Molino et al. (2009; 2012) also developed a log-linear regression model (with Poisson distribution) to estimate pedestrian counts at signalized intersections in Washington, D.C. While 15-min pedestrian counts served as the dependent variable, the independent variables of the model included land use variables (e.g. commercial, residential) and characteristics of the day (e.g. day of the week, time of the day). Using the parameter estimates of the model and follow-up adjustment procedures, a total number of miles traveled were estimated “…by multiplying the total number of pedestrians by the mean width of all the sampled signalized intersections.” This result was then used as an exposure measure in pedestrian crash rate computation.

Several others have followed similar approach in estimating bicycle or pedestrian volume exposure measures using direct demand models. For example, Qin and Ivan (2001) estimated a general linear model for rural areas of Connecticut. Their dependent variable was the weekly pedestrian volume crossing the street at 32 different sites. The estimated model indicated that the number of lanes, area type, and sidewalk system significantly associated with the estimated pedestrian volume, which was proposed to be used in analyzing pedestrian fatality and injury rates in a follow-up research. Abasahl (2013) developed linear and log-linear models for both pedestrian and bicycle volume estimation. Their dependent variables were pedestrian and bicycle volumes collected at 92 (signalized) intersections in four Michigan cities. The estimated model volumes were then used as exposure measures in crash analysis for each nonmotorized mode and for all cities in the study area. Using a count database of 954 observations and 471 locations, Hankey and Lindsey (2016) employed a stepwise linear regression model that allowed for varying spatial scale of independent variables including land use and transportation network variables. Relying on the modeled estimates of bicycle traffic from this latter work, Wang et al. (2016) then estimated peak-hour bicycle traffic volumes for many segments in Minneapolis. The model results were then converted to bicycling volume for intersections and used for computing bicycle crash rates by intersections and segments. Radwan et al. (2016) used a stepwise regression model to estimate intersection pedestrian volume (for 52 sample intersections). The model was based on three main independent variables: daily traffic volumes, distance crossed, and population. The model estimates were then used to classify intersections across the state of Florida to compute the statewide averages for pedestrian crash rates at intersections.

Strauss et al. (2013; 2014) used a relatively improved version of modeling to estimate nonmotorized demand. Strauss et al. (2013) developed a bivariate Bayesian Poisson model to simultaneously estimate cyclists’ injury occurrence and bicycle activity at 647 signalized intersections on the island of Montreal, Quebec, Canada. In a follow-up study, Strauss et al. (2014) applied their Bayesian modeling methodology as part of a multimodal approach aimed at examining the safety at intersections for both nonmotorized and motorized traffic. After model calibration, the study compared injury and risk between modes and intersections by using the “expected number of injuries (obtained from the models) per million cyclists, pedestrians or motor-vehicle occupants per year” as the expected risk.

Regional Travel Demand Models

The state of practice mainly consists of regional travel demand models that are based on traditional trip-based forecasting models. The trip-based models generally consist of four main steps: trip generation, trip distribution, mode share, and traffic assignment. During the assignment step, the predicted traffic volumes are assigned to the individual network links (usually based on shortest distance). TAZs are the most commonly used geographic units to inventory existing and future demographic data required for modeling purposes. Therefore, trip-based models are particularly limited in estimating nonmotorized travel due to their coarse level of spatial analysis structure. Schneider et al. (2009) and Griswold et al. (2011) indicated that such regional demand models are not adequate in capturing the fine-grained differences in intersection-level bicycle or pedestrian activity. Kuzmyak et al. (2014) pointed out that “…if these models are used to account for nonmotorized travel, it is typically limited to the trip generation step; nonmotorized trip productions and attractions are estimated, but they are then removed from the remainder of the analysis, which focuses on motor vehicle trips”. Furthermore, Aoun et al. (2015) emphasized the lack of these regional models in capturing recreation trip purpose, which is a key consideration in pedestrian and bicyclist trip rates.

To overcome these limitations and increase their sensitivity to pedestrian and bicyclist trips, several enhancements have been made to trip-based models, such as developing an enhanced trip generation model sensitive to land use factors or an enhanced auto ownership model as input to nonmotorized trip production. Emerging tour- or activity-based models also provide superior alternatives to traditional four-step models since they are based on individuals rather than trips, and the spatial resolution can be reduced to a smaller level of geography (such as parcels instead of TAZs).

Special Focused Models

The literature has also witnessed various versions of specially focused models that can be particularly beneficial to be applied at the corridor and subarea planning level. These are primarily variations of focused regional model approaches. For example, the scenario planning tools, which heavily depend on the usage of GIS-based modeling methodologies, are used to estimate nonmotorized travel under alternative land use and transportation investment scenarios. The GIS-based walk-accessibility model developed by the NCHRP 08-78 research team provided an enhanced example to such tools, expanding their capability by estimating pedestrian trip tables using GIS-derived walk-accessibility scores. This model uses GIS to compute measures of accessibility to or from any point by all modes and attraction types, and then estimates mode split and creates walk trip tables by purpose (Kuzmyak et al. 2014).

Similarly, the pedestrian trip generation and flow models are among the examples of focused regional models specifically focusing on pedestrian travel and smaller geographical zones rather than TAZs. The two examples of this type of modeling include PedContext and its sequel Model of Pedestrian Demand (MoPeD), which were developed through the University of Maryland’s National Center for Smart Growth. While PedContext is not publically available, MoPeD is an open source model designed to be used by practitioners and non-experts. The model requires data on vehicle ownership, street connectivity, parcel-level land use, Census population and employment, and travel survey. Kuzmyak et al. (2014) indicated that, compared to the NCHRP 08-78 GIS-based walk-accessibility model, MoPeD is limited since it only generates walk trips instead of developing an overall trip generation and mode choice model. In particular, MoPeD uses pedestrian analysis zones (which are block or street-level) as the level of analyses. The main modeling structure follows regional four-step process, but employed only for walk trips, yielding to a link and intersection activity levels for walk trips (see http://kellyjclifton.com/products/moped/).  Clifton et al. (2008) indicated that the objective of the MoPeD study was to “develop a method to estimate pedestrian demand or pedestrian volumes on a network – in order to evaluate pedestrian risk exposure in Maryland communities”.

Network analysis models can also be discussed under these special focused models since they usually use variations of the four-step modeling approach for trip generation and distribution (Raford and Ragland 2006). These models are based on a representation of a pedestrian network. They are used to estimate volumes for specific facility types (e.g. street segments or intersections) over an entire area of interest (e.g. neighborhood or city). Space Syntax is one of the most well-known examples of network analysis models, which was first developed in mid 1980s in London. The model framework has been widely used in planning projects in Europe and Asia, but relatively unknown in the United States (Raford and Ragland 2004; McDaniel et al. 2014). Kuzmyak et al. 2014 indicated two potential reasons for its relatively minimal usage: 1) The information on its special software is limited; and 2) The process is not intuitive to transportation planners since its process does not follow traditional trip generation and distribution steps. Instead, it uses spatial characteristics and relationships to explain the route chosen. Raford and Ragland (2004) described space syntax as a “suite of modeling tools and simulation techniques used to analyze pedestrian movement and to predict pedestrian volume”. The authors listed seven steps for creation of a space syntax predictive model (Raford and Ragland, 2006):

  1. Base data collection
  2. Pedestrian route network modeling
  3. Processing for movement potentials
  4. Collection of pedestrian counts to calibrate the model
  5. Addition of land use, transit and other variables
  6. Testing the accuracy of the model
  7. Forecasting future pedestrian volumes based upon network changes

In one of their relevant studies, Raford and Ragland (2004) applied the space syntax model to estimate pedestrian volumes at 670 intersections (from a total of 42 intersections and 94 counts) for the City of Oakland, California. The output volumes were then used in a safety analysis for the city’s first pedestrian master plan. In specifics, the relative risk used for the safety analysis was computed for each intersection by dividing the annual pedestrian-vehicle collisions by exposure measure (i.e. average annual pedestrian volume).

Finally, there have been few efforts using the underlying principles of discrete choice models to model crossing behavior based on the information about the crossings and crossing behavior during the trip as an exposure measure. Lassarre et al. (2007) explained that “accident risk exposure can be estimated for any location (micro-environment) along a pedestrian trip, where a pedestrian is likely to cross”. This then requires development of a pedestrian behavior choice model for each location along an entire trip. Lassarre et al. (2007) used this concept in developing a framework for modeling pedestrian’s exposure considering their crossing behavior. A nested logit model was utilized for developing a hierarchical choice structure between junctions and mid-block locations. The independent variables used included origins and destinations, traffic characteristics (e.g. traffic volume), and pedestrian facilities (e.g. pedestrian signal). The final estimated crossing probabilities were then used to compute accident risk exposure. Using a sequential logit model, Papadimitriou et al. (2012) followed a similar approach by first estimating crossing probabilities associated with each alternative location along a trip, and then estimating pedestrian exposure to risk based on this crossing behavior. A pilot model was applied to a pedestrian trip in Athens, Greece.

Simulation-based Traffic Models

Benefiting from advances in computation capabilities, simulation-based traffic models have also evolved substantially over the last two decades. These models differ from the travel demand models as they often use the traffic volumes output of the relevant travel demand models as inputs to the traffic simulation models (DKS et al. 2007). These models can be applied at microscopic, macroscopic, or mesoscopic levels. For example, Abdelghany et al. (2012) developed a mesoscopic simulation-based dynamic trip assignment model for large-scale pedestrian networks. Their model is responsive to predict pedestrian responses to changes in design, operational conditions, and crowd management. The authors indicated that “the model can configure the study area in the form of a network and represent pedestrian demand at the individual level”. Hong et al. (2016) developed a pedestrian exposure model based on a hybrid microsimulation-statistical model that also accounts for heteroscedasticity and spatial correlation. They used data composed of pedestrian dynamics, pedestrian area dynamics, and network topology measures. The authors tested their methodology for modeling 688 crosswalks in Seattle Washington using various network typology measures integrating centrality, clustering, attractiveness, etc.

In fact, the recent literature has witnessed development of various agent-based microsimulation models aimed at capturing pedestrian activity in an area. Raford and Ragland explained that these models are based on “simulation of individual pedestrian movement in crowds (“agents”) based on complex behavioral rules and environmental modeling”. Greene-Roesel et al. (2007) described these models as using “flow principles from physical science to model pedestrian behavior in confined spaces such as interior of shopping mass or subway stations, on a single or small number of streets, or within building interiors”. Microsimulation models provide highly accurate, detailed, and visually strong pedestrian flow. On the other hand, they are complex and require significant input data and special resources, such as specialized software (e.g. VISSIM, PARAMICS, etc.) and unique expertise (Raford and Ragland 2006; Greene-Roesel et al. 2007).

Innovative Methods to Estimate Facility-Specific Exposure

Thus far, this chapter has focused on data sources and methods used by practitioners and researchers to directly measure or estimate facility-specific pedestrian and bicyclist exposure. However, there are several innovations that, if fully developed in the next five to ten years, have the potential to dramatically improve pedestrian and bicyclist activity data to estimate exposure for safety analyses. The next several sections describe these innovative methods and data sources.

Crowdsourcing from Mobile Devices

Many pedestrians and bicyclists carry a highly-advanced, location-aware sensor everywhere they go: their GPS-enabled smartphone (http://www.pewinternet.org/fact-sheet/mobile/). There are many commercial applications that rely on knowing the location of a smart phone to provide location-based services (such as Google Maps and Traffic, Facebook, TripAdvisor, etc.), and these companies are using petabytes of historical smartphone location information and advanced analytics to estimate activity levels at businesses and other points of interest (e.g., see Google’s Popular Times feature). Other companies (such as Streetlight Data, Cuebiq, etc.) also work in this location analytics domain, and try to provide insights in real-world consumer behaviors and trends. It is important to note that some population groups (such as the elderly, poor, young children) may be underrepresented in smartphone activity data.

This type of smartphone monitoring is considered passive, in that the smartphone owner/user does not have to initiate an app in order for the smartphone location to be tracked (although location services does have to be enabled within the app’s settings). With passive monitoring, the smartphone routinely stores and sends the smartphone location without owner/user intervention. Sometimes this occurs even when the smartphone app is not open or active (assuming that locations services are enabled). As one might expect, passive and anonymous monitoring of thousands or millions of location-aware smartphones over several months or years could lead to significant improvements in pedestrian and bicyclist activity data.

Another approach to smartphone monitoring is becoming more common in pedestrian and bicyclist planning. This approach, sometimes called active monitoring, relies on a smartphone owner/user to initiate a fitness-based app (such as Strava) or other trip/activity collector app (such as Moves) to gather walking, jogging, running, or bicycling activity. In other cases, a wearable fitness device (such as Fitbit) could be used to gather route choice information from pedestrians or bicyclists. The user agreement for these apps includes a provision for the app to anonymously collect and reuse the smartphone location data. To date, Strava has been the most active app developer to resell this pedestrian and bicyclist activity data to public agency for planning purposes. Relative walking and bicycling activity levels of Strava users (i.e., Strava Heat Map) can be seen online at http://labs.strava.com/heatmap.

Both of these smartphone monitoring approaches have their strengths and limitations. Passive monitoring can provide larger and less biased samples of pedestrians and bicyclists, simply because it does not require the smartphone owner/user to initiate app-based location monitoring. But in some situations, it may be difficult to differentiate travel mode (e.g., bicyclists in slow-moving motor vehicle traffic). Active monitoring (such as Strava) can provide more detailed data about activity type (e.g., walking, jogging, bicycling, off-road bicycling) as well as demographics. However, the activity data from active monitoring is typically from a much smaller and more biased sample (e.g., recreation-based activity) (Griffin and Jiao 2016).

Advanced Video Image Processing

Image processing on video data or Google Street View uses machine learning algorithms to accurately identify a non-motorist from the field-of-view of video or images. Video-based and image-based human detection and counting have been studied by many researchers (Dalal and Triggs, 2005; Gallahar and Chen, 2009; Tan et al., 2011; Yin et al., 2015; Qi et al., 2016; Tome et al., 2016;). The common method applied in these studies by extracting the features from the images at first, and then utilizing different machine learning algorithms to perform a precise classification. Although researchers developed many high quality algorithms, exploration on pedestrian detection and count is still attracting many researchers as there is an ongoing demand for more accurate estimates.

The algorithm developed by Dalal and Triggs (2005) is considered as one of the most popular pedestrian detection algorithm. The detection was done by the histogram based on the gradient direction. Tan et al. (2011) developed semi-supervised elastic net to count pedestrians. The research team used sequential information between unlabeled samples and their temporally neighboring samples to perform the study. The developed method showed superior predictions than the other state-of-the-art studies. Yin et al. (2015) used Google Street View images to extract pedestrian counts using machine vision and learning technology. The reliability tests results showed that the developed method was adept in counting pedestrians with a reasonable level of accuracy. Qi et al. (2016) proposed a sparse representation based approach for pedestrian detection from thermal images. In this study, the researchers first adopted the histogram of sparse code to represent image features and then detected pedestrian with the extracted features in different frameworks. This study validated the approach by comparing with three widely used features: Haar wavelets, histogram of oriented gradients, and histogram of phase congruency. The results showed the superiority of the approach developed in this study. Tome et al. (2016) used the popular modern algorithm deep convolution neural network to detect pedestrians. The research team tested the developed algorithm on the core hardware of autonomous car technology.

Pedestrian Pushbutton Activations

Several studies have examined the feasibility of using pedestrian crosswalk pushbuttons to estimate pedestrian counts at intersections. In this method, the number of crosswalk actuations must be factored up to account for cases in which more than one pedestrian crosses during each pedestrian crosswalk phase.  Several studies attempt to characterize this relationship between signal actuations and actual pedestrian counts.  In a limited pilot study in Portland, Blanc et al. (2015) finds 1.24 pedestrian crossings per pedestrian actuator press. Kothuri et al. (2016) also measure pedestrian actuated signal phases in Portland, Oregon. They show that is feasible to capture pedestrian actuated signal phases using existing technology and infrastructure for the evaluation of facility improvements. In NCHRP Report 797, Ryus et al. (2014) provide a thorough overview of actuation technologies for counting pedestrians and bicyclists.   Day et al. (2014) present a statistical regression method to predict single actuator counts as a function of weather, time of day, facility characteristics, and land use. While signal actuation counts are a noisy estimate of pedestrian counts, Day et al. argue that the method is useful for the measurement of relative changes in exposure due an intervention. It is our conclusion that more research is needed before signal actuation methods are widely implemented.

Naturalistic Data Collection for Pedestrians and Bicyclists

Naturalistic data collection refers to the unobtrusive observation of behaviors. This approach has been used in many fields outside of transportation for quite some time. Advances in sensor, computer, and telecommunication technologies now provide a method for automatically collecting detailed, objective information about a person’s driving performance (LeBlanc et al. 2006, 2007) and have led to a number of naturalistic driving studies (Campbell, 2012). At the same time, most have concentrated on motor vehicles, including cars, trucks, and motorcycles, and the primary focus has been on understanding safety-related behaviors rather than estimating exposure.

Thus, it is not surprising that most of the available exposure data on pedestrians and bicycles comes from sources other than naturalistic data collection (Vanparijs et al. 2015). These include various estimation models using: survey data such as the NHTS (Clifton and Krizek 2004; Salon and Handy 2014); data from pedestrian and bicyclist counts for relatively small areas and taken during one or a few points in time (Gallop et al. 2011; Griswold et al. 2011); and data for other location-specific characteristics such as employment, household size, and land use (Schneider et al. 2009).

In the limited cases in which naturalistic data have been used to examine pedestrian and/or bicycle behavior, the focus has typically not been on exposure per se but rather risk factors for collisions or crashes. Some of these studies involved the installation of unobtrusive video cameras at intersections to examine unsafe behaviors of cyclists such as red light running or helmet use (Johnson et al. 2008; Pai and Jou 2014; Wu et al. 2012). Most of the other studies involved the instrumentation of bicycles themselves; several of these latter studies are highlighted below.

Johnson et al. (2010) used helmet-mounted video cameras to capture the behaviors (head checks, reactions, maneuvers) of cyclists in Melbourne, Australia, that were associated with collisions, near collisions, and other events. Their data were limited to 127 hours and 38 minutes for 13 participants, and were not intended to generate information on exposure for this or a wider set of cyclists.

Dozza et al. (2012) described their efforts to adapt naturalistic driving methods to bicycles in Gothenburg, Netherlands, including the equipment used, data produced, tools and algorithms for visualizing and pre-processing the data, methods of analysis, and lessons learned. Results were reported for 12 participants, with data collection occurring over two weeks for each participant. The authors noted that their study complements the analyses of critical safety events by Johnson et al. (2010), by adding kinematics, location, and information about distraction by cyclists, leading to a better understanding of crash causation and providing a basis for deriving measureable safety indicators for development of intelligent countermeasures.

In related work, Werneke and Dozza (2014) used naturalistic cycling data from Gothenburg, in combination with crash, insurance, exposure, weather, road, and interview data to better understand single bicycle crashes and the factors contributing to them. According to the authors, naturalistic cycling studies can be used to fill in the gaps in reporting for single bicycle crashes and allow for the capture bicyclists’ behavior before and during safety critical events, overcoming the limitation of post-crash data sources.

Building on methods used in the 100-car study conducted by the Virginia Tech Transportation Institute, Jahangiri et al. (2015) developed a smaller data acquisition system (consisting of several sensors measuring acceleration, speed, and current location) that can be used to study bicyclists’ behavior and more specifically to develop bicycle violation prediction models. They also incorporated connected vehicle technology to facilitate communication within the system.

Investigators in Stockholm, Sweden, also used global positioning system logging devices and cameras to identify accessibility and safety problems among a sample of 16 commuter cyclists, as well as generate an accessible geographical interface for use as traffic planning tool (Gustafsson and Archer 2013). Specific study aims were to: identify the frequency and location of safety and accessibility/mobility problems and their causes; map and document these problems; help inform policy and strategy development; and help traffic planners better understand the problems faced by commuter cyclists on a daily basis. Participants were asked to ride 17 major cycle routes during the morning and afternoon peak traffic hours, resulting in the identification of more than 500 safety and accessibility/mobility problems.

In another naturalistic cycling study, Schleinitz et al. (2015) equipped three types of bicycles (conventional, pedelecs, and S-pedelecs) with a data acquisition system consisting of sensors to measure speed and distance, and two cameras. Using data collected from 85 cyclists over a period of 4 weeks each, the authors examined differences in speed and acceleration between bicycle types, as well effects of age and infrastructure features (e.g., bicycle lane, path, part of road used by cars, and foot path); they were not able to assess the implication of results on crash risk but speculated that higher speeds are associated with more severe crash outcomes.

A few studies used data from the naturalistic driving studies focusing on motor vehicles to examine various aspects of pedestrian/bicycle safety. For example, Lin et al. (2015) used data from the Strategic Highway Research Program 2 Naturalistic Driving Study to examine interactions between drivers and various pedestrian features at selected signalized intersections through which they drove. However, their interest was in assessing the effectiveness of selected pedestrian features in improving pedestrian safety and not estimating exposure. Other investigators used data from a 110-car naturalistic driving study in Indianapolis, Indiana, to identify potential conflict situations involving pedestrians and vehicles (Du et al. 2013; Tian et al. 2014); they concluded that this approach holds promise for studying pre-crash scenarios for vehicle to pedestrian collisions.

Collectively, the studies highlighted above suggest a promising start for applying naturalistic data collection methods to pedestrians and bicyclists to better understand safety-related problems, especially factors associated with crash and near crash risk. However, the use of naturalistic observation for these segments of the road user population is not without its challenges. For example, Dozza et al. (2012) pointed out that bicycle data loggers, unlike those used in cars or trucks, need to be small, light, waterproofed, tolerant to shocks and vibrations, low cost, and use as little energy as possible. In addition, they noted that higher drop-out rates might be expected for bicycling studies, compared to naturalistic driving studies, due to inclement weather and more convenient alternatives to bicycling. They also cautioned that privacy may be more of a concern for participants in naturalistic cycling studies, especially when cycles are taken indoors with cameras still recording, and that data loggers on cycles are more exposed to theft and tampering, especially when left outside and unattended.

The lack of studies using naturalistic data collection for the sole purpose of estimating exposure raises concern about whether this approach is a practical approach to measure pedestrian and bicyclist exposure. It may be more fruitful to explore naturalistic observation as a tool for supplementing other approaches to estimating exposure. For example, there may be a role for naturalistic data collection in helping to calibrate or validate estimation models for exposure and provide more robust measures of risk.

Summary and Conclusions

This chapter summarized numerous examples of exposure analyses that were conducted on specific transportation facilities. The facility-specific exposure analyses were most often used to identify high-priority locations for pedestrian and bicyclist safety improvements, and were typically conducted for an entire city. In some cases, the facility-specific information was also aggregated to provide overall trends for certain road types or for subareas within a city.

Data Sources and Methods

Most of the facility-specific exposure analyses used pedestrian and bicyclist count data from one or both of these sources:

For direct measurement of pedestrian and bicyclist counts, much progress has been made in the past ten years. Several companies now offer automated count equipment that helps to make pedestrian and bicyclist counting more efficient and cost-effective. The NBPD provided early guidance and helped to promote count data collection. Since then, FHWA has included a chapter specifically devoted to nonmotorized traffic monitoring in their 2013 TMG, and this chapter is likely to be updated in 2017. In 2014, NCHRP Report 797 Guidebook on Pedestrian and Bicycle Volume Data Collection was published and is a comprehensive resource on pedestrian and bicycle count data collection.

For estimating pedestrian and bicyclist counts, direct demand models have been the most widely used models for facility-specific exposure estimation thus far, and typically use regression analysis to relate directly measured counts to other measured attributes of the adjacent environment (e.g., land use and form, street type, etc.). Assuming that these measured attributes are available citywide, the regression model allows one to extend the sample of facility-specific counts to all facilities citywide. This chapter provided details on other types of modeling approaches that have been used on a limited basis, or could be used for facility-specific exposure estimation. These approaches include regional travel demand models, GIS-based models, trip generation and flow models, network analysis models, discrete choice models, and simulation-based traffic models.

Exposure Measures

Similar to the areawide exposure analyses, the units used in facility-specific exposure measures varied widely. Since the primary data source was pedestrian and bicyclist count data (rather than surveyed trip data in areawide exposure analyses), the units of exposure typically were a volume count for specified time period, or a distance traveled (calculated by multiplying a count by a street crossing width or road segment length). In a few cases where the exposure values were very high, the exposure was given in units of 1 million or 10 million (e.g., 1 million pedestrian miles traveled, 10 million pedestrian crossings).

Unlike areawide exposure analyses, several of the facility-specific exposure analyses did account for the interaction of nonmotorized and motorized traffic in their exposure measure. Several analyses computed the product of pedestrian and/or bicyclist traffic and motorized traffic ((P or B) × V) at intersections or other street crossings. A few other analyses used exposure measures like pedestrian crossings (or pedestrian miles) per entering motorized vehicle. Thus, having more granular exposure data on specific facilities does provide a better opportunity to quantify the level of interaction between pedestrian or bicyclist traffic and motorized vehicle traffic.

 < Previous Table of Content Next >
Page last modified on March 31, 2017
Safe Roads for a Safer Future - Investment in roadway safety saves lives
Federal Highway Administration | 1200 New Jersey Avenue, SE | Washington, DC 20590 | 202-366-4000