Guidebook on Pedestrian and Bicycle Volume Data Collection (2014)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

57 c h a p t e r 4 4.1 Chapter Overview This chapter discusses two types of factors that can be applied to raw count data: correction factors and expansion factors. Correction factors account for systematic inaccuracies in auto- mated counter technology. Expansion factors are applied to short-duration counts to estimate volumes over longer periods. The process of identifying potentially incorrect counts (e.g., due to a sensor being blocked, a battery becoming depleted, or other problems that prevent a sensor from “viewing” the activity scene correctly) is different from the process of correcting counts that have been collected by a fully engaged sensor, which is covered in this chapter. Techniques for identifying and removing potentially incorrect counts are discussed in Sec- tion 3.3.9, Cleaning and Correcting Count Data. 4.2 Sources of Counter Inaccuracy Several systematic sources of error are inherent in automated pedestrian and bicycle counts. This chapter presents methods for correcting raw automated count data to achieve more accurate volume estimates, but making a “correction” does not necessarily make a count perfectly accu- rate. The following are potential sources of counting errors that arise from automated counters. 4.2.1 Occlusion Some counter technologies count users who cross an invisible screenline. When two or more people cross the line simultaneously, an undercount occurs because the device only detects the person nearest the sensor. In previous research, this effect has been found to become more Adjusting Count Data Chapter 4 Topics • Sources of automated counter errors • Measured accuracy and precision of automated sensor technologies • Correction factors for automated counters, used to adjust raw counts to more closely represent the ground truth • Expansion factors, applied to short-duration counts to estimate volumes over longer periods of time • Example application of applying correction and expansion factors Source: NCHRP 07-19 field testing video. Example of occlusion of a passive infrared sensor. The red triangle indicates the sensor’s detection zone.

58 Guidebook on pedestrian and Bicycle Volume Data collection pronounced with higher volumes (i.e., errors are non-linear) (Ozbay et al. 2010, Schneider et al. 2012). This effect was observed for the passive infrared, active infrared, and radio beam sensors tested in NCHRP Project 07-19. The effects for passive infrared sensors, however, were somewhat ambiguous: a strong occlusion effect appeared for one product, while the other product did not demonstrate this effect as strongly. 4.2.2 Environmental Conditions Environmental conditions, such as weather and lighting, may cause counting inaccuracies in different counting technologies, because of particular characteristics of those technologies. Hot Temperatures Passive infrared and thermal sensors detect people based on the temperature gradient between persons and the background. If the ambient temperature is close to that of the surface of a human being, the gradient may be weak and hence detections may be missed. However, this effect was not observed during NCHRP Project 07-19 when temperatures exceeded 90°F. How- ever, some vendors advise that passive infrared sensors in direct sunlight can become overheated and significantly overcount. Cold Temperatures Andersen et al. (2014) performed controlled tests of pyroelectric (passive infrared) sensors in a very cold setting. They found that errors increased when a person wore a heavily insulat- ing down jacket and passed at the edge of the stated detection range of the counter. However, errors were not substantially greater than the vendor-specified ±5%. Passive infrared sensors tested as part of NCHRP Project 07-19 did not suffer accuracy problems at temperatures below 30°F, based on 11 hours of data with a mean temperature of 20°F and a minimum temperature of 10°F. Pneumatic tubes are suspected to undercount in cold temperatures as the rubber hardens, but this effect has not been formally documented. Pneumatic tubes are not recommended for use during snow or leaf-fall conditions—they can easily be damaged or dislodged by snow plows or street sweepers. Precipitation Precipitation has been suggested as a potential error source for all types of optical sensors (e.g., passive infrared, active infrared, and automated video). This effect was not observed with the passive and active infrared sensors tested by NCHRP Project 07-19, although only limited data Two Classes of Adjustment Factors Two types of factors can be applied to count data when developing volume estimates: • Correction factors are developed from validation counts and account for system- atic inaccuracies in counter technology. These factors are used to adjust the raw counts to more closely represent the ground truth. • Expansion factors are applied specifically to short-duration counts to estimate volumes over longer periods.

adjusting count Data 59 were collected during heavy precipitation events. Anecdotal reports from active infrared sensors have suggested very high overcount rates in heavy rain and snow. These errors are high enough (i.e., 3,000–5,000 people/hour) that they can be easily identified and cleaned out of the data. Lighting Low lighting has also been suggested as a problem for optical sensors; however, this effect was not seen in NCHRP Project 07-19 for active or passive infrared sensors. 4.2.3 Counter Bypassing Even though a counter may accurately count the pedestrians or bicyclists that pass through its detection zone, it may still not count all of the users if it is possible for users to bypass the detection zone. For example, inductive loop sensors may not cover the entire facility width, leaving “blind spots” on the edges of the path. A beam-type sensor mounted to a pole within the walkway will miss pedestrians who walk on the other side of the pole. Therefore, site selection should consider how easily persons can avoid being counted, also keeping in mind that some types of counters are more obvious than others, which may cause bicyclists to avoid them as potential hazards. 4.2.4 Mixed-Traffic Effects Two sets of bicycle-specific pneumatic tubes were tested by NCHRP Project 07-19 at a mixed- traffic site, Rue Milton in Montreal. The tubes at this site had higher accuracy and consistency results than bicycle-specific pneumatic tubes tested overall by NCHRP Project 07-19, despite a very large range of bicycle volumes. The data collection site experiences relatively low motor vehicle traffic volumes, so additional research is needed to explore the accuracy rates of pneu- matic tubes on shared-use lanes with higher motor vehicle volumes and speeds. Hjelkrem and Giæver (2009) tested two models of pneumatic tubes in mixed traffic and found bicycle count accuracy rates of -27.5% and -1.9%. They also tested inductive loops in mixed traffic and found a bicycle count accuracy rate of -16.5%. ViaStrada (2009) tested two models of inductive loops in New Zealand at four on-road sites, with accuracy rates ranging from -10% to +4%. Nordback et al. (2011) tested inductive loops capable of distinguishing bicycles from other vehicles on shared roadways and found a bicycle count accuracy rate of +4%. The TMG (FHWA 2013) suggests that magnetometers might not perform well for bicycle counting in mixed traffic with motor vehicles. 4.3 Measured Counter Accuracy Table 4-1 summarizes the accuracy and consistency of all the sensor technologies tested by NCHRP Project 07-19. Accuracy refers to how close, on average, the automated count is to real- ity. Consistency refers to the counter’s ability to reproduce consistent levels of accuracy across multiple observation periods. Three metrics are provided: • Average Percentage Deviation (APD) represents the overall divergence from perfect accuracy across all data collected. This metric does not differentiate over- from undercounting, which may tend to cancel each other out, resulting in the metric making a technology look more accurate than it actually is under some circumstances. • Average of the Absolute Percent Difference (AAPD) is a measure of the counter’s consistency. Greater consistency (i.e., a lower AAPD value) is useful, because it makes it possible to apply a single adjustment factor with confidence that the result will consistently be closer to the actual Source: NCHRP 07-19 field testing video. Example of a missed bicycle detection. The inductive loops capture bicyclists in the left-side bicycle lane, but not bicyclists riding with traffic on the right side of the street.

60 Guidebook on pedestrian and Bicycle Volume Data collection count. At higher values of AAPD, the counter is less consistent in its detection error, which means that more error will remain in the results after applying an adjustment factor. • Pearson’s Correlation Coefficient (r) is a measure of linear correlation between two variables. The values presented here are calculated between the hourly manual and automated counts for a given technology. Pearson’s coefficient can take values anywhere between -1 (perfectly negative linear correlation) and +1 (perfectly positive linear correlation). The closer this value is to +1, the more appropriate it is to use a multiplicative adjustment factor. The project’s final report (Ryus et al. 2014) describes how these metrics are calculated. Figure 4-1 demonstrates how to interpret APD, AAPD, and r values. This figure presents the passive infrared and active infrared sensor data collected during NCHRP Project 07-19. These devices have similar APDs, indicating similar levels of accuracy, but the passive infrared data Table 4-1. Accuracy and precision rates of sensor technologies tested by NCHRP Project 07-19. Sensor Technology APD AAPD Pearson’s r Hours of Data Hourly Volume (Avg./Max.) Passive infrared 8.75% 20.11% 0.9502 298 240 / 846 Acve infrared 9.11% 11.61% 0.9991 30 328 / 822 Radio beam 18.18% 48.15% 0.9503 95 129 / 563 Bicycle specific pneumac tubes 17.89% 18.50% 0.9864 160 218 / 963 Inducve loops (detecon zone)* 0.55% 8.87% 0.9938 108 128 / 355 Inducve loops (including bypass errors)* 14.08% 17.62% 0.9648 165 200 / 781 Piezoelectric strips 11.36% 26.60% 0.6910 58 128 / 283 Combinaon (pedestrian volume) 18.65% 43.78% 21.37% 47 176 / 594 Notes: APD = Average percentage deviaon, AAPD = average of the absolute percent difference, r = Pearson’s Correlaon Coefficient, Avg. = Average, Max. = Maximum. *Detecon zone results refer to the accuracy of the device with respect to the bicycle volume that passed through its detecon zone. Errors are larger when comparing the device’s count to the actual volume on the facility, including bicyclists that bypassed the detecon zone. Undercounting and Overcounting Undercounting is when fewer bicyclists and/or pedestrians are counted than actu- ally passed a site. Overcounting is when a technology records more bicyclists and/ or pedestrians than actually passed a site. Undercounting can and should be expected for even the most effective auto- mated counting technologies. It can occur for a number of reasons, such as occlu- sion (e.g., when pedestrians walking side-by-side pass a counter and the sensor detects one instead of two people), environment (e.g., a cold morning on which pneumatic tubes may not compress sufficiently to record the bicyclists riding across them), and other similar situations. Overcounting is a more significant issue because it is the result of false detec- tions. This is likely due to broken equipment (e.g., damage during transport) or improper installation (e.g., counter not calibrated correctly, site condition such as windows create false positives for infrared counters).

adjusting count Data 61 have a much higher AAPD. Similarly, the active infrared data have a correlation coefficient extremely close to 1. This difference can be seen in the plot as the greater degree of spread in the data. The lower AAPD indicates that one can have more confidence in the adjusted count result after applying an adjustment factor to the active infrared data, compared to the passive infrared sensor data. 4.4 Counter Correction Factors Correction factors correct for systematic counting errors associated with a particular technology, such as those associated with occlusion. Applying a correction factor produces a better estimate of the “true” count of pedestrians or bicyclists who passed through the counter’s detection zone. As discussed in Section 4.2.3, even a perfectly accurate counter may not count all facility users, if users can bypass the counter’s detection zone. These correction factors do not account for bypass errors, which are highly site specific. When working with technologies susceptible to bypass errors, count program managers are encouraged to develop site-specific factors to account for typical levels of bypass errors. 4.4.1 Correction Factors Developed Through NCHRP Project 07-19 Testing Table 4-2 presents simple multiplicative factors developed from the field tests conducted by NCHRP Project 07-19. These are multipliers applied to the raw count to estimate the true count. For example, if the raw count was 100 bicycles in an hour and the counting technology in use has a correction factor of 1.20, the estimate of the true count would be 120. The project’s final report (Ryus et al. 2014) provides details on how these factors were developed. When possible, products from more than one vendor were tested for each technology evalu- ated. TRB does not endorse specific products; therefore, individual product results are anony- mized in the table. Where results varied significantly between different products implementing the same technology, users are advised to use site- or product-specific calibration to develop accurate correction factors, because the errors seem to indicate potentially significant variation in performance among the vendor products available at the time that testing occurred. Site-specific calibration is also recommended when technologies shown in Table 4-2 were only tested in one location, resulting in insufficient data on whether or not the results are representative of the technology overall. The process of generating local correction factors is described in Section 4.4.2. Source: NCHRP Project 07-19 testing. (a) Passive Infrared (b) Active Infrared Figure 4-1. Comparative accuracy and precision of passive infrared and active infrared sensors.

62 Guidebook on pedestrian and Bicycle Volume Data collection Comparison with Previous Research Previous research has primarily focused on active and passive infrared, inductive loops, and pneumatic tube sensors. Consistent definitions of error rates and data collection protocols have not been used, which makes direct comparisons diffi- cult. NCHRP Project 07-19’s findings show similar or improved accuracy, compared with previous research. Passive infrared sensors were previously shown to undercount non-linearly (Schneider et al. 2012, Ozbay et al. 2010), which this study agrees with. The aver- age error rate of -9% determined by NCHRP 07-19 was roughly similar to or slightly better than findings in previous research, and Product A’s performance was better than this, suggesting that innovations in product design may have occurred since previous research was conducted. Previous work on active infrared (Jones et al. 2010) suggested a non-linear adjustment function, which was also found by NCHRP 07-19. Inductive loops appear to perform similar to or slightly better than the results presented in Nordback et al. (2011), although the detection zone was not clearly defined in the earlier study. Pneumatic tube findings are similar to those presented in Hjelkrem and Giæver (2009) and ViaStrada (2009). Table 4-2. Simple counter correction factors developed by NCHRP Project 07-19. Sensor Technology Adjustment Factor Hours of Data Passive infrared 1.137 298 Product A 1.037 176 Product B 1.412 122 Acve infrared* 1.139 30 Radio beam 1.130 95 Product A (bicycles) 1.470 28 Product A (pedestrians) 1.323 27 Product B 1.117 40 Bicycle specific pneumac tubes 1.135 160 Product A 1.127 132 Product B 1.520 28 Surface inducve loops 1.041 29 Embedded inducve loops 1.054 79 Piezoelectric strips* 1.059 58 Combinaon (pedestrians) 1.256 47 Notes: *Factor is based on a single sensor at one site; use cauon when applying.

adjusting count Data 63 Multiplicative factors, as shown in Table 4-2, are the easiest form of correction factor to inter- pret and estimate. For most applications, using simple multiplicative factors is sufficient. How- ever, in some cases, automated data can be corrected to be closer to the ground truth volume by introducing higher order terms, interaction terms with environmental factors affecting counter accuracy, or both. More complicated functional forms have been estimated using the NCHRP Project 7-19 data. A selection of these is presented in Table 4-3 and the full set can be found in the project’s final report (Ryus et al. 2014). The functions presented here are limited by the amount of data collected during the project and are recommended to be revisited as more validation data becomes available. For an example application of the functions presented in Table 4-3, take the third passive infrared equation. In equation form, this would be ( ) = × − × + × × 1.125 3.358 10 0.015 2 4 Manual Count Automated Count Automated Count Automated Count Facility Width The models presented in Table 4-3 are shown with the Akaike Information Criterion (AIC), a measure of model fit used in comparing multiple models for the same dataset. The AIC repre- sents the level of unexplained variation in the data and is penalized by adding additional terms to the model—hence, the lower the AIC value, the better fitting the model. However, AIC values are only meaningful when used in comparing multiple models on the same dataset, not as an absolute measure of fit. As discussed in Section 4.2, despite previous findings in the literature in some cases, various environmental factors were not found to affect accuracy, including the following: • Hot and cold temperature effects with active and passive infrared sensors, • Cold temperature effects with pneumatic tubes, and • Lighting effects with active and passive infrared sensors. Technology Automated Count Automated Count2/104 Facility Width () × Automated Count Temperature ( F) × Automated Count AIC Passive infrared 1.137 3288 1.313 3.995 3258 1.125 3.358 0.015 3237 Acve infrared† 1.139 232 1.100 0.787 230 1.413 0.868 0.004 219 Radio beam 1.130 987 1.857 0.053 986 Pneumac tubes* 1.135 1584 Inducve loops 1.050 829 0.906 0.685 771 Piezoelectric† 1.059 607 1.562 3.246 594 Notes: *Bicycle-specific products. †Factor is based on a single sensor at one site; use cauon when applying. AIC = Akaike Informaon Criterion. Table 4-3. Counter correction factors developed by NCHRP Project 07-19.

64 Guidebook on pedestrian and Bicycle Volume Data collection Possible explanations for differences between the NCHRP Project 07-19 results and previous results in the literature on environmental effects on count accuracy include (1) improvements in how vendors have implemented technology and (2) insufficient observations in the NCHRP Project 07-19 dataset for specific environmental conditions (e.g., heavy rain or heavy snow). Based on the NCHRP Project 07-19 results, a simple multiplicative factor is sufficient in most cases for improving the accuracy of data collected with automated counters. However, the more complicated correction functions in Table 4-3 are recommended to be considered for higher volume sites when using passive infrared sensors, active infrared sensors, inductive loops, and piezoelectric strips; to correct for facility width when using passive infrared and radio beam sensors; and to correct for temperature when using active infrared sensors (although the effect is not strong). In either case, using a locally developed, site-specific correction factor or function is preferable to using a factor from Table 4-2 or 4-3, because a locally developed factor will account for the unique characteristics of a particular counting product installed at a particular site. Sec- tion 4.4.2 describes how to create site-specific correction factors. Additionally, it is important to develop factors to account for bypass errors when using devices susceptible to this effect (e.g., inductive loops and pneumatic tubes). Inductive loops yielded a wide range of bypass error rates during NCHRP Project 07-19 testing, which appeared to vary according to individual site characteristics. The sites with the highest bypass error rates were two on-street bicycle lanes (APD = -46.62% and -25.49%), where bicyclists frequently traveled into the shared-use lane, and the Midtown Greenway in Minneapolis (APDs = -37.32% and -20.29%) where both sets of inductive loops being tested were substantially narrower than the facility width. Lower bypass errors occurred on the Loyola multi-use path in Davis, California (APD = -1.56%); on the Key Bridge between Arlington, Virginia, and Washington, D.C.; (APD = 7.95%), and the Rue University cycle track in Montreal, Quebec (APD = -2.44%). The latter facilities had inductive Why Develop Site-Specific Factors? The correction factors presented in Tables 4-2 and 4-3 were developed from NCHRP 07-19 field testing. Depending on the specific counting technology, the factors are based on data from 1–6 sites using 1–2 products implementing the technology. However, different errors can occur, based on a site’s unique traffic and environ- mental characteristics. For example, sites where occlusion is more likely to occur (e.g., due to high volumes, wide facilities, frequent grouping of pedestrians) will tend to require a larger upward adjustment. The NCHRP 07-19 testing also found substantial differences in the required cor- rection factor between different products implementing the same sensor tech- nology. It can also be expected that vendors will continue to work to improve their products and that new products will come onto the market. Finally, while the NCHRP 07-19 testing generally found linear relationships between correction factors and volume, there is the potential for a non-linear relationship to exist. Therefore, it is recommended that organizations develop site- or device-specific correction factors for their counters.

adjusting count Data 65 loops that extended over most of the facility width, as well as channelization features that forced bicyclists to ride over the sensors. 4.4.2 Developing Site-Specific Correction Factors Developing a local correction factor requires conducting manual counts and comparing the results to the automated counter data. The comparison data can be collected with traditional in-the-field manual counts or by developing manual counts from video footage. Conducting manual counts in the field can be more straightforward, but is prone to error during high- volume situations and must be completed in real time (Diogenes et al., 2007). Video-based man- ual counts require field trips to set up and take down the video camera, but the resulting video can be sped up or slowed down to accurately and efficiently accommodate high- and low-volume time periods. Appendix C provides the count protocol used in this study, which includes details of how to conduct quick computer-based manual counts from video images. One of the primary challenges in collecting data for developing correction factors is synchro- nizing the counter’s clock with the video camera’s clock. If there is a high degree of confidence that the two clocks are synchronized, then shorter count intervals (e.g., 15 minutes) are recom- mended, assuming the counter can produce the shorter interval. This approach reduces the time required to develop the correction factor, compared to using 60-minute count intervals. However, if there is a lower degree of confidence in the synchronization of the two clocks, longer count intervals (e.g., 60 minutes) should be used. The reason for the different count intervals is best explained through an example. If a counter and camera happen to be about 20 seconds off, this condition may result in one or two persons miscategorized by time in each interval. At a location with 120 bicycles an hour, a 15-minute count might be approximately 30 bicyclists. However, if two bicyclists are missed in 15 minutes, the error in the “ground truth” count would be approximately 6%, while for an hour interval, missing two bicyclists would result in an error of slightly less than 2%. A minimum of 30 time periods worth of ground truth data (i.e., approximately 8 hours of counts when 15-minute data are collected, or 30 hours of counts when 60-minute data are Device Accuracy and Detection Zones To determine a counter’s accuracy, it is necessary to have a clear understanding of its detection zone. For example, inductive loops will capture bicyclists that ride over or in the immediate vicinity of the loops. To measure the accuracy of the loop technology, only those bicyclists that rode through the loops’ detection zone should be compared to what the loops sensed. A product’s vendor should be able to specify the product’s detection zone. To determine how effectively a counter captures the complete picture of bicycle or pedestrian activity at a specific site, the site’s true bicycle and/or pedestrian volume (as determined from manual counts in the field or from video) should be volumes recorded by the counter. If relatively large volumes of pedestrians or bicyclists are not being counted because they are not passing through the coun- ter’s detection zone, a different sensor technology, or a different implementation of the same technology, may be needed to capture an accurate picture of the site’s activity.

66 Guidebook on pedestrian and Bicycle Volume Data collection collected) is recommended when developing correction factors. Time periods should include a range of volumes, including some time periods when peak volumes occur. These ground truth validation counts can be collected in person or using video footage, as detailed in Appendix C. Once the ground truth data have been collected, they can be plotted against the automated counter’s counts, as illustrated in Figure 4-2. If the counts generally appear to fall along a straight line, as in Figure 4-2(a), then the correction factor is calculated as the ground truth count divided by the recorded count. If the counts appear to curve or follow some other non- linear shape, as in Figure 4-2(b), then statistical methods will be needed to fit a curve to the pattern (a spreadsheet’s curve-fitting tools may be sufficient for this purpose). Finally, if the count pattern appears to be more of a “cloud,” as in Figure 4-2(c), then the counting device may not be installed or calibrated properly and should be adjusted prior to collecting new ground truth data. As previously noted, the accuracy of any counting technology with a limited detection zone (e.g., inductive loops) can be quantified in one of two ways: (1) comparing counts over the entire facility width or (2) comparing counts just within the counter’s detection zone. The former is the preferred option (provided there is a predictable detection aversion factor) when developing site-specific cor- rection factors, because the resulting factor will account for errors resulting from the sensor itself as well as site-specific errors resulting from persons passing outside the device’s detection zone. The lat- ter is the preferred option when a device will be used for short-term counts at a variety of locations. 4.5 Expansion Factors Expansion factors are used to estimate pedestrian or bicycle volumes under conditions differ- ent than actually counted. Similar to the well established procedures for estimating long-term motorized traffic volumes, permanent counters are used to develop these factors. Expansion factors include the following types of adjustments: • Temporal adjustments. Temporal adjustments are used to estimate volumes at a different time, or for a longer time period, than was counted. A common application is to expand a short- term count to an estimate of annual volume. • Environmental adjustments. Environmental adjustments are used to estimate what the counted volume would have been under different conditions than occurred during the count. For example, a count taken during rainy, hot, or windy conditions could be adjusted to estimate the volume that would have been seen on a good weather day. Figure 4-2. Examples of potential scatterplots when developing correction factors. (a) Approximately linear (b) Non-linear (c) Cloudy

adjusting count Data 67 • Land use and facility type adjustments. These adjustments can be used to account for differ- ences in volumes attributable to differences in the surroundings of a count site, compared to a continuously counted control site. Expansion factors are typically applied to sites sharing an activity profile (e.g., commuter vs. recreational route or shopping district vs. residential area) similar to that of the continuously counted site(s) used to develop the expansion factor. Groups of sites sharing similar profiles are also known as factor groups; Chapter 4 of the Traffic Monitoring Guide (FHWA 2013) provides more details on how to identify factor groups. 4.5.1 Temporal Adjustment Factors Temporal adjustment factors are used to account for peaking patterns—the tendency for pedestrian or bicycle volumes to be distributed unevenly throughout the day, week, or year. For example, high pedestrian volumes may be present on sidewalks in a CBD at 5 p.m., with low volumes present at 3 a.m. A popular recreational trail may have higher bicycle volumes on weekends than weekdays. Adjustment factors can be applied to short-duration counts to estimate volumes over a longer time. For example, if counts are collected at a site for 1 week in January, the data represent 1 week of activity, but not necessarily a representative or “average” week. Therefore it is not appropriate to multiply the count total by 52 and expect to arrive at a reasonable annual estimate. Depending on the region and associated climate where the count was taken, seasonal impacts on bicycling and walking will vary significantly. Therefore, it is necessary to develop a seasonal adjustment factor based on continuously monitored control sites so as to provide a more accurate under- standing of how a week in January compares to a typical week during the year. At the most basic level, permanent count data can be collected at one location in a city, with the observed year-round patterns used to expand short-term counts at other sites in the city. However, when possible, multiple permanent count sites should be used because various activity profiles inevitably exist within a given city. An ideal number of permanent count stations has not been identified. The Traffic Monitoring Guide (FHWA 2013) recommends three to five continu- ous count stations per factor group, but recognizes that the limiting factor in most cases will be the available traffic monitoring budget. Estimating Pedestrian Volumes Around a University Campus Researchers at the University of California, Berkeley used a combination of short- duration (2-hour) manual counts and continuous automated counts to estimate pedestrian volumes around the campus periphery as a proxy for exposure. Three passive infrared sensors were mounted at locations near the edges of cam- pus for 15 months to estimate temporal adjustment factors specific to the three edges. Two-hour pedestrian crossing counts were collected at most intersections around the campus. Pedestrian volumes for a 10-year period were then estimated at each intersection based on the short-duration intersection counts and the adjustment factor for the appropriate campus edge. The 10-year volumes were used to normalize the number of pedestrian-involved crashes at each intersection to identify the riskiest locations for pedestrians around the campus periphery.

68 Guidebook on pedestrian and Bicycle Volume Data collection Temporal expansion factors are simple to apply. A look-up table is generated from the per- manent count data, giving the percentage of the total volume observed during each time period. For example, Table 4-4 contains illustrative expansion factors from a site in San Francisco and a site in Arlington, Virginia, both of which count bicycle volumes in a bicycle lane. These factors could be used to adjust monthly counts at sites having similar characteristics to one of these loca- tions. Based on the data shown in the table, bicycle volumes at the Arlington site show a stronger seasonal variation than those at the San Francisco site. These factors can be refined and improved over time by analyzing data from more sites over a greater period time, ideally reaching a point where the activity profile becomes highly predict- able and thus more reliable for factoring short-duration counts. In further refining these factors, it is important to consider how the surrounding land uses will influence bicycle and pedestrian activity. For example, recreational areas are likely to have different temporal peaking character- istics for pedestrian and bicycle activities than CBDs. One would expect to see higher peaking characteristics for recreational pedestrian and bicycle activity during non–work days and hours (e.g., weekends and evenings), and one would expect to see higher peaking characteristics for a CBD during weekday commuting periods. An expansion factor is calculated as follows for a given site: 1 Volume Volume t ii N t ∑γ = = where gt = the expansion factor for time period t, Volumei = the volume during time period i, Volumet = the volume during time period t, and N = the set of all time periods i. In Table 4-4, N is the set of all months in the year. However, N could well be the set of all hours in a day, days in a week, or any other conceivable time period. As an example of how to use this formula, the expansion factor for the San Francisco site in January is calculated as follows: the sum of the monthly volumes over the course of the year was 552,592; the monthly volume for Month San Francisco Arlington, Virginia Monthly Volume Percentage of Annual Volume Expansion Factor Monthly Volume Percentage of Annual Volume Expansion Factor January 44,143 8% 12.52 2,985 5% 20.85 February 44,254 8% 12.49 2,450 4% 25.40 March 38,600 7% 14.32 3,112 5% 20.00 April 49,366 9% 11.19 5,581 9% 11.15 May 55,143 10% 10.02 6,298 10% 9.88 June 53,335 10% 10.36 6,408 10% 9.71 July 52,788 10% 10.47 6,815 11% 9.13 August 54,522 10% 10.14 7,697 12% 8.08 September 52,054 9% 10.62 7,701 12% 8.08 October 42,393 8% 13.03 6,136 10% 10.14 November 35,651 6% 15.50 3,983 6% 15.62 December 30,343 5% 18.21 3,062 5% 20.32 Total 552,592 100% 62,228 100% Table 4-4. Example temporal expansion factor calculation.

adjusting count Data 69 January was 44,143; and (552,592) / (44,143) = 12.52, which is the expansion factor for January shown for the San Francisco site in Table 4-4. Once a set of expansion factors has been developed, the factors can be used to expand a short- term count at a comparable site. For example, 1 month’s worth of pneumatic tube data might be collected in June at another site in San Francisco with characteristics similar to the permanent count site. First, the raw data from the counter would be corrected to account for systematic errors associated with the tube counter, using data from Table 4-2, Table 4-3, or a local correc- tion factor, if available. The corrected counts could then be multiplied by the expansion factor for June—in this case, 10.36—to estimate annual bicycle volumes at the short-term count site. If the corrected monthly count was, for example, 31,570 bicyclists, then the estimate of annual average bicycle traffic (AABT) would be 31,570 × 10.36, or 327,065. As noted in Section 3.2.3, short-term counts are most accurately expanded when the original count covers a substantial period of time (e.g., weeks). Even so, the estimated long-term volume may be substantially different from the true volume. Improvements on basic day-of-week and month-of-year expansion factors can be made in several ways. First, additional multiplicative fac- tors can be applied to correct for weather patterns or land use characteristics. Second, separate fac- tor groups can be estimated to account for different activity patterns based on land use and facility type. These types of factors are discussed in following subsections. Third, a day-of-year expansion factor can be applied in lieu of the traditional combination of day-of-week and month-of-year factors. This approach has been shown to substantially outperform the traditional approach in both bicycle-only monitoring situations and in mixed bicycle and pedestrian situations (Hankey et al. 2014, Nosal et al. 2014). Further details of this approach are provided in Appendix D. 4.5.2 Land Use Adjustment Factors Land use adjustment factors can be used to account for differences in volumes that are attrib- utable to differences in the surroundings of a count site, compared to a continuously counted To Round or Not To Round? Remember that any volume estimates based on extrapolation are just that— estimates. They are by no means precise, and to present them as such can be misleading. When presenting volume estimates, they should be rounded and specified as approximate values. For example, if the 327,065 bicyclists/year volume presented in this section were calculated, it should be presented as “approximately 327,000/ year.” This will help to avoid giving false impressions of the degree of accuracy of the estimates. However, if volume estimates are being used as an intermediate step in a cal- culation (such as when they are used for exposure in a pedestrian crossing risk evaluation), rounding should not be performed. There is no reason to say that 327,000 bicyclists/year is a more accurate value than 327,065 bicyclists/year, so the unrounded value should be passed to the risk calculation. Specific applications (e.g., reporting to the FHWA) may specify that volume should not be rounded (or vice versa), and specific agencies may have their own adopted rounding procedures (often based on AASHTO [2009] guidance).

70 Guidebook on pedestrian and Bicycle Volume Data collection control site. This approach is more robust for estimating pedestrian volumes than bicycle vol- umes, because bicycle trips tend to be longer and more “through traffic” is counted that does not necessarily have much relation to the surrounding land use characteristics. These factors are separate from the time-of-day factors. Possible land uses to control for include broad characterizations (i.e., residential, industrial, and CBD) (Hocherman et al. 1988), socioeconomic variables (e.g., population density, employ- ment density, and median household income) (Ryan et al. 2014), or degree of land use mixture (e.g., entropy index, dissimilarity index, interspersion index, juxtaposition index, and contagion index) (Cervero 1989, Cervero and Kockelman 1997, Hess et al. 2001). 4.5.3 Weather Adjustment Factors Counts can be adjusted based on the weather patterns for the day on which counts are col- lected. For example, if volumes in a city are observed to drop to 60% of “normal levels” during rainy weather, and short-duration counts are conducted on a rainy day, the estimate of annual traffic should be adjusted upward by 67% (i.e., 1/0.6) to account for weather. Weather-related phenomena have been documented in the literature. Table 4-5 presents a summary of weather-related effects on volumes. The literature review found in this project’s final report (Ryus et al. 2014) summarizes the findings of the studies cited in Table 4-5. If continuous counts are available for the entire year, the day-of-year method described in Appendix D can be used, because it incorporates weather effects that may have occurred on any given day. 4.6 Example Application of Factor Adjustment Methods The following example uses 1 month of data collected in October 2013 using a passive infra- red sensor at the study site, located on a multi-use path on a university campus. This example is intended as a simplified hypothetical exercise of working with raw data to arrive at an estimate of annual volumes. Step 1. Check Data for Anomalies Figure 4-3 plots the raw data downloaded from the counter at the study site for the month of October 2013. The counter appears to have been out of commission for a few days—it is unlikely that the volumes would be 0 for such a long period. To correct, one approach is to impute from sur- rounding data. However, one should be careful not to interpolate from a special event, weather anomaly, or other special situation that might lead to drastically “abnormal” volumes. In this case, the analyst is reasonably certain that the data from the previous and following week- ends reflect typical conditions and can be used. So, the average values for each of those days are taken. For example, the volume from 10–11 a.m. on Sunday, October 13 is estimated as the average of the volume from 10–11 a.m. on Sunday, October 6 (46) and the volume from 10–11 a.m. on Sunday, October 20 (41), resulting in an estimated volume of 44. These imputed values are shown by the blue dashed line in Figure 4-4. (In this simplified example, only the preceding and following Sunday are used; the approach recommended in Section 3.3.9 would use 10–11 a.m. data from the four preceding and four following Sundays.)

Table 4-5. Weather-related effects on non-motorized volumes documented in the literature. Weather Factor Source Descripon User Type Effect Rain Cameron, 1977 Shopping districts inSea le, WA P 0.05 in/day reduced traffic by 5% below average in summer, no effect in December Precipita€on events Aultman Hall et al.,2009 CBD in Montpelier, VT P 13% lower average volumes during precipita€on Cloudy Schneider et al., 2009 Counts throughoutAlameda County, CA P Mul€plica€ve factor of 1.05 Cool temperatures ( 50°F) Schneider et al., 2009 Counts throughout Alameda County, CA P Mul€plica€ve factor of 1.02 Hot temperatures ( 80°F) Schneider et al., 2009 Counts throughout Alameda County, CA P Mul€plica€ve factor of 1.04 (hours 1200–1800) Mul€plica€ve factor of 0.996 (hours 0000–1200 and 1800–2400) Rain Schneider et al., 2009 Counts throughoutAlameda County, CA P Mul€plica€ve factor of 1.07 Temperature (<28°C) Miranda Moreno and Nosal, 2011 Separated bicycle facili€es in Montreal, QC B 10% increase corresponds to 4–5% volume increase Temperature 28°C and rela€ve humidity 60% Miranda Moreno and Nosal, 2011 Separated bicycle facili€es in Montreal, QC B Volume decrease of 11–20% Humidity Miranda Moreno andNosal, 2011 Separated bicycle facili€es in Montreal, QC B 100% increase corresponds to 43–50% decrease in volume 3 hour lagged precipita€on Miranda Moreno and Nosal, 2011 Separated bicycle facili€es in Montreal, QC B 25–36% reduc€on in volume 1 hour lagged precipita€on Chapman Lah€ and Miranda Moreno, 2012 Pedestrians in Montreal, QC P 8% decrease on weekdays; 11% decrease on weekends Precipita€on Flynn et al., 2012 Bicyclists via survey B Twice as high likelihood with no morning precipita€on Temperature Lewin, 2011 Bicycle volumes B Highly correlated (R2 = 0.50) Notes: P = pedestrian, B = bicyclist, CBD = central business district.

72 Guidebook on Pedestrian and Bicycle Volume Data Collection Step 1A. Establish Site-Level Data Correction Factors Manual counts can be conducted at the study site and compared with the automated data from the counter. These counts are used to estimate a site-specific correction factor, as discussed in Section 4.4.2. Manual counts are collected in 15-minute intervals for 6 hours at the study site, for a total of 24 data points from which to calculate a correction factor. (This simplified example uses fewer data points than the recommended 30.) Figure 4-5 compares the manual and automated counts. The dashed line corresponds to per- fect accuracy. In other words, data points (corresponding to 15-minute count intervals) falling below the dashed line represent periods when net undercounts occurred, while data points fall- ing above the line represent periods when net overcounts occurred. The thick solid line shows Figure 4-4. Raw count data with missing data imputed. Figure 4-3. Raw data from the counter.

Adjusting Count Data 73 the adjustment function, which is a mirror image of a best-fit line through the data points. Automated counts should be multiplied by the slope of the thick solid line to bring them closer to the perfect accuracy line. Based on the slope of the adjustment function shown in Figure 4-5, a correction factor of 1.167 (e.g., 70/60) is estimated. Step 2. Correct Data with an Appropriate Correction Factor In the next step, the raw count data are corrected for undercounting, using either the site- specific correction factor developed in Step 1A, or using a general factor or function such as those presented in Tables 4-2 and 4-3. To use a correction factor, multiply the raw automated count values by the adjustment factor. If local data suggest a non-linear relationship between the manual and automated data and, therefore, that an adjustment function would be more appropriate, the raw automated count values and any other variables used in the function are provided as inputs to the function. Figure 4-6 shows the corrected count data, after the correction factor of 1.167 that was deter- mined in Step 1A is applied. Step 3. Expand to Annual Volumes In this step, unless the annual volumes at the study site are known, an assumption must be made about which factor group the site belongs to. Given that this site is on a university campus (where people have very irregular schedules), its volume profiles do not fit cleanly into a “traditional” utili- tarian or recreational pattern. These monthly counts should therefore be expanded using continu- ous data from an automated count station on a mixed-use path at another university or college in a similar climate (or, even better, another site on the same university campus). To demonstrate how to apply these expansion factors, assume the volumes shown in Table 4-6 have been counted by an automated count station on another mixed-used path on the same university campus. The corrected monthly volume for the study site in October was 50,232. Given that this site is on the same university campus and on the same type of facility (a mixed-use path) as the automated count station site, the study site’s October volume can be multiplied by the count sta- tion’s expansion factor for October to obtain an estimate of AADT. In this case, 50,232 × 10.431 = 523,970, which is rounded to the nearest thousand (i.e., 524,000) in consideration of the various assumptions involved. Figure 4-5. Example calculation of a local correction factor.

74 Guidebook on Pedestrian and Bicycle Volume Data Collection Next, consider what could be done if only the first 8 days in October had been counted at the study site, resulting in a corrected volume of 14,031 for those days. One way to handle this would be to assume that all of the 31 days in October are roughly equal in terms of volume and expand the corrected volume by the proportion of the month it represents. In this case, the estimated annual volume would be (14,031) × (31/8) × (10.431), or approximately 567,000. This result is different than the estimate that used a full month’s worth of data. An alternative approach, given the availability of daily volumes from the automated count station, would be to use the day-of-year approach (described in Appendix D). From Table 4-6, the annual volume at the count station was 460,531. If the total volume at the count station for the first 8 days of October was 11,073, then the AADT for the study site could be estimated as (14,031) × (460,531/11,073), or approximately 539,000, which is closer to the estimate based on an entire month’s worth of data. Monthly Volume Fracon of Year Extrapolaon Factor January 30,513 6.63% 15.087 February 43,101 9.36% 10.681 March 51,029 11.08% 9.021 April 52,049 11.31% 8.845 May 60,324 13.10% 7.631 June 29,210 6.35% 15.760 July 24,021 5.22% 19.165 August 21,031 4.57% 21.889 September 56,201 12.21% 8.191 October 44,134 9.59% 10.431 November 38,519 8.37% 11.951 December 10,219 2.22% 45.049 Note: These data have been invented for this example and should not be used as extrapolaon factors for an actual facility. Table 4-6. Example count station data for university campus mixed-use path. Figure 4-6. Count data after applying a correction factor.