Methods and Technologies for Pedestrian and Bicycle Volume Data Collection: Phase 2 (2017)

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.

NCHRP 07‐19(02) Final Report 1 Abstract  This report documents the research conducted by NCHRP Project 07‐19(02), Methods and Technologies for Pedestrian and Bicycle Volume Data Collection: Continuation. This continuation project tested and evaluated automated count technologies that capture pedestrian and bicycle volume data, focusing on technologies that came onto the market too late to be included in the Phase 1 research (NCHRP Project 07‐19). The report presents combined results for a range of technologies tested by both the Phase 1 research and the continuation research (Phase 2). The research evaluated automated nonmotorized count technologies in different settings, including ranges of temperature, varying weather conditions, mixed traffic conditions, mixed travel directions, and facility types (e.g., roadways, multiuse paths), to determine their accuracy and reliability in different contexts. This report documents the research findings on the accuracy and consistency found for the different automated count technologies. It provides a complete account of the process used to select technologies for testing, identify test sites, and evaluate the effectiveness of the technologies. It is clear from the testing that careful site selection plays an important role in the ultimate accuracy of the collected count data. It is also critical for practitioners to calibrate the counters they install at specific sites to obtain the most accurate and reliable results.

NCHRP 07‐19(02) Final Report 2 Summary  NCHRP Project 07‐19 evaluated six automated count technologies that capture pedestrian and/or bicycle volume data. The results of that research were presented in NCHRP Report 797: Guidebook on Pedestrian and Bicycle Volume Data Collection (Ryus et al. 2014a) and NCHRP Web‐Only Document 205: Methods and Technologies for Pedestrian and Bicycle Volume Data Collection (Ryus et al. 2014b). The research conducted by NCHRP Project 07‐19 is referred to as Phase 1 in this report. The research described in this report, NCHRP Project 07‐19(02), evaluated five additional automated count devices, representing four different detection technologies, that entered the market after the start of Phase 1 or were otherwise unable to be studied during Phase 1. This continuation research is referred to as Phase 2 in this report. This summary section summarizes the technologies tested and test sites used during Phase 2, and updates NCHRP Web‐Only Document 205 by providing the combined results of the research efforts. Chapter 1 provides background information about NCHRP Project 07‐19. Chapters 2 and 3 focus on the project’s testing approach and findings, combining results from Phases 1 and 2. Chapter 4 presents the research conclusions and suggestions for additional future research. COUNTING TECHNOLOGIES TESTED  The counting devices tested in Phase 2 represent products and counting technologies that entered the market too late to be included in the Phase 1 testing effort, technologies for which only a limited amount of data were able to be collected during the original research, or both. Based on conversations with product vendors, discussions with the project panel, and the available budget, the following technologies were selected for testing as part of Phase 2:  Thermal imaging camera. These devices combine the technologies of passive infrared detection and automated counting from imaging. The camera, mounted overhead for the tested application, detects the infrared radiation (i.e., heat) given off by pedestrians and bicyclists, and the system counts the number of heat‐emitting objects that pass through a defined detection zone within the camera’s field of view.  Radar. These devices operate by emitting electromagnetic pulses and deducing information about the surroundings based on the reflected pulses. The device that was tested is designed to be buried in the pavement. Although there are also above‐ground radar devices on the market, none were tested in this study.  Bicycle‐specific pneumatic tubes. Pneumatic tubes detect the pulses of air generated when a vehicle or bicycle rides over the tube. Many transportation agencies already use standard pneumatic tubes, designed for motor vehicle counting, as part of their motorized counting programs, and are familiar with how to install and use them. However, this project investigated pneumatic tubes with a smaller profile that are specifically designed to count bicycles. Bicycle‐specific pneumatic tubes are designed to differentiate between bicycles

an al  P ru co se th p Figure S‐ are define (a) (c Figure S‐ d motor veh l vehicles wi iezoelectric nning over mbination p parate pass at pass the s assing the co 1 illustrates d and illust  Thermal imag ) Bicycle‐speci 1.  Phase icles in mix th a classific strips. Thes them. One of edestrian an ive infrared ensor, and t unter. the sensor te rated in NCH ing camera (c fic pneumatic   2 Counting ed traffic, pr ation schem e strips emi the two dev d bicycle co sensor detec he differenc chnologies t RP Report 7 ircled) tubes  Technologi 3 oducing eith e that separa t electrical s ices tested d unter. The p ts the heat e e in the two ested in Pha 97 (Ryus et a es Illustrate N er counts of tely identifi ignals when uring this re iezoelectric mitted by bo counts gives se 2. Techno l. 2014a). (b (d) Piezoe and pa d  CHRP 07‐19 bicycles alon es the numb deformed by search was strips detect th pedestria the number logies teste ) Radar (circle lectric strips ( ssive infrared  (02) Final R e, or counts er of bicycle bicycle wh part of a bicycles, a ns and bicy of pedestria d in Phase 1 d) rectangle)  (circled)  eport of s. eels clists ns only

NCHRP 07‐19(02) Final Report 4 Table S‐1 summarizes the amount of data evaluated in Phase 2, categorized by the environmental and user volume conditions under which the data were collected. While efforts were made to test the technologies under a range of conditions, no cold‐weather testing occurred in Phase 2 due to a limited project budget and scope. Table S‐1.  Counting Technologies Tested in Phase 2 by Environmental and User Volume  Conditions  Pneumatic Tubes  Condition  Radar  Thermal  Camera  Piezo‐ electric  Strips  Passive  Infrared  Bike‐ Specific  Standard  Total hours of data evaluated  32  28  78  39  43  17  Temperature (°F) (mean/SD)  60 / 5  54 / 12  73 / 13  73 / 13  56 / 10  64 / 4  Temperature range  (°F)  48–72  40–83  48–93  48–93  40–74  58–72  Hourly user volume  (mean/SD)  69 / 50  100 / 97  90 / 49  112 / 54  76 / 54  61 / 32  Hourly user volume range  1–205  8–340  21–181  36–210  8–223  15–157  Nighttime hours  2  5  4  2  6  0  Rain hours  3  10  2  1  8  2  Cold hours (<30 °F)  0  0  0  0  0  0  Hot hours (>90 °F)  0  0  8  4  0  0  Thunder hours  0  1  2  1  1  0  Note:  SD = standard deviation.  SITE SELECTION  Three jurisdictions were selected for testing the devices in Phase 2 on the basis of (1) the researchers having existing staff contacts at the jurisdictions who could facilitate obtaining any required permits to install the devices, and (2) the locations being close to the vendors of the devices, allowing easier coordination for installation, calibration, and (if necessary) troubleshooting. These jurisdictions were:  Arlington County, Virginia;  Washington, D.C.; and  Oakland, California. The Arlington site was the same site on the Four Mile Run multiuse trail used in Phase 1. The site was selected as both piezoelectric counters being tested were already located there, one having been installed by Arlington County prior to Phase 1 and the other having been installed (but not able to operate properly) during Phase 1. The test sites in the other two cities were on‐road bicycle facilities selected in consultation with local agency staff and the count device vendor on the basis of:

NCHRP 07‐19(02) Final Report 5  From the researchers’ point of view, locations with relatively high bicycle volumes, to allow the devices’ accuracy and precision to be tested under a range of volume conditions;  From the jurisdictions’ point of view, locations that would provide useful data for them in the future, as they would take over ownership of the counter following the tests; and  From the vendors’ point of view, locations that complied with the vendors’ guidance for suitable installation locations for the sensors and their supporting technology (e.g., detection area dimensions, access to power, radio communication ranges). The bicycle‐specific pneumatic tubes, being portable, were tested at each location where one of the other devices was being tested, to maximize the data produced by each hour of validation counts (i.e., each hour of validation counts could be compared to the counts produced by multiple count devices at a site). EVALUATION CRITERIA  Technologies were primarily evaluated for their accuracy in classifying and tabulating bicycle and pedestrian volumes (based on intended use). Accuracy was evaluated by comparing the count recorded by a given automated counter to ground truth counts produced by manually reducing video data from the site. Chapter 3 provides a comprehensive explanation of how ground truth counts were conducted. In addition to accuracy, counting technologies were also assessed for ease of implementation, labor requirements, security from theft or vandalism, maintenance requirements, software requirements, cost, and flexibility of downloading and working with the count data. Readers are encouraged to review NCHRP Report 797 (Ryus et al. 2014a) for guidance on these practical aspects of using count technologies. OVERALL FINDINGS FROM PHASES 1 AND 2  The analysis conducted for each of the counting technologies was accomplished in three steps in both Phases 1 and 2: (1) graphical (exploratory) analysis, (2) accuracy calculations, and (3) development of correction factors. Table S‐2 summarizes the combined accuracy and precision findings for the technologies tested during Phases 1 and 2. The table includes updated results for devices tested during Phase 1, where appropriate, incorporating both new data from Phase 2 and the errata for Phase 1 (http://onlinepubs.trb.org/onlinepubs/nchrp/nchrp_rpt_797errata.pdf). NCHRP Web‐Only Document 205: Methods and Technologies for Pedestrian and Bicycle Volume Data Collection (Ryus et al. 2014b) describes the technologies and test sites used in the original research, as well as all other aspects of the Phase 1 research. Greater detail regarding the findings for each technology is provided in Chapter 4 of this report. Consistent with NCHRP practice, no product names or manufacturers are identified by name.

NCHRP 07‐19(02) Final Report 6 Table S‐2.  Counting Technology Key Findings: Combined Results from Phases 1 and 2  Technology  Mode  Research  Phase  APD  AAPD  WAPD r N  Average  Hourly  Volume  Passive infrared  C  1, 2  −3.5%  22.5%  −9.5%  0.938  398  258  Product A  C  1 8.7% 22.2% −1.6% 0.949  244  279 Product B  C  1 −26.0% 26.4% −27.0% 0.982  115  263 Product C  C  2 −13.5% 13.5% −13.6% 0.988  39  113 Active infrared  C  1 −6.6% 7.3% −7.6% 0.998  34  327 Thermal imaging camera  B, P  2 5.5% 22.5% 2.7% 0.912  28  101 Bicycle−specific tubes  B  1 −19.8% 22.2% −17.1% 0.979  262  167 Product A  B  1 −9.5% 10.8% −11.1% 0.993  172  203 Product B  B  2 −69.1% 69.1% −59.6% 0.841  47  117 Product C  B  2 −7.3% 16.6% −9.4% 0.920  43  76 Standard tubes  B  2  −15.2%  17.1%  −17.9%  0.936  17  62  Radar  B, P  1  22.7%  27.8%  14.2%  0.918  31  72  Surface inductive loops  B  1  3.8%  10.5%  4.8%  0.959  136  155  Embedded inductive loops  B  1 0.3% 7.6% −3.1% 0.997  29  145 Surface inductive loops   (facility counts)  B  1  139.5%  159.7%  −11.5%  0.971  136  183  Embedded inductive loops    (facility counts)  B  1  −10.8%  49.6%  −35.3%  0.980  66  186  Piezoelectric strips  B  1, 2 −4.0% 4.5% −4.1% 0.995  120  105 Product A  B  1, 2 −3.4% 3.7% −3.4% 0.997  81  112   Product B  B  2 −5.2% 6.1% −5.8% 0.994  39  91 Radio beam  C  1 −9.6% 9.7% ‐11.1% 0.991  56  321 Notes:   Mode: B = bicycle, C = combined bicycle and pedestrian, P = pedestrian  APD = average percentage deviation, AAPD = average of the absolute percent difference, WAPD = weighted average percentage  deviation, r = Pearson’s Correlation Coefficient, N = number of hours evaluated, Average volume = hourly average pedestrian and  bicycle counts based on video observation.  Facility count statistics reflect both errors inherent to the counting device or technology, and bypass errors (i.e., missed detections  due to bicyclists traveling outside the device’s detection area).  1A negative APD indicates undercounting of device.  2AAPD weights overcounting and undercounting as absolute percentages.  3WAPD accounts for the low‐volume bias of the AAPD measure by weighting the AAPD based on the ground truth volume.  4Values of Pearson’s r closer to +1 indicate a stronger positive correlation.  As shown in Table S‐2, four accuracy and precision measures were calculated for each technology:  Average Percentage Deviation (APD) represents the overall divergence from perfect accuracy across all data collected;  Weighted Average Percentage Deviation (WAPD) is similar to APD but is volume‐weighted to remove bias caused by large percentage deviations during low‐volume hours;  Average of the Absolute Percentage Deviation (AAPD) addresses the issue of overcounts and undercounts canceling each other out in the calculation of APD; and  Pearson’s Correlation Coefficient (r) tells how correlated the ground truth volume and the automated count are; a value of +1 indicates that one perfectly predicts the other, although the two counts do not necessarily have to be equal. A value close to +1 suggests that one can reasonably estimate the true volume by multiplying the automated count by a multiplicative adjustment factor.

NCHRP 07‐19(02) Final Report 7 Chapter 3 provides details about the calculation of these accuracy and precision measures. The following subsections summarize accuracy and precision results by counter technology. Passive Infrared (Phases 1 and 2)  Based on the Phase 1 literature review and practitioner survey, passive infrared appears to be the primary sensor technology used at present in the United States for single‐mode and mixed‐mode (pedestrians and bicyclists) environments. Devices using this technology are relatively easy to install; however, care should be given to background conditions that may trigger false detections, such as the presence of windows or other reflective surfaces that can accumulate heat in the sun. Occlusion (i.e., where one person blocks another from the sensor’s view when both pass the counter’s sensor at the same time) was found to occur with higher user volumes, resulting in undercounting. Overall, the testing found a weighted average undercounting rate of 9.5% and a total deviation of 22.5%. Passive infrared sensors from three different vendors (Products A, B, and C) were tested, and a large difference between the accuracy of Products A and C and Product B was found. Product A had a weighted average undercount of 1.6%, Product B had a weighted average undercount of 27.0%, and Product C had a weighted undercount of 13.6%. The undercounting rate includes instances of overcounting that offset the undercounting that occurs. The total deviation indicates the absolute deviation from the actual pedestrian and bicycle volumes counted; therefore, the absolute sum of the under‐ and overcounting amounts to a 22.5% deviation from the actual. Active Infrared (Phase 1)  One active infrared sensor was tested. The Phase 1 research found these sensors are used less commonly than passive infrared sensors, however, the active infrared sensor was found to be fairly accurate with high consistency. It was moderately easy to install, but special attention should be given to align the transmitter and the receiver. In the original research, the device was found to have a weighted undercount rate of 7.6% with a total deviation from actual counts of 7.3%. The undercounting rate includes instances of overcounting that offset the undercounting that occurred. The total deviation indicates the absolute deviation from the actual volumes counted; therefore, for the active infrared sensor, the absolute sum of the under‐ and overcounting amounts to a 7.3% deviation from actual. Radar (Phase 2)  A single radar device was tested in this study, embedded in a bicycle lane. Some inaccuracies were observed that appear to be attributable to motor vehicles encroaching in the bicycle lane, so care should be taken when installing these devices to locate them at sites with minimal encroachment. The device tested was purposely installed in a mixed traffic environment under the advisement of the product vendor. There was an observed weighted average overcounting rate of 14.2%, and a total deviation of 27.8%. However, this overcount rate could be specific to this study site. Further research is needed to produce more generalizable results. As the tested product is designed to be embedded in the pavement, it is more labor‐intensive and intrusive to install than many other counting technologies.

NCHRP 07‐19(02) Final Report 8 Thermal Imaging Camera (Phase 2)  A single thermal imaging camera was tested on a two‐way cycle track. A weighted average overcount rate of 2.7% was observed, with a total deviation of 22.5%. Bicycle‐Specific Pneumatic Tubes (Phases 1 and 2)  Bicycle‐specific pneumatic tubes can be used in mixed traffic to count bicycles only. Bicycle‐specific tubes were installed and tested primarily on multiuse paths and bicycle lanes, except for two sets installed on a shared‐use lane with relatively low motor vehicle traffic and two installations across on‐road bicycle facilities and adjacent general traffic lanes. The tubes are relatively easy to install but ongoing, routine checks of the site are recommended to make sure the tubes have not become dislodged, as Phase 1 and 2 testing indicated that bicycle‐specific tubes dislodged more easily than standard tubes. This consideration is particularly important in mixed traffic settings. The net weighted average accuracy of bicycle‐specific pneumatic tubes showed an undercount by an average of 19.8%. The research team tested pneumatic tubes from three vendors and found differences in the accuracy of Products A, B, and C. Product A had a net weighted undercount of 11.1%, Product B had a net weighted undercount of 59.6%, and Product C had a net weighted undercount of 9.4%. Practitioners are encouraged to evaluate specific products to understand their relative accuracy as differences between products can vary widely for a given technology. The total deviation from the actual counts was found to be 22.2% on average (considering all products). The total deviation for Product A was 10.8%, 69.1% for Product B, and 16.6% for Product C. Standard Pneumatic Tubes (Phase 2)  One set of standard pneumatic tubes (designed to detect both motor vehicle traffic and bicycle traffic) was installed in a shared‐lane configuration, covering both a bicycle lane and the parallel general purpose travel lanes. The sensor was the same as bicycle‐specific Product C (above), except the tubes were replaced with standard tubes after the original bicycle‐specific tubes were dislodged early in the test. This alternate configuration was not recommended by the vendor, but the results were fairly promising. A weighted average bicycle undercount rate of 17.9% was observed, with a total deviation of 17.1%. Radio Beam (Phase 1)  The original research tested radio beam technology at five locations: four multiuse path sites and one wide sidewalk site. Two of the devices installed on multiuse paths distinguished bicyclists from pedestrians, while the other three devices reported aggregate bicycle and pedestrian volumes. There were substantial data errors observed in the output of the devices that output disaggregate totals, with long periods of 0 counts interspersed with unreasonably high volumes. These results, therefore, have not been reported as the counters do not appear to have been functioning correctly. Additionally, the devices being tested required the installer to specify the start time and binning

NCHRP 07‐19(02) Final Report 9 interval, which was done incorrectly at one site, such that the device could not easily be validated in the same effort as the other devices at the site. The remaining two radio beam sensors functioned very well, with an average undercount rate of 11.5% and very high correlation (0.98) with the manual counts. Further testing is needed for the device that performs mode differentiation. One of the devices tested had a relatively narrow maximum facility width that could be counted, which severely constrained the number of potential sites at which it could be tested. Inductive Loops (Phase 1)  The original research evaluated the effectiveness of inductive loops on multiuse paths and on‐street bicycle facilities. Inductive loops for more permanent installations are embedded into the pavement and therefore tend to be more labor‐intensive and intrusive to install than many other counting technologies. Semi‐permanent inductive loops are also available, which are affixed to the pavement surface with a rubberized compound. The advantage of semi‐permanent loops is that the majority of the hardware (aside from the loops themselves) can be reused at future sites, so multiple sites can be sampled for a lower cost compared to installing permanent loops. Semi‐permanent loops are designed for approximately 6 months of installation. Inductive loops have a visible detection area on the pavement. This is different than some of the other sensor technologies, such as passive infrared, that establish an invisible screen line across a facility. Depending on the site characteristics and how the inductive loops are installed, the detection zone for the loops may not extend the entire width of the facility (e.g., an on‐street bicycle lane). This instance can result in bypass errors, where bicyclists ride through the count site but do not pass through the inductive loop detection zone and therefore are not counted. As a result, it is important to develop site plans and confirm that the facility width can be fully covered by the inductive loops. Where installations unavoidably result in uncovered portions of the facility, it may be possible to develop a bypass rate based on manual observation, but this is less desirable. The inductive loop technology was found to have a high degree of accuracy and consistency for counting the bicyclists riding through the detection zone. In these cases, Phase 1 found a weighted average overcounting of 4.8% and an average total deviation from the actual counts of 10.5%. Larger undercounting will occur at sites where the bicycle travelway is wider than the detection zone, as a result of bypass errors. On‐street installations where bicyclists may not always travel in the bicycle lane pose the most challenging context for establishing a sufficiently wide detection zone. Piezoelectric Strips (Phases 1 and 2)  Two sets of piezoelectric strips were tested, both installed along the same multiuse path. One sensor was pre‐existing at the time of the Phase 1 research, while the second sensor was installed as part of the Phase 1 research, but was not able to export data until a software issue was corrected during Phase 2. Both sensors were found to be highly accurate and precise, with a pooled weighted

NCHRP 07‐19(02) Final Report 10 undercount of 4.5% and a total deviation of 4.5%. The separate results for each product were similar, with respective values of (−3.4%, 3.7%) for Product A and (−5.8%, 6.1%) for Product B. Additional testing in on‐street configurations may be warranted, but for multiuse path settings, piezoelectric strips worked well at the single site used in this research. Piezoelectric strips are more labor‐intensive and intrusive to install than many other counting technologies. Combination (Phases 1 and 2)  The research included two combination systems. These systems are able to distinguish pedestrian and bicyclist volumes by taking counts from the passive infrared sensor (which aggregates pedestrians and bicyclists together) and subtracting the count from a second sensor (which only captures bicyclists) to develop the pedestrian count. The second sensor was an inductive loop in the device tested in Phase 1 and a piezoelectric sensor in the device tested in Phase 2. This research focused on evaluating individual sensor technologies, so each component was included with its respective sensor technology category (passive infrared, inductive loops, or piezoelectric strips). The inductive loops and piezoelectric strips were analyzed by comparing against the count of number of bicyclists, while the passive infrared sensors were analyzed by comparing against the sum of the number of pedestrians and the number of bicyclists. However, the combination devices were also evaluated on the accuracy and consistency of their estimates of pedestrian volumes, with an average undercount of 18.65% and total deviation of 43.78%. CONCLUSIONS  Example Adjustment Factors  Various Poisson regression model forms were tested for correcting counted volumes to actual volumes. Tested correction functions included environmental factors (such as temperature), which can be interpreted as modifications on the accuracy rates under the given conditions. In many cases, however, a simple intercept‐only model with an automated count offset (i.e., multiplicative factor) may be the best option for practitioners. Because the intercept (i.e., the fixed adjustment to the count) was close to zero in nearly all cases, it could be neglected, with only the multiplicative factor being used to estimate the true count. More details on the modeling process are provided in Chapter 3 of this report. Table S‐3 presents a sample of model intercepts and adjustment factors derived for each of the sensor technologies tested during the Phase 1 and 2 research. In cases where multiple products representing the same technology were tested, individual anonymized product results are presented along with the overall results for the technology. To apply adjustment factors, one would multiply the automated count data by the appropriate “adjustment factor” from Table S‐3. For example, an hourly count of 100 people from the active infrared sensor tested by the project would correspond to a best estimate of 108 people. These adjustment factors are presented as examples of typical adjustment factors developed for various technologies. Practitioners should develop equipment‐ and site‐specific adjustment factors for their automated count deployments and not

NCHRP 07‐19(02) Final Report 11 rely on these “general” factors for purposes of calibrating count data, as myriad factors may affect the accuracy of any given deployment. Table S‐3.  Example Counter Adjustment Factors: Combined Results from Phases 1 and 2  Sensor Technology  Research  Phase  Intercept  Adjustment  Factor  Hours of   Data  Active infrared*  1  0.079  1.082  34  Thermal imaging camera*  2 −0.026  0.974  28  Passive Infrared A  1  0.016  1.016  244  Passive Infrared B  1  0.314  1.369  115  Passive Infrared C*  2  0.146  1.157  39  Radar*  2 −0.162  0.851  32  Radio beam  1  0.118  1.125  56  Inductive loops A  1 −0.046  0.955  136  Inductive loops B  1  0.032  1.032  29  Piezoelectric strips A*  1, 2  0.035  1.035  81  Piezoelectric strips B*  2  0.060  1.061  39  Pneumatic tubes A (bicycle‐specific tubes)  1  0.117  1.124  172  Pneumatic tubes B (bicycle‐specific tubes)  1  0.543  1.721  47  Pneumatic tubes C (bicycle‐specific tubes)  2  0.078  1.081  43  Standard tubes*  2  0.197  1.217  17  Note:  *Factor is based on a single sensor at one site. As shown in Table S‐3, no tested technology or product worked perfectly; all would require some adjustment to the raw count data to produce a best estimate of the nonmotorized traffic volume. At the same time, it was possible to develop a reasonable adjustment factor for all of the tested devices—even those with relatively high over‐ and undercounting rates. It can also be seen in Table S‐3 that the accuracy of the counted volumes varied substantially between products when multiple products were tested for a given sensor technology. This finding suggests that a specific vendor’s implementation of a technology (e.g., the algorithm used to decide whether a detection should be registered) can be as important as the technology itself in determining accuracy. Given these findings, as well as the knowledge (discussed below) that site‐specific conditions can influence count accuracy, it is recommended that users develop their own local correction factors for their devices whenever possible, particularly when Table S‐3 indicates substantial over‐ or undercounting associated with a particular technology or substantial differences in accuracy between products. Factors Influencing Accuracy  Counter accuracy varied notably depending on site‐specific characteristics. Site‐specific factors influencing the accuracy of the counts included proper calibration and installation of the devices, and taking care to avoid situations that can result in over‐ or undercounts. For example, passive

NCHRP 07‐19(02) Final Report 12 infrared sensors are susceptible to false positives when windows, mirrors, or other reflective surfaces are positioned behind the pathway being counted. Counters can also be subject to bypass errors, where a pedestrian or bicyclist is able to go around the counter’s detection zone and avoid being counted. Thoughtful selection of counting locations that minimize opportunities for avoiding the counter can help minimize bypass errors. More detailed installation guidance is provided in Chapter 5 of NCHRP Report 797 (Ryus et al. 2014a). For screenline‐based technologies (e.g., radio beam, passive infrared), occlusion is an unavoidable and predictable factor contributing to systematic undercounting. The degree to which occlusion contributes to undercounting is a factor of pedestrian and bicycle grouping (i.e., groups of persons traveling side‐by‐side). All of the tested counting technologies that are subject to occlusion effects showed a linear relationship between counted and actual volume that could be corrected using a simple multiplicative factor. Factors Not Found to Influence Accuracy  Several factors anticipated to affect the accuracy of counting technologies were not evident in the testing. For example, concern has been expressed that the age of inductive loops influences their accuracy. However, the inductive loops tested by this project included loops that were 2 and 3½ years old, and diminished quality was not detected in those loops’ counting accuracy. Similar concerns have been expressed related to the age of pneumatic tubes. However, over the six‐month duration of the Phase 1 testing, count quality was not observed to decline over time. Proper maintenance practices are important to ensure consistent reliable performance and it is recommended that the user follow vendor‐specified practices for maintaining and replacing equipment. For example, pneumatic tubes may degrade after being installed in motor vehicle traffic for prolonged periods. The testing parameters did not allow for exploration of extreme heat or cold temperatures (over 100 °F or under 20 °F). No clear impact or effect of temperature on the accuracy of the technologies was found. However, for the temperature ranges captured in the research, no impact on the accuracy of the tested devices was observed. Similarly, there was limited observation of rain and snow events during testing that did not allow for any discernable effect on count accuracy, but those events that did occur did not appear to influence the quality of the data. Given the lack of conclusive evidence, it is recommended that the user follow vendor‐specified practices for deployment and installations under varying weather conditions or climatic regions. Recommendations for Practitioners  It is recommended that practitioners calibrate and conduct their own ground truth count tests for the automated technologies they deploy at a given site or set of sites. This project’s research results are intended to provide information to practitioners on the types of technologies that may be most promising for a specific circumstance, use, or location where automated count technology is being considered. More details on factors such as cost, maintenance, and other installation considerations are not covered in detail in this report, but are covered extensively in Chapter 5 of NCHRP Report 797 (Ryus et al. 2014a).

NCHRP 07‐19(02) Final Report 13 The project’s accuracy findings should not be blindly applied to other sites than those at which these technologies were tested at, and it should not be assumed that the same degree of accuracy will occur at other site locations or with other products. Practitioners can use the research approach described in this report and NCHRP Report 797 to, on a smaller scale, test and evaluate the performance of their own automated count technologies. Particularly when using a product for the first time, it is important for practitioners to work with the vendor to thoroughly understand the product’s installation, operation, and maintenance requirements, including those of any required accessories (e.g., software, communications equipment). As with any product, obtaining other users’ experiences with specific products and specific vendors’ customer support can be highly useful when making decisions on which product(s) to purchase.