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 31 Chapter 3: Combined Findings from Phases 1 and 2 A major component of the research involved field testing a variety of commercially available pedestrian and bicycle counting technologies by comparing the counts produced by the technologies with manual counts taken from video footage. The manual counts were assumed to represent correct, or “ground truth,” counts. Counting technologies were then evaluated for accuracy (average error rate across all time periods) and consistency (degree to which similar accuracy rates are repeated for different time periods) based on comparison to the ground truth data. The term precision is also used to describe counting consistency. All of the manual counts were observed from videos taken at each test site. Videos were typically recorded for two periods of up to a week at each of the sites in the study, in an effort to test the technologies under a diverse set of environmental conditions, within the limitations of the project budget. For example, a set of test sites were selected in Davis, CA in Phase 1 because it has hot weather and high bicycle volumes, and a set of test sites were selected in Minneapolis because it has high volumes and cold temperatures. Digital copies of the videos were shipped from the test sites on DVDs and flash drives to the data reduction team in Berkeley. DATA ANALYSIS  Data analysis involved three phases: graphical (exploratory) analysis, accuracy calculations, and correction functions. Graphical Analysis  The first phase of the data analysis process involved plotting manual (ground truth) versus automated counts for each technology. For example, the initial plots depict the manual count values on the x‐axis versus the automated count values on the y‐axis (at 1‐hour resolution, which has been used for all of the analysis). Figure 3‐1 provides an example. The “overcounting” region in this figure shows hours where the automated count exceeded the ground truth. The “undercounting” region shows hours when the automated count was less than the ground truth. The diagonal dashed “perfect accuracy” line indicates hours when the automated and ground truth counts matched each other. Note that when a data point falls on this line, undercounting and overcounting could be occurring that cancel each other out (e.g., four missed detections and four false positives), resulting in an automated count that matches the ground truth count.

Figure 3‐ The graph example, adjusted intended multiuse tubes ove Figure 3‐ 1.   Exam ical analysi the pneuma after the firs for a mixed path site. Pr rcounted, bu 2.   Befor Adjus ple of a Gra s also shows tic tubes on t two days o traffic situat ior to the adj t overall cou e and After  tment  phical Displ patterns in the Midtown f data collect ion, which re ustment, the nt accuracy Comparison 32 ay of Count  the data in te Greenway i ion during P quires diffe tubes unde improved. F  of Pneuma N Data  rms of accu n Minneapo hase 1. Thes rent sensitiv rcounted. Fo igure 3‐2 de tic Tube Ac CHRP 07‐19 racy and con lis had their e tubes wer ity settings t llowing the picts this ef curacy with (02) Final R sistency. As sensitivity e initially han for a adjustment, fect.  Sensitivity  eport an the

NCHRP 07‐19(02) Final Report 33 Accuracy Calculations  Four accuracy and precision measures were calculated for each technology. Average percentage deviation (APD) and weighted average percentage deviation (WAPD) are both measures of accuracy, while average of the absolute percentage deviation (AAPD) and Pearson’s Correlation Coefficient (r) are measures of precision. These measures are described in more detail below. All analyses were carried out using the Python programming language. Average Percentage Deviation (APD) Average Percentage Deviation (APD) represents the overall divergence from perfect accuracy across all data collected. This is calculated as: ܣܲܦ ൌ 1݊෍ ܣ௧ െ ܯ௧ ܯ௧ ௡ ௧ୀଵ where At is the automated count for time period t, Mt is the manual (ground truth) count in period t, and n is the total number of periods analyzed. This metric has the advantage of providing insight into what correction factors can be used for a given technology (as discussed in greater detail below), but does not provide as much detail on overall accuracy. In particular, overcounts and undercounts in different time periods can cancel each other out. Weighted Average Percentage Deviation (WAPD) To account for the low‐volume bias of the APD measure, a volume‐weighted accuracy measure is also calculated, as: ܹܣܲܦ ൌ ෍ ቆܣ௧ െܯ௧∑ ܯ௝௡௝ୀଵ ቇ ௡ ௧ୀଵ WAPD is considered to be more reliable than APD, as it is not sensitive to deviations in low‐volume hours. For example, consider two hours with ground truth volumes of 1 and 100 people. If there is a single false positive and it occurs during the first time period, the APD would be 50%, and the WAPD would be 0.99%. If the same false positive occurred in the second time period, the APD would be 0.5% and the WAPD would still be 0.99%. Thus, APD is highly sensitive to stochastic variation during low‐volume periods, and is primarily included here to allow comparisons with other studies. Average of the Absolute Percentage Deviation (AAPD) AAPD helps to remedy the undercount/overcount cancellation problem with the APD. ܣܣܲܦ ൌ 1݊෍ ฬ ܣ௧ െ ܯ௧ ܯ௧ ฬ ௡ ௧ୀଵ

NCHRP 07‐19(02) Final Report 34 By taking the absolute values, over and undercounts of the same magnitude no longer balance each other out, but rather both count toward the total accuracy. However, this measure has the difficulty that percentage errors at low volumes can bias the results, as for example an overcount of 1 on a ground truth volume of 1 is calculated as a 100% overcount, whereas an overcount of 1 on a ground truth volume of 100 is calculated as a 1% overcount. Pearson’s Correlation Coefficient (r) Pearson’s r tells how correlated two variables are with each other, where r = +1 is total positive correlation, r = −1 is total negative correlation, and r = 0 is no correlation. With automated counters, the value of r will ideally be +1 between the ground truth volume and the automated count. That is, one perfectly predicts the other, although the two counts do not necessarily have to be equal. A correlation coefficient close to r = +1 suggests that one can reasonably estimate the volume by multiplying the automated count by a multiplicative adjustment factor. Pearson’s coefficient is calculated as: ݎ ൌ ∑ ሺܯ௧ െܯഥሻሺܣ௧ െ ̅ܣሻ ௡௧ୀଵ ඥ∑ ሺܯ௧ െ ܯഥሻଶ௡௧ୀଵ ට∑ ሺܣ௧ െ ̅ܣሻଶ௡௧ୀଵ Correction Functions  A number of accuracy correction functions were estimated for each technology. As the value being modeled is an integer variable, correction functions are modeled using Poisson regression estimated using maximum‐likelihood estimation. A log‐link function is considered, and the automated count is used as an “exposure” variable. This is expressed mathematically as: ܯ௜ ~ Poissonሺߤ௜ሻ lnሺߤ௜ሻ ൌ lnሺܣ௜ሻ ൅ ߙ௜ ൅ ࢼࢄ࢏ where  Mi is the manual (i.e. ground truth) count for observation hour i  ߤ௜ is the (conditional) Poisson rate for the observation hour i  Ai is the automated count for the observation hour i  ߙ௜ is the intercept, an estimated parameter corresponding to the baseline miscount rate  ࢼ is a vector of additional estimated parameters corresponding to the effects of various factors on counter accuracy  ࢄ࢏ is a vector of values for additional factors, such as precipitation or lighting conditions, thought to potentially effect the accuracy of the counter The effect of modeling the accuracy in this way is that, by exponentiation, the Poisson rate can be expressed as:

NCHRP 07‐19(02) Final Report 35 ߤ௜ ൌ ܣ௜݁ఈ೔ାࢼࢄ࢏ ൌ ܣ௜݁ఈ೔Π௞݁ఉೖ௑ೖ೔ That is, the rate is assumed to vary linearly with the automated count. If only an intercept (ߙ௜ሻ is included in the model, the resulting value of ݁ఈ೔ is a simple multiplicative factor by which the automated count should be adjusted. In this formulation, any additional factors included in Xi will also have a multiplicative effect on the accuracy. For example, in studying the passive infrared technology, one of the estimated models includes an intercept and a fixed effect for the specific product with which the observation was collected. The intercept point estimate is 0.0388, with a value of 0.276 for the Product B term. The resulting correction function can thus be interpreted as ܯ௜ ൌ ܣ௜݁଴.଴ଷ଼଼ ൌ 1.040 ൈ ܣ௜ for Product A (the base), and ܯ௜ ൌ ܣ௜݁଴.଴ଷ଼଼݁଴.ଶ଻଺ ൌ 1.040 ൈ 1.318 ൈ ܣ௜ ൌ 1.37 ൈ ܣ௜ for Product B. Similar interpretations can be made for other effects. For each dataset, models were compared on the basis of the Akaike Information Criterion (AIC). The AIC is a measure of model fit which penalizes models based on the number of estimated parameters. For a given dataset, models with lower AIC values can be interpreted as fitting the data better. However, the AIC does not provide any information on absolute goodness‐of‐fit (as measures such as R2 do). R2 is not used in this evaluation because it is not reliable for non‐Gaussian models. Models were also evaluated based on the significance (compared against 0 using a t‐test) of individual parameter estimates. ANALYSIS BY TECHNOLOGY TYPE  Radar  Qualitative Experience For this study, a single radar unit was tested in a bicycle lane at the Oakland site. The bicycle lane is adjacent to a parking lane on the right and to two parallel vehicular travel lanes on the left, as shown previously in Figure 2‐6. In addition, the site is located on a slight curve in the roadway. An analysis of the data indicates the tested device overcounts at this site. Motorized vehicles were observed encroaching the bicycle lane (“cutting the corner”); parking activity will also be registered as a bicycle detection. As the counter is buried in the pavement, its detection zone was estimated from the device’s technical specifications, which required some subjectivity on the part of the data collector to determine whether a given bicyclist traversed the detection zone or skirted its edge. It could not be determined with certainty whether the sensor’s detection area extended beyond the bicycle lane (in which case motorized vehicles at the edge of their lane could be detected) or fell short of the edge of the bicycle lane (in which case a bicycle riding along the edge of the lane might not be detected). If a bicyclist riding along the lane line (a frequent occurrence, likely due to the well‐utilized on‐street parking at the site) was detected by the sensor, but the data collector misidentified this event as being outside the detection zone, it would result in an apparent

overcoun eliminate Based on motor ve understa character the detec passes th either mo where th In additio repeater the repea Accuracy In Figure count of b truth to b encroach apparent One outli automate would be removing (a) Bicy Figure 3‐ t, when in fa these error this experie hicles are lik nding of the istics of the tion zone. M rough the de stly segrega ere is not mu n to conside and receiver ter can be lo and Consis 3‐3, two acc icyclists pas e a count of ments into t accuracy, su er was obser d count data an unreason this outlier clists Passing  3.  Accur ct the error s, but there i nce, care sho ely to freque technology, reflected rad isidentificati tection zone ted facilities ch lane chan ration of the component cated from b tency uracy plots a sing through bicyclists pa he bicycle la ggesting tha ved at coord . This data p ably high nu dramatically Through Dete acy Plots fo was in the m s some possi uld be taken ntly encroac it classifies v ar pulse, wh ons appear t , so this tech (e.g., cycle t ging behavi sensor loca s of the syste oth the sens re shown. T the detecti ssing throug ne. The inclu t the sensor inates (1, 76 oint is from mber for th improves th ction Zone r Radar  36 anual valida bility of thei when insta h the senso ehicles as bi ich is govern o occur mos nology is lik racks, off‐str or, so that m tion, site sel m can be loc or and the r he plot on th on zone. The h the detect sion of these may in fact b ) that would 6:00 a.m. on is time perio e results. (b) Bicy N tion. All pos r occurrence lling radar c r’s detection cycles or car ed in part b t frequently ely to work eet paths), o otor vehicle ection also n ated, as the eceiver. e left consid plot on the ion zone, plu car encroac e detecting likely be fla a rainy Sun d at this loca clists Plus Car CHRP 07‐19 sible efforts . ounters to av zone. From s based on t y the cross‐s when a sma best for coun r in shared l s are identif eeds to cons re are limits ers the grou right consid s a count of hments imp these events gged in a qu day, when 76 tion. As sho  Encroachmen (02) Final R were made t oid sites wh the authors’ he ectional ma ll section of ting bicycle anes at loca ied correctly ider where t on how far a nd truth to b ers the grou car roves the . ality check o bicyclists wn in Table ts in Detection eport o ere ss in a car s in tions . he way e a nd f the 3‐1,  Zone 

NCHRP 07‐19(02) Final Report 37 Table 3‐1.  Accuracy and Consistency Values for Radar Sensor  Dataset  APD  AAPD  WAPD  r  N  Average  Hourly  Volume  Bicyclists  256%  261%  18%  0.89  32  70  Bicyclists  (no outlier)  23%  28%  14%  0.92  31  72  Bicyclists +  perceived car  encroachments  30%  62%  −12%  0.94  32  93  Notes:   APD = average percentage deviation, AAPD = average of the absolute percent difference, WAPD = weighted average percentage  deviation, r = Pearson’s Correlation Coefficient.  Effects of Environmental Conditions The radar sensor tested in this study was installed in Oakland, California and data were collected at times of year when there is relatively little variation in environmental conditions. Due to the limited amount of data collected in this study for this device, it is not possible to identify any effects of environmental conditions. The only observation of note is the one outlier hour discussed above, which was collected during heavy rain; however, this is a single observation, and therefore does not provide any significant statistical evidence of performance under these conditions. Other observations collected in the rain by this device were highly accurate, but there were only three rainy hours total in the dataset. Correction Functions Using the Poisson formulation above yields a multiplicative correction factor of 0.851 (95% CI: 0.816, 0.887). However, it is important to note that this result is based on just one site at which the particular mechanics of interaction with the adjacent traffic lane appear to have a substantial impact on the results (resulting in more overcounting than would generally be expected). Site‐ specific factors should be estimated for any new installation, and additional research is needed to generate more general adjustment factors for this technology. Thermal Imaging Camera  Qualitative Experience One thermal imaging device was tested, at one location. This device required hardwiring into the electrical grid, which may limit its applicability to locations with access to electrical power. While this technology is designed to collect a variety of data types, including multimodal counts, this project only tested its ability to count bicyclists in a dedicated bicycle facility – specifically, a two‐ way cycle track segment with minimal intrusion by motor vehicles. This technology offers the ability to selectively identify events of interest to monitor. For this application, two overlapping bounding boxes were defined within the camera’s field of view, with specified directionality. A count was recorded when a bicyclist passed through a bounding box in

the direct the data c than the b bounding sources o differenti Accuracy The therm The obse r=0.912 w Figure 3‐ Regressi Based on with the c there app multiplic develope AIC, and h ion specified ollector to m icycle or pe boxes that w f error. Futu ate bicyclist and Consis al imaging rved accurac ith a sampl 4.  Accur on Correctio visual inspe aveat that th ears to be li ative factor. d; as shown ad significa for that box itigate bypa destrian faci ere defined re studies co s in mixed tr tency counter has y and precis e size of N=2 acy Plot for  ns ction, it appe is result is b mited utility However, m in Table 3‐2 nt p‐values. . This featur ss errors by lity alone. H only extend uld vary thi affic or to di fairly high ac ion metrics 8. Thermal Im ars that the ased on a si to developin odels that in ,these show 38 e allows for defining the owever, due ed to the ed s configurati fferentiate b curacy with were APD=5 aging Senso thermal infr ngle counter g correction cluded rain a ed marginal N directional bounding b to the limite ge of the cyc on to test th icyclists from high consist .5%, AAPD= r  ared camera at a single s functions m nd nighttim improvemen CHRP 07‐19 counting, an ox to cover a d scope of th le track, to a e device’s ab pedestrian ency, as sho 22.5%, WAP ’s accuracy ite. In light o ore comple e (as binary ts to overal (02) Final R d may also a wider exten is testing, th void conflat ility to s. wn in Figure D=2.7%, and function is li f this findin x than a simp variables) w l fit, based on eport llow t e ing 3‐4. near, g, le ere the

NCHRP 07‐19(02) Final Report 39 Table 3‐2.  Correction Functions for Active Infrared Sensor  Intercept  Night  Rain  AIC  ‐0.0263 (0.163)  ‐‐‐  ‐‐‐  400  ‐0.0398 (0.045)  ‐‐‐  0.146 (0.021)  397  ‐0.00567 (0.786)  ‐0.105 (0.0295)  ‐‐‐  397  Notes:   AIC = Akaike Information Criterion. Numbers in parentheses are P values; coefficients with a P value of 0.050 or less are significant at  a 95% confidence level.  Night is an indicator variable that is 1 when the starting time for the counting period is between sunset and sunrise, and 0 otherwise.  Effects of Different Conditions Because only one thermal imaging sensor was tested in the research, there was not enough variation in the data to ascertain whether any site‐level factors had an impact on the accuracy. The following factors were hypothesized to have an effect: False Positives in Heavy Precipitation The modeling suggests that controlling for rain can yield a slight improvement in overall correction function fit. However, the estimated parameter values counterintuitively suggest that rainy hours are better expressed as undercounting than overcounting. This result suggests the need for further testing in inclement weather, which might best be accomplished with a disaggregate study design. Temperature The hypothesized effect of temperature on the operation of this technology is a “masking” of objects when the ambient temperature approaches the object temperatures, or approximately 98oF. The maximum observed temperature for this technology during this study was 83oF, so this effect could not be evaluated. Passive Infrared  Qualitative Experience Passive infrared sensors made up a significant proportion of the data in the study. According to the practitioner survey conducted during Phase 1 (Ryus et al. 2014b), these sensors are in wide use for collecting pedestrian volume data in single‐mode environments (e.g., sidewalks) and for collecting combined bicycle and pedestrian volume data in mixed‐mode environments. As documented by the Phase 1 literature review (Ryus et al. 2014b), multiple studies have evaluated the accuracy of passive infrared sensors. The degree of difficulty associated with installing passive infrared sensors varied by product but was generally fairly simple. Temporary installations involved either bracketing a box with the sensor inside to an existing fixed object, such as a street sign pole, or screw‐mounting a small sensor into a wooden surface. Permanent passive infrared sensor installations typically involve

NCHRP 07‐19(02) Final Report 40 sinking a wooden post into the ground adjacent to the facility being counted, as is common when used as part of a combination sensor unit. Due care should be taken with the site design and installation for passive infrared sensors to avoid directing the sensor toward backgrounds that are likely to trigger false detections. Examples of problematic backgrounds include heavy foliage, windows, or background traffic or other movements. One example of an installation‐related issue from the testing experience involved the sensor being directed toward a planter box below a window, resulting in reflective interference that inflated counts. This problem was exacerbated with high ambient air temperatures that likely heated the foliage to a temperature approaching that of a human body. Another difficulty with passive infrared sensors is undercounting due to occlusion. These counters were observed to perform worse at higher volumes, especially Product B. Accuracy and Consistency The findings on passive infrared sensors corroborate the findings of previous studies. The passive infrared sensors demonstrated an overall average undercount rate (APD) of 10.0% and an AAPD of 17.0%. Figure 3‐4 provides a combined graph of the accuracy of all three tested passive infrared products, as well as accuracy graphs for the individual products. A less accurate count rate was observed with one product compared to the other two, as demonstrated in Figure 3‐4 and Table 3‐2. All three products appear to follow a roughly linear profile, but one seems to have a lower slope (i.e., farther from 45°), suggesting that errors for this product propagate with volumes at a higher rate than the other products. Note that Product C, which is the new addition from Phase 2, was only tested at relatively low traffic volumes at one site (as part of a combination counter on a multiuse path), so direct comparison with the other two products is somewhat limited.

Figure 3‐ The fact t technolog installatio However counts ov obtained correctio needed to correctio Device‐Sp Evaluatin these tech calculate 5.  Accur hat the prod y plays an im n are impor , as can be in er a variety from any of n factor need adjust the o n factors. ecific Accur g accuracy a nologies, bu separate fac (a)  acy Plots of ucts had diff portant rol tant factors ferred from of volumes, the tested pr ed to adjust ther produc acy and Con nd consisten t given the v tors for each Combined Acc (b) Accuracy   Passive Infr erent error e in the coun to investigat Figure 3‐5, a which mean oducts by si one produc ts’ counts, it sistency cy in aggreg aried result device prod 41 uracy of All Th of Individual P ared Senso rates indicat ter’s accura e prior to pu ll three prod s that reason mply applyi t’s counts is is recomme ate is useful s for this spe uct and dev N ree Products roducts  rs  es that a ven cy and thus rchasing an ucts produc ably accurat ng a correcti substantially nded that us for evaluati cific techno ice tested. Ta CHRP 07‐19 dor’s implem product sele d deploying ed relatively e volume es on factor. Be different th ers develop ng the overa logy, it is imp ble 3‐3 show (02) Final R entation of ction and this technolo consistent timates coul cause the an the facto their own lo ll performan ortant to s the result eport this gy. d be rs cal ce of s.

NCHRP 07‐19(02) Final Report 42 Table 3‐3.   Accuracy and Consistency Metrics on a Site‐ and Device‐Specific Basis for  Passive Infrared Sensors  Site  Product  APD  AAPD  WAPD  r  N  Average Hourly  Volume  Overall  Average  ‐  −10%  17% −11%  0.96  392  261  Overall  Product A  A −1%  12% −4%  0.98  238  284  L St  A −2%  10% −3%  0.94  39  537  Four Mile  Run  A  2%  8%  3%  0.90  42  171  Fell Street  A −5%  10% −3%  0.97  26  43  15th  Avenue  A −5%  5% −4%  0.99  16  367  Key Bridge  A −12%  13% −13%  0.99  48  393  Berkeley  A −14%  14%  14%  0.96  6  151  Sycamore  A  14%  15%  12%  0.99  24  160  Loyola  A −18%  19% −18%  0.86  12  64  Overall  Product B  B −26%  26% −27%  0.98  115  263  Midtown  Greenway  B −26%  26% −27%  0.97  24  436  Four Mile  Run  B −21%  22% −21%  0.91  42  171  Key Bridge  B −29%  29% −30%  0.98  31  351  Sycamore  B −33%  33% −31%  0.85  18  96  Overall  Product C  C −14%  14% −14%  0.99  39  113  Four Mile  Run  C −14%  14% −14%  0.99  39  113  Notes:  APD = average percentage deviation, AAPD = average of the absolute percent difference, WAPD = weighted average percentage  deviation, r = Pearson’s Correlation Coefficient, N = number of detectors, Average volume = hourly average pedestrian and bicycle  counts based on video observation.  Correction Functions The various correction function forms that were tested for passive infrared counters are presented in Table 3‐4. Including a product fixed effect yields significantly improved results, further confirming that the particular implementation of the technology can affect results. Controlling for the product, significant effects on accuracy are found due to lighting conditions (“night”), rain, and hot weather. The positive signs on all of these parameters suggest that these conditions lead to

NCHRP 07‐19(02) Final Report 43 greater degrees of undercounting, as the adjustment factor is positive. Similarly, Products B and C appear to have lower overall accuracy than Product A, again as suggested by the positive signs. Table 3‐4.  Regression Correction Functions Tested for Passive Infrared Sensors  Intercept  Night  Rain  Temperature Cold  Hot  Product  B  Product  C  AIC  0.117  (0.000)  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  6265  −0.0391  (0.055)  ‐‐‐  ‐‐‐  0.00219  (7.13e−15)  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  6205  0.0388  (0.000)  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  0.276  (0.000)  0.107  (0.000)  4737  0.0375  (0.000)  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  0.0972  (0.000)  0.276  (0.000)  0.0926  (0.000)  4723  0.0372  (0.000)  ‐‐‐  ‐‐‐  ‐‐‐  0.00828  (0.66)  0.0973  (0.000)  0.276  (0.000)  0.0929  (0.000)  4725  0.0346  (0.000)  ‐‐‐  0.03  (4.46e−3)  ‐‐‐  ‐‐‐  ‐‐‐  0.279  (0.000)  0.111  (0.000)  4731  0.033  (0.000)  ‐‐‐  0.0317  (2.69e−3)  ‐‐‐  ‐‐‐  0.0999  (0.000)  0.28  (0.000)  0.096  (0.000)  4716  0.0363  (0.000)  0.0397  (4.09e−3)  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  0.277  (0.000)  0.108  (0.000)  4731  0.0298  (0.000)  0.0443  (0.0014)  0.0338  (0.0014)  ‐‐‐  ‐‐‐  0.103  (0.000)  0.282  (0.000)  0.097  (0.000)  4708  Notes:   AIC = Akaike Information Criterion. Numbers in parentheses are P values; coefficients with a P value of 0.050 or less are significant at  a 95% confidence level.  Hot, cold, rain, and dark are indicator variables that are 1 when the condition is met and 0 otherwise. Night is defined as the starting  time for the counting period being between sunset and sunrise. “Cold” represents temperatures below 30 °F and “hot” represents  temperatures over 90 °F.  Environmental Condition Effects For passive infrared counters, the following effects of weather and other environmental conditions were hypothesized, with the conclusions enumerated below: Worse performance at higher volumes, due to a higher incidence of occlusion This effect appears to be more of a problem with Product B than Products A and C. Product A actually has negative coefficients in the automated count, which suggests that its performance is slightly better at high volumes. The magnitude of this term is small, however. Product B, on the other hand, demonstrates a stronger adverse effect of high volumes. This is a problem that has been documented for passive infrared counters in previous literature (Schneider et al. 2012, Ozbay et al. 2010), but it appears that it is a surmountable problem, given Product A’s high accuracy even with high volumes.

Worse pe distinguis There is m 3‐6, there than in th temperat problem most of th difficult t On the ot at very hi very cold attributed current r enough c recomme winter te Figure 3‐ Worse pe The data Figure 3‐ rformance at hing people f inimal evid are limited e overall da ures above 9 is that high t e data in th o conclusive her hand, th gh volumes. temperatur to heavily i esearch beca lothing to ha nded that pr mperatures 6.  Passiv rformance in collected by 7. temperatur rom the bac ence of this data in the “ taset. In this 0 F, and “M emperatures e high tempe ly discern an e testing fou This is in co es and high v nsulating clo use the tem ve an effect— actitioners u monitor dev e Infrared A heavy rain a NCHRP Proj es approachi kground effect sugges hot” temper plot, “Cold” id” refers to have a depr rature regim y difference nd no effect nflict with re olumes, wh thing. It is s peratures w temperatu sing passive ice performa ccuracy as  nd/or snow ect 07‐19 we 44 ng that of a ted in Table ature catego refers to tem anywhere in essing effec e did not ha s in accuracy of freezing te cent researc ere many mi uspected tha itnessed wer res did not d infrared se nce during c a Function o due to false p re sparse fo N human body, 3‐4. Howev ry, with a m peratures b between. O t on nonmot ve very high . mperatures h (Andersen ssed detecti t this effect e not low en rop much be nsors in loca oldest cond f Temperat ositives r heavy rain CHRP 07‐19 due to diffic er, as demon uch smaller v elow 30 F, “ ne difficulty orized volum volumes. It on detectio et al. 2014) ons occurred was not obs ough to war low 10 F. I tions with e itions to ens ure  and snow, a (02) Final R ulties strated in F olume rang Hot” refers t in assessing es, such tha is therefore n accuracy, e conducted , which was erved in the rant heavy‐ t is xtreme cold ure accuracy s indicated i eport igure e o this t ven at . n

Figure 3‐ However observed detection However undercou The testin whether operate b Active I Qualitati The testin device ap with prev The activ transmitt 7.  Accur Range , during peri . On the surf s on passive , in the regre nting, as inc g was not a snow affects y detecting b nfrared  ve Experien g only inclu peared to fu ious experie e infrared se er and recei acy Compar   ods of rain, n ace, this wou infrared sen ssion analys luding the ex ble to captur device perfo ody heat. ce ded one acti nction fairly nce with the nsor is mod ver have to b ison for Pas o substantia ld appear to sors, but fur is, it appears istence of ra e significant rmance. It s ve infrared s accurately, w technology erately easy e installed s 45 sive Infrare l overcount refute the p ther testing that rain m in as a facto snow event eems unlike ensor, a loan ith very hig (Lindsey et to install (no eparately an N d Sensors b s or other pe roposition t may be need ight actually r does impr s, to allow a ly, given tha from the U h consisten al., 2012). ground cut d aligned w CHRP 07‐19 y Rain and T rformance i hat rain trig ed to confir lead to a gr ove model fi determinati t passive infr niversity of M cy. This expe ting required ith each othe (02) Final R emperatur ssues were gers false m this findin eater degree t. on about ared sensor innesota. T rience fits w ), although r. eport e  g. of s his ell the

Accuracy The activ 3‐8. In pa WAPD = ‐ hours of d Figure 3‐ Regressi Based on again wit finding, th simple m were test fit was w Table 3‐5 In 0.079 0.0 0.079 0.09 Notes:   AI a 9 Ni and Consis e infrared se rticular, volu 7.62%, r = 0 ata. 8.  Accur on Correctio visual inspe h the caveat ere appears ultiplicative ed, both sep orse as indic . Corre tercept  3 (6.21e−17)  816 (0.17)  5 (1.75e−16)  06 (0.242)  C = Akaike Inform 5% confidence le ght is an indicator tency nsor has fair me estimat .9979) with acy Plot for  ns ction, it appe that this is o to be limite factor. Mode arately and a ated using A ction Functi −0.005 −0.01 ation Criterion. Nu vel.   variable that is 1  ly high accu es were foun a gradually i Active Infra ars that the nly based on d utility to d ls including s a linear su IC, as seen in ons for Acti Night  ‐‐‐  ‐‐‐  96 (0.909)  21 (0.857)  mbers in parenth when the starting 46 racy with ve d to be very ncreasing un red Sensor active infrar a single cou eveloping co temperature m, but no sig Table 3‐5. ve Infrared  eses are P values  time for the coun N ry high cons precise (AP dercount. T ed accuracy nter at a sin rrection fun and night t nificant val Sensor  Temper ‐‐‐ −3.11e−05 ‐‐‐ −0.000142 ; coefficients with ting period is bet CHRP 07‐19 istency, as s D = ‐6.61%, A hese values function is e gle site. In li ctions more ime (as a bin ues were fou ature     (0.968)     (0.885)   a P value of 0.050 ween sunset and  (02) Final R hown in Figu APD = 7.33 are based on xtremely lin ght of this complex tha ary variable nd and over  or less are signif sunrise, and 0 oth eport re %, 34 ear, n a ) all AIC  266  268  268  270  icant at  erwise. 

NCHRP 07‐19(02) Final Report 47 Effects of Different Conditions Because only one active infrared sensor was tested in the research, there was not enough variation in the data to ascertain whether any site‐level factors have an impact on the accuracy. Further, the inclement‐weather data for this counter were quite sparse. The following factors were hypothesized to have an effect: Occlusion effects that increase with increasing volumes Occlusion does not appear to be a factor with increased volumes, given that undercount rates do not increase with volume, despite the observed site including relatively high volumes. False positives in heavy precipitation No time periods with heavy rain or snow were captured on camera for the site with the active infrared sensor, so no conclusions can be drawn on this topic from this study. While previous work has suggested that these events may trigger false positives, this effect has not been documented (Bu et al., 2007). Temperature Temperature does not appear to affect accuracy for the active infrared sensor. Pneumatic Tubes  Qualitative Experience Pneumatic tubes were tested primarily on multiuse paths or bicycle lanes in this study. Bicycle‐ specific pneumatic tubes, thinner and smaller than vehicle counting tubes, were primarily used, although a set of standard tubes (traditional motor vehicle tubes) was used at one site (Oakland) after the bicycle‐specific tubes were dislodged early in the study. The researchers experienced issues with the smaller bicycle‐specific tubes being dislodged at multiple sites during Phases 1 and 2, with various combinations of products and installation teams. The Oakland bicycle‐specific tubes were therefore replaced with standard tubes to reduce the chance of the tubes coming loose a second time. The tests of the third pneumatic tubes product (Product C) at on‐street locations (Oakland and 15th Street NW in Washington, D.C.) included installing the pneumatic tubes across both the bicycle facility (lane/cycle track) and the adjacent general purpose travel lanes in the same direction to test the ability of bicycle‐specific tube counters to distinguish bicycles from motor vehicle traffic. During Phase 1, only one mixed traffic site produced usable data for pneumatic tubes, and the bicycle volumes on that facility were higher than the motor vehicle volumes. The two on‐street facilities used in Phase 2, on the other hand, had substantially higher motor vehicle volumes than bicycle volumes, which is expected to be a more common situation for observing bicycles in mixed traffic. One site (15th Avenue in Minneapolis) proved problematic in multiple ways. First, during the initial data collection phase, the tube fasteners failed, causing the tubes to become free. After the tubes were reinstalled, they did not appear to function well, as shown in Figure 3‐9. These data are from

two sets o tubes we longer pr inaccurat occasiona However counting. overall da Figure 3‐ Accuracy Figure 3‐ Figure 3‐ C with bic and site, a f pneumatic re damaged operly funct e counts is t lly occluded , this does no The data fro ta plot in Fi 9.  Accur and Consis 10 plots accu 11 plots accu ycle‐specifi nd Table 3‐ tubes, one i during the p ional when t hat 15th Ave the camera t seem like t m these sen gure 3‐10) a acy Plot for  tency racy results racy results c and standa 5 provides a nstalled in e eriod when t hey were re nue has fair ’s view of the he sole expl sors have be s the cause o Pneumatic  for all tubes by the three rd tubes atta ccuracy and 48 ach bicycle l hey became installed. An ly high bus a counter for anation, give en omitted f f the errors Tubes at 15 that were t products te ched). Figur consistency N ane. One pos free to the e other possib nd truck tra 1–2 minute n the severi rom all follo is uncertain. th Avenue S ested (minus sted (includ e 3‐12 plots statistics for CHRP 07‐19 sible explan xtent that th le explanati ffic, and thes s in a 15‐min ty of both un wing analys ite  the 15th Av ing a compa accuracy re all tubes. (02) Final R ation is that ey were no on for the e large vehic ute count p der‐ and ov is (and the enue site). rison of Prod sults by prod eport the les eriod. er‐ uct uct

Figure 3‐ The pneu counter t work. It c accuracy Universit were test tubes we sometime tubes or n crossed c 10.  Accur matic tube d echnologies. an be seen in line are Fell y in Montrea ed, and it als re installed w s difficult fo ot because lose to the e acy Plot for  ata present This result s Figure 3‐12 Street in San l (Product B o has the hig ithin the bi r data collec the tubes we nd of the tub Pneumatic  a number of uggests that that the sit Francisco a ). Rue Milton hest observ cycle lane bu tors to deter re at the edg es. 49 Tubes   distinct patt there are st es with the h nd Rue Milto is a mixed t ed bicycle vo t not into th mine wheth e of the field N erns, in cont rong site‐ an ighest diver n in Montre raffic site on lumes. On F e shared‐use er the bicycl of view, an CHRP 07‐19 rast with th d device‐sp sion from th al (Product which pneu ell Street, th lane. For th ist rode ove d many bicy (02) Final R e other teste ecific effects e perfect A), and Rue matic tubes e pneumatic is site, it wa r the pneum clists’ traject eport d at s atic ories

Figure 3‐11.  Accuracy Plot for Pneumatic  50 Tubes by Pr N oduct  CHRP 07‐19(02) Final Report

Figure 3‐12.  Accuracy Plot for Pneumatic  51 Tubes by Pr N oduct and S CHRP 07‐19 ite  (02) Final Report

NCHRP 07‐19(02) Final Report 52 Table 3‐6.  Accuracy and Consistency Values for Pneumatic Tubes by Product and Site  Site  Product  APD  AAPD  WAPD  r  N  Average  Volume  Overall Average ‐  −14.0%  16.6%  −14.1%  0.986  259  166  Bicycle‐specific  tubes −14.0%  16.6%  −14.0%  0.986  242  173  Overall Product A  A  −9.5%  10.8%  −11.1%  0.993  172  203  University  A  0.5%  1.2%  0.6%  0.998  17  206  Key Bridge  A  −11.1%  11.3%  −12.4%  0.981  48  97  L Street  A  −6.7%  10.8%  −10.4%  0.991  45  62  Fell Street  A  −20.2%  20.2%  −23.1%  0.987  23  248  Clarendon  A  −21.1%  21.1%  −22.9%  0.999  3  23  Milton  A  −7.7%  7.7%  −9.2%  0.997  36  503  Overall Product B  B  −52.6%  52.6%  39.9%  0.970  28  99  University  B  −34.3%  34.3%  −34.8%  0.805  11  199  Eastbank  B  −64.4%  64.4%  59.4%  0.781  17   34  Overall Product C  C  −7.3%  16.6%  −9.4%  0.920  43  76  Four Mile  C  8.9%  9.2%  7.6%  0.996  22  75  15th Street NW  C  −24.2%  24.2%  −26.6%  0.976  20  81  Oakland  C  −28.6%  28.6%  −28.6% 1  14  Standard Tubes  C  −15.2%  17.1%  −17.9%  0.936  17  62  Oakland  C  −15.2%  17.1%  −17.9%  0.936  17  62  Notes:  APD = average percentage deviation, AAPD = average of the absolute percent difference, WAPD = weighted average percentage  deviation,  r = Pearson’s Correlation Coefficient, N = number of detectors, Average volume = hourly average pedestrian and bicycle  counts based on video observation.  Effects of Environmental Conditions Frozen tubes The research team hypothesized that pneumatic tubes will become less sensitive in very cold temperatures, as the rubber in the tubes hardens. However, very little data were available in cold temperatures (below 30 F), so these conditions were not extensively tested. It bears mentioning that in many cases, these conditions should not be experienced anyway, as pneumatic tubes are generally not intended to be used during winter conditions. Reduced accuracy due to aging tubes The rubber components of a pneumatic tube sensor system are a consumable that must be periodically replaced as they wear out. Tubes can develop cracks, holes, and weak spots that result in miscounts. Cases of total failures are fairly obvious to detect—no counts are produced. However, it was suspected that tubes might lose some accuracy as they age. To test this, one set of bicycle‐

NCHRP 07‐19(02) Final Report 53 specific tubes was left installed on the Midtown Greenway for the duration of the study (~5 months). Accuracy rates were not substantially worse for this set of tubes than for others in the study. However, tubes installed in mixed traffic likely experience more rapid degradation based on the size and frequency of vehicles traveling across the tubes. Reduced accuracy in mixed traffic Two field tests of Product A were conducted in Montreal at the Rue Milton site, as well as field tests of Product C on Grand Avenue in Oakland and on 15th Street NW in Washington, D.C. At the Rue Milton site, tubes were deployed across a bicycle lane in one direction and across a shared lane in the opposite direction. At the Grand Avenue and Washington, D.C. test sites, the tubes were installed across a bicycle facility the two adjacent motor vehicle travel lanes to simulate the effects of a shared travel lane while allowing multiple counters to be tested at the same location. In the regression table below (Table 3‐7), it can be seen that including an indicator variable for a counter being installed on a shared lane does not improve model fit, suggesting that there is no significant effect on accuracy. Regression Corrections Table 3‐7 provides the parameters associated with the various regression models that were tested. Significant differences were found between Product A (the base), Product B, and Product C using standard tubes (but not Product C using bicycle‐specific tubes). Mixed traffic lanes, as discussed above, do not appear to have an effect. Temperature (when included as a linear term) does appear to have a significant effect on accuracy when controlling for differences between products, although this finding is somewhat equivocal as the mechanism behind the effect is not clear. However, a substantially better in‐sample fit (measured by AIC) is seen for this model.  Table 3‐7.  Correction Functions for Pneumatic Tubes   Intercept  Shared Lane  Temperature  Tubes B  Tubes C  Tubes C  Standard  AIC  0.152  (0.000)  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  ‐‐‐  4485 0.168  (0.000)  ‐‐‐  −0.000233  (0.693)  ‐‐‐  ‐‐‐  ‐‐‐  4486 0.117  (0.000)  ‐‐‐  ‐‐‐  0.484  (0.000)  −0.0189  (0.301)  0.0795  (0.0113)  3782 0.117  (0.000)  ‐‐‐  ‐‐‐  0.484  (0.000)  0.00406  (0.801)  ‐‐‐  3788 0.121  (0.000)  −0.00669  (0.515)  ‐‐‐  0.481  (0.000)  −0.0197  (0.281)  0.0826  (9.22e−3)  3784 0.507  (0.000)  ‐‐‐  −0.00566  (0.000)  0.529  (0.000)  −0.0792  (0.000)  0.052  (0.0991)  3711 Notes:   AIC= Akaike Information Criterion. Numbers in parentheses are P values; coefficients with a P value of 0.050 or less are significant at  a 95% confidence level.  Shared lane is an indicator variable that is 1 when both motorized vehicles and bicycles cross the tubes and 0 otherwise. 

NCHRP 07‐19(02) Final Report 54 Radio Beam  Qualitative Experience The research team faced significant difficulties evaluating the accuracy of one of the radio beam sensors. The product that distinguished pedestrians from bicyclists (Product A) had to be mounted on both sides of the facility, with a maximum separation of 10 feet, which severely limited the choice of sites where the counter could be tested. The product specifications available online, used in selecting test sites, had indicated that the maximum separation was 12 feet, but the material that arrived with the product specified only 10 feet, which required the team to locate new test sites for the product, which ended up being narrow sections of multiuse paths. The narrower paths were also associated with lower pedestrian and bicycle volumes, which reduced the volume range over which the device could be tested. Product A also defaulted to beginning its count immediately when initiated, rather than aggregating into bins beginning on the hour, creating difficulty synchronizing the data collection intervals with the ground truth video. This was a setting that could be altered, but required going into an “advanced settings” menu, which most of the installers did not realize. This meant that the automated counts collected corresponded to different time periods than the manual counts, as all other devices counted in 15‐minute or 1‐hour periods that began on the hour. However, this setup issue is a minimal problem in terms of collecting volume data in general, and in fact this device had far more flexibility than the others tested (e.g., time bins of any integer number of minutes, delayed count starts). Accordingly, counts had to be repeated for a number of radio beam counters, for example counting from 12:05–12:19, 12:20–12:34, 12:35–12:49, and 12:50–13:04. One counter used in this study was inadvertently set to count in 61‐minute intervals, so manual counts for this site were redone using intervals corresponding to the same time periods. Finally, substantial data errors were encountered with the installations of Product A, and therefore results for Product A have been removed from the analysis. It is suspected that these data errors were due to the devices being installed on bridges containing large metal components, which was not a condition identified in the product instructions, but which the research team learned retrospectively may affect this technology. Accuracy and Consistency Figure 3‐13 shows the accuracy plot for radio beam Product B, which provides a total of all passers‐ by. As shown in Table 3‐8, the device had high accuracy and precision over a large range of traffic volumes, and the results were similar across all sites in the study. The results for Product B presented in the Phase 1 report were based on raw count data that had not been adjusted for daylight savings time; this problem has been corrected in the results presented below.

Figure 3‐   Table 3‐8 Dataset  Overall  Average  Four Mil Midtown 5th Aven Notes:   AP de Effects of The prim as is the c lighting, a variation radio bea between 13.  Accur .  Accur Pro e Run    ue  D = average perc viation, r = Pearso Environme ary hypothe ase with all nd rain wer in the weath m sensors in modes. acy Plot for  acy and Con duct  AP ‐  −10 B  −6. B  −11 B  −12 entage deviation,  n’s Correlation Co ntal Conditi sized source screenline se e not suspec er to evalua adverse we Radio Beam sistency Va D  AA .0%  10 0%  6 .0%  11 .0%  12 AAPD = average o efficient .  ons of error for nsors. Radi ted to be pro te any poten ather condit 55  Product B lues for Rad PD  W .0%  − .0%  − .0%  − .0%  − f the absolute per radio beam o beams are blems. Duri tial effect of ions, and for N io Beam Pr APD  11.0%  6.0%  12.0%  13.0%  cent difference, W sensors prio not optical d ng testing, th rain. Furthe radio beam CHRP 07‐19 oduct B  r  0.99  1.00  0.98  0.98  APD = weighted  r to this stud evices, so te ere was ins r research is counters th (02) Final R N  Ave Ho Vo 56  3 20  1 24  4 12  3 average percenta y was occlu mperature, ufficient warranted at distinguis eport rage  urly  lume  21  56  36  66  ge  sion, for h

Inductiv Qualitati Inductive Both perm When eva addresse always co which is l and insta eliminate First, con 14. The lo is located located in maneuve the loop s Fell Stree also be co Figure 3‐ Second, c Minneapo zones. Tw surface lo between testing w errors for e Loops  ve Experien loops were anent and t luating indu d. Bypass err ver the enti ikely a resul llation and h d. sider the em op is install immediatel the bicycle rs between b ensor. Simil t and are con nsidered a b 14.  Poten onsider the lis. This site o sets of ind ops. Both se their edges a here the stri eastbound ce tested at a n emporary in ctive loops, ors arise be re width of t t of traffic pa elp minimiz bedded loop ed on a green y after a bicy lane. High bi icyclists, wh arly, cyclists tinuing stra ypass error tial Bypass  embedded a is a very wi uctive loops ts of inducti nd the sides pe was shift bicyclists us umber of sit ductive loop a special typ cause of the he facility. Cy tterns and m e this error, installation ‐painted bic cle route tur cycle volum ich can resu who have m ight) someti . Error Source nd surface lo de multiuse were instal ve loops are of the path, ed 1 to 2 feet ing the edge 56 es during th s were teste e of error, by sensor’s spa clists somet icro‐design but in most o on Fell Stre ycle lane on ns left onto es on this fac lt in bypass ade a right t mes ride on s on Fell St op detectors path, with m led at this sit centered in a result of th into the ped line as a gui N is project, bo d. pass error, tially limited imes ride ar elements of n‐street cas et in San Fra the left side Fell Street. T ility result i errors if the urn onto Fel the right sid reet  installed on arked separ e: a set of pe the bicycle fa e restriping estrian zon de; the bypa CHRP 07‐19 th on‐ and o needs to be detection z ound the ed the facility. es it cannot ncisco, show of a one‐wa he inductiv n frequent p passing man l Street (or w e of the stre Midtown G ated pedestr rmanent loo cility, but h of the edgel e. This gap r sses seemed (02) Final R ff‐street fac specifically one: loops d ge of the loo Good site de be fully n in in Figur y street. Thi e loops are assing euvers occu ere already et, which can reenway in ian and bicy ps and a set ave small ga ine just prio esulted in by to occur mo eport ilities. o not ps, sign e 3‐ s site r at on cle of ps r to pass re

frequentl estimatin difficult t Finally, co Arlington The path path. Hen Figure 3‐ Accuracy Figure 3‐ volume a plots rep Midtown of nearly lower vol volume a slightly a y at high vol g the accura o determine nsider the e , Virginia, sh has a constr ce, this site 15.  Induc and Consis 16 compares nd the overa resent two s Greenway d 800 bicyclis umes, the fa ppears to inc ffected. umes. Note t cy of the loo the exact ed mbedded lo own in Figu ained width does not exp tive Loop Te tency the differen ll facility vol lightly differ ata collectio ts per hour), cility counts rease with v hat this effe ps when bicy ge of the det op detector re 3‐15. The (8 to 10 feet erience subs sting on Ke ces in accur ume (includ ent data sets n process. Th which likely are underco olume, whe 57 ct was also p clists were ection zone on the Key B facility is a m ), and the lo tantial bypa y Bridge  acy and cons ing bypass e , due to the a is is a very h biases the a unted. As ex reas detectio N roblematic f riding direct in the camer ridge betwe ultiuse path ops cover ro ss errors. istency base rrors). It is i forementio igh‐volume ccuracy dow pected, und n zone accu CHRP 07‐19 or data colle ly over them a footage. en Washingt on the side ughly the en d on the det mportant to ned difficult site (peak v nwards. Ho ercounting o racy appear (02) Final R ction in term , in that it w on, D.C. and of the bridg tire width of ection zone note that th ies with the olumes obse wever, even f the facility s to be only v eport s of as e. the ese rved at ery

Figure 3‐ Table 3‐9 (i.e., cons table, ind substanti Table 3‐9 Notes:  AP de Dataset  All data  All surfa Loyola  Universit All embe Sycamor Clarendo Key  Fell  (a) Detect 16.  Accur provides the idering only uctive loop s al difference . Accur D = average perc viation, r = Pearso Ty − ce  S S y  S dded  E e  E n  E E E ion Zone Only  acy Plots fo accuracy an those bicycl ensors are b between th acy and Con entage deviation,  n’s Correlation Co pe  AP   1.5   0.3   7.9   −5.0   1.9   6.6   −27.4   7.5   −9.6 r Inductive L d consistenc es that passe oth very acc e surface ind sistency Va AAPD = average o efficient.  D  AA %  8 %  7 %  10 %  5 %  8 %  7 %  27 %  7 %  9 58 oops  y statistics f d through a urate and ve uctive loops lues for Ind f the absolute per PD  W .3%  .6%  .8%  .3%  .5%  .5%  .4%  − .7%  .6%  N (b) By or inductive loop’s detec ry precise. T and the emb uctive Loop cent difference, W APD  −0.6%  −3.1%  6.2%  ‐4.7%  0.1%  5.4%  29.7%  6.2%  −9.5%  CHRP 07‐19 pass Errors In loops, witho tion zone). A here does n edded indu s (Detection APD = weighted  r  0.990  0.997  0.974  0.988  0.989  0.997  1.000  0.994  0.998  (02) Final R cluded  ut bypass err s shown in ot appear to ctive loops.  Zone Accu average percenta N  A H V 134  29  12  17  105  24  3  48  26  eport ors the be a racy)  ge  verage  ourly  olume  136  145  51  211  133  141  25  97  193 

NCHRP 07‐19(02) Final Report 59 Table 3‐10 presents accuracy and consistency statistics for inductive loops when considering facility volumes (i.e., including bypasses) as the “ground truth.” The facility volume was defined on a site‐by‐site basis. For multiuse paths, all bicyclists were counted while riding on the path. For bicycle lanes, the volume included all bicyclists riding along the street. Accordingly, the facility‐level accuracy and consistency values presented here should not be taken as general truths, but rather as documentation of the range of values encountered for the sites on which inductive loops were installed in this study. While these are important performance considerations, they will vary substantially by site and therefore need to be evaluated on a case‐by‐case basis when using devices that do not cover the entire width of a facility. Overall consistency rates are fairly high (Pearson’s r > 0.95 for all sites), while accuracy rates varied substantially between sites. The multiuse path sites and cycletrack site had high accuracy rates, particularly when the inductive loops spanned the entire path width. The Midtown Greenway’s inductive loops were both narrower than the path, with the surface loops being narrower than the embedded loops, and hence these loops undercounted the total facility volume. Similarly, the on‐ street sites were subject to bypass errors from bicyclists riding outside of the bike lane. Table 3‐10.  Accuracy and Consistency Values for Inductive Loops (Facility‐Level Accuracy)  Notes:  E = embedded, S = surface.  APD = average percentage deviation, AAPD = average of the absolute percent difference, WAPD = weighted average percentage  deviation, r = Pearson’s Correlation Coefficient, N = number of detectors, Average volume = hourly average pedestrian and bicycle  counts based on video observation.    Inductive loop bypass error can be mitigated in a number of ways, including selecting and designing sites with widths compatible with the loop detector parameters and avoiding locations that are unconstrained, allowing bicycle traffic to travel outside the intended travelway. However, these measures cannot necessarily be applied to on‐street facilities where travel paths cannot be easily constrained and bicyclists may choose and be permitted to travel outside the marked bicycle lane. Dataset  Type  APD  AAPD  WAPD  r  N  Average  Hourly  Volume  All data  −  90.4%  123.8%  −19.4%  0.950  202  184  All surface  S  −10.8%  49.6%  −35.3%  0.980  66  186  Loyola  S  −1.6%  8.2%  −3.3%  0.977  12  56  L Street  S  25.6%  98.6%  −31.6%  0.500  20  79  Midtown  S  ‐35.4%  35.4%  −38.1%  0.991  34  295  All embedded  E  139.5%  159.7%  −11.5%  0.971  136  183  Sycamore  E  6.6%  7.5%  5.4%  0.997  24  141  Clarendon  E  −46.6%  46.6%  −48.0%  0.996  3  33  Key  E  7.5%  7.7%  6.2%  0.994  48  97  Midtown  E  −17.9%  17.9%  −19.4%  0.996  31  321  Fell  E  −25.6%  25.6%  −26.2%  0.991  26  236 

NCHRP 07‐19(02) Final Report 60 For these installations, site‐specific correction factors should be developed to account for bypass errors and to improve the estimation of facility volumes. Effects of Environmental Conditions The age of inductive loops has been suggested as a potential issue for loop accuracy. However, this was not detected in the testing. One set of embedded loops that was tested was 3½ years old and another was 2 years old. However, as shown in Figure 3‐15, which includes data from all of the counters, no data seem to be of especially bad quality despite these older systems being included in the study. Correction Functions As the graphical exploration and accuracy metrics discussed above suggest, inductive loops work very well when considering the accuracy through the “detection zone.” This is additionally borne out by estimating correction functions, which when only an intercept was concluded showed both a very low parameter estimate (0.0065) and a 95% confidence interval on either side of 0, suggesting that no correction is needed. Correction functions are not separately considered for the facility counts, because the rates of this are largely a function of site‐specific conditions, and therefore should be evaluated on a case‐by‐ case basis. General findings on the degree of bypass are an area for potential further study. Piezoelectric Strips  Qualitative Experience Two piezoelectric strip sensors were tested in this study, both located at the Four Mile Run site. One was a pre‐existing installation. The other was installed during Phase 1, but experienced technical problems that were fixed by the product’s new vendor in time for Phase 2. Accuracy and Consistency Figure 3‐17 presents accuracy plots for the two tested devices. The results for Product A that were published in the Phase 1 reports failed to correct for daylight savings time; this error has been corrected in the results presented below, and the data for the product have been supplemented with additional data collection during Phase 2.

Figure 3‐ Both of th consisten are some are prom Table 3‐1 Product  Overall  Product  Product  Notes:  AP de co Effects of While the precipita condition Regressi As noted are neede tested bu 17.  Accur e piezoelect cy over the r what limited ising for this 1. Accur AP −4.0 A  −3.4 B  −5.2 D = average perc viation, r = Pearso unts based on vid Environme re was limit tion on accur s. on Correctio above, the o d, and parti t was not fou acy Plot for  ric strip sen ange of volu , to the exten technology. acy and Con D  AA 0%  4.5 0%  3.7 0%  6. entage deviation,  n’s Correlation Co eo observation.  ntal Conditi ed adverse w acy. Further ns verall high a cularly not fo nd to be sig Piezoelectr sors tested i mes observ t that this t sistency Va PD  0%  0%  10%  AAPD = average o efficient, N = num ons eather obse research sh ccuracy of th r environm nificant. The 61 ic Strips  n this study ed, as indicat est was only lues for Pie WAPD  −4.10%  −3.40%  −5.80%  f the absolute per ber of detectors, rved, there ould be cons is technolog ental conditi pooled mod N showed very ed in Table conducted o zoelectric St r  0.995  0.997  0.994  cent difference, W  Average volume did not appe idered on th y suggests t ons. A fixed el with just CHRP 07‐19 high accura 3‐11. While n a single m rips  N  Av 120  81  39  APD = weighted  = hourly average  ar to be any e effects of hat no subst effect for the an intercept (02) Final R cy and these finding ultiuse path erage Hourl Volume  105  112  91  average percenta pedestrian and bi effect of very cold antial correc product wa is estimated eport s , they y  ge  cycle  tions s to

NCHRP 07‐19(02) Final Report 62 have an intercept value of 0.0415 (95% CI: 0.024, 0.059), suggesting a multiplicative adjustment factor of 1.04 (1.03, 1.06). Combination Counter (Pedestrian Estimates)  Qualitative Experience Combination counters use multiple technologies to generate separate estimates of pedestrian and bicycle volumes. In this study, two combination counter technologies were tested. Both technologies used passive infrared detection to get aggregate counts, but one used inductive loops to get separate bicycle counts and the other used piezoelectric strips. The tested devices output estimates of pedestrian and bicycle volumes by default. These sensors have been evaluated separately in the previous sections, considering the sum of the automated and pedestrian count as the passive infrared sensor’s ground truth and considering the bicycle count as the inductive loop’s ground truth volume. Qualitative experiences have already been discussed, but here the accuracy and consistency of the counters’ inferred pedestrian volumes are discussed explicitly. It should be noted that the error in the pedestrian count depends in part on the errors in the bicycle and total counts (as the pedestrian volume is estimated by subtracting the bicycle count from the total count), but the focus here is on evaluating the need to correct the count data produced by a combination counter. Accuracy Description Figure 3‐18 shows the pedestrian volume estimate accuracy based on data for both sites. Figure 3‐ 19 distinguishes the data between the two products, while Table 3‐12 gives calculated accuracy and consistency metrics. These devices appear to work well on the whole, with a high consistency rate (Pearson’s r = 0.992). The Sycamore Park detector and the piezoelectric strips‐based detector both had fairly poor accuracy and precision metrics, but these are largely attributable to the relatively low pedestrian volumes observed at these sites. As previously discussed, a small number of detection errors at a low‐volume site results in a large percentage deviation.

Figure 3‐ Figure 3‐ The insta due to oc 18.   Accur 19.  Accur Count llation at the clusion effec acy Plot for  acy and Con ers Compar Key Bridge ts) with a hi Pedestrian  sistency Plo ing Two Pro site, on the o gh rate of co 63 Volumes Es ts for Pede ducts  ther hand, s nsistency, as N timated fro strian Volum hows net un indicated in CHRP 07‐19 m Combina es from Co dercounting Table 3‐12. (02) Final R tion Counte mbination  (as expecte eport rs  d

NCHRP 07‐19(02) Final Report 64 Table 3‐12.  Accuracy and Consistency Metrics for Pedestrian Volumes from Combination  Counters  Regression Corrections Table 3‐13 shows the correction functions developed for pedestrian volume estimates using combination counters. The analysis was restricted to considering an intercept‐only model with a binary factor representing the product type. It appears that considering the two products separately produces substantially better results. Caution is advised with these results, however, due to the limited range of conditions under which observations were taken, particularly the limited volume range for the piezoelectric‐based sensor. Table 3‐13.  Correction Functions Estimated for Pedestrian Volumes from Combination  Counters  Intercept  Piezo Sensor  AIC  0.199 (0.000)  ‐‐‐  1412  0.179 (0.000)  0.424 (0.000)  1283  Dataset  APD  AAPD  WAPD  r  N  Average  Hourly  Volume  All data  8.8%  54.8%  −18.0%  0.992  111  140  Combination passive  infrared/inductive loops  36.6%  61.4%  −16.4%  0.991  72  204  Sycamore  145.7%  148.3%  60.9%  0.751  24  19  Key Bridge  −17.9%  17.9%  −18.9%  0.982  48  296  Combination passive  infrared/piezoelectric  −42.7%  42.7%  −45.3%  0.648  39  22 

NCHRP 07‐19(02) Final Report 65 SUMMARY  Table 3‐14 provides a combined comparison of accuracy and consistency values by site, product, and counting technology. Table 3‐14.  Accuracy and Consistency Values for all Technologies by Site and Product  Technology  Subset  APD  AAPD  WAPD  r  N  Average  Hourly  Volume  Passive Infrared  All data  −3.5%  22.5%  −9.5%  0.938  398  258  Product A  8.7%  22.2%  −1.6%  0.949  244  279  Product B  −26.0%  26.4%  −27.0%  0.982  115  263  Product C  −13.5%  13.5%  −13.6%  0.988  39  113  Four Mile (A)  2.1%  8.0%  2.9%  0.897  42  171  Four Mile (B)  −21.1%  22.1%  −20.9%  0.913  42  171  Four Mile (C)  −13.5%  13.5%  −13.6%  0.988  39  113  Key (A)  −12.3%  12.6%  −12.7%  0.988  48  393  Key (B)  −28.6%  28.6%  −30.1%  0.982  31  351  L Street  53.1%  63.2%  4.5%  0.865  45  474  Loyola  −18.5%  19.5%  −18.1%  0.863  12  64  Sycamore (A)  14.1%  14.7%  12.0%  0.993  24  160  Sycamore (B)  −32.8%  32.8%  −30.8%  0.852  18  96  15th Ave  −4.9%  5.4%  −4.1%  0.995  16  367  5th Ave  10.0%  22.4%  1.4%  0.872  31  294  Fell  −4.6%  9.6%  −3.0%  0.968  26  43  Active Infrared  All data  −6.6%  7.3%  −7.6%  0.998  34  327  Thermal Camera  All data  5.5%  22.5%  2.7%  0.912  28  101  Radar  All data  23.0%  28.0%  14.0%  0.920  31  72  Pneumatic  Tubes   All data  −15.3%  17.7%  −14.8%  0.986  279  160  Product A  −9.5%  10.8%  −11.1%  0.993  172  203  Product B  −45.9%  45.9%  −41.9%  0.889  47  117  Product C   (bike‐specific tubes)  −5.4%  14.8%  −7.5%  0.938  43  76  Product C  (standard tubes)  −15.2%  17.1%  −17.9%  0.936  17  62  Clarendon  −21.1%  21.1%  −22.9%  0.999  3  23  Key  −11.1%  11.3%  −12.4%  0.981  48  97  L Street  −6.7%  10.8%  −10.4%  0.991  45  62  Midtown 1†  4.3%  33.3%  30.0%  0.901  30  298  Midtown 2†  39.3%  39.3%  46.9%  0.982  9  12  University (A)  0.5%  1.2%  0.6%  0.998  17  206  University (B)  −34.3%  34.3%  −34.8%  0.805  11  199  Eastbank  −49.4%  49.4%  −46.6%  0.789  36  92  Fell  −20.2%  20.2%  −23.1%  0.987  23  248  Milton 1  −6.6%  8.9%  9.8%  0.979  18  497  Milton 2  −8.2%  8.2%  9.9%  0.996  18  497    15th St NW  −19.9%  20.3%  −22.7%  0.981  20  81    Oakland  −15.2%  17.1%  −17.9%  0.936  17  62    Four Mile  8.9%  9.2%  7.6%  0.996  22  75 

NCHRP 07‐19(02) Final Report 66 Technology  Subset  APD  AAPD  WAPD  r  N  Average  Hourly  Volume  Inductive Loops  All data  0.6%  8.9%  7.6%  0.994  108  128  Surface loops  0.3%  7.6%  5.7%  0.997  29  145  Embedded loops  0.6%  9.4%  8.4%  0.993  79  122  Sycamore  7.9%  8.2%  7.7%  0.973  18  81  Loyola  7.9%  10.8%  9.8%  0.974  12  51  Clarendon −27.4%  27.4%  29.7%  1.000  3  25  Key  8.0%  8.1%  6.0%  0.996  31  92  University −5.0%  5.3%  5.0%  0.988  17  211  Fell −9.5%  9.5%  9.5%  0.998  27  194  All data* −14.1%  17.6%  23.6%  0.965  165  200  Surface loops* −20.1%  21.6%  29.3%  0.942  59  222  Embedded loops* −10.7%  15.4%  19.9%  0.990  106  187  Sycamore*  7.9%  8.2%  7.7%  0.973  18  81  Loyola* −1.6%  8.2%  8.4%  0.977  12  56  Clarendon* −27.3%  27.3%  29.7%  1.000  3  25  Key*  8.0%  8.1%  6.0%  0.996  31  92  Midtown (surface)* −37.5%  37.5%  41.3%  0.994  30  298  Midtown (embedded)*  −20.6%  20.6%  23.0%  0.998  27  330  University* −2.4%  2.8%  2.8%  0.992  17  206  Fell* −25.5%  25.5%  26.1%  0.990  27  237  Piezoelectric  All data −4.0%  4.5%  −4.1%  0.995  120  105  Product A −3.4%  3.7%  −3.4%  0.997  81  112  Product B −5.2%  6.1%  −5.8%  0.994  39  91  Radio Beam  All data (Product B) −9.6%  9.7%  −11.1%  0.991  56  321  Midtown −11.1%  11.4%  −11.5%  −0.982  24  436  5th Ave −12.2%  12.2%  −13.5%  0.977  12  366  Four Mile Run −6.2%  6.2%  −6.4%  0.999  20  156  Combination  (ped)  All data  8.8%  54.8%  −18.0%  0.992  111  140  Inductive loop–based  36.6%  61.4%  −16.4%  0.991  72  204  Sycamore  145.7%  148.3%  60.9%  0.751  24  19  Key Bridge  −17.9%  17.9%  −18.9%  0.982  48  296  Piezoelectric‐based  −42.7%  42.7%  −45.3%  0.648  39  22  Notes:   *Denotes values calculated using facility counts (i.e., including bypass errors).  †Denotes value not included in overall accuracy calculations due to identified sensor problems.  (A), (B), (1), and (2) represent different products implementing a given sensor technology.  APD = average percentage deviation, AAPD = average of the absolute percent difference, WAPD = weighted average percentage  deviation, r = Pearson’s Correlation Coefficient, N = number of detectors, Average volume = hourly average pedestrian and bicycle  counts based on video observation.  Table 3‐15 provides sample adjustment factors for each technology, based on the experience of the field testing. The factors are simple multiplicative factors—that is, a multiplier that is 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 an adjustment factor of 1.20, the estimate of the true count would be 120. Although a number of different types of models for correcting counts were tested by the research, multiplicative adjustment factors provided the best combination of prediction accuracy and simplicity of application. The implication is that count errors increased at a linear rate

NCHRP 07‐19(02) Final Report 67 for the technologies tested. Where multiple products representing a given technology were tested, Table 3‐15 also provides product‐specific, anonymized results. Table 3‐15.  Counter Sample Adjustment Factors Developed by NCHRP Project 07‐19  Sensor Technology  Intercept  Adjustment Factor  Hours of Data  Active infrared*  0.079  1.082  34  Thermal imaging camera* −0.026  0.974  28  Passive infrared  0.100  1.106  398   Product A  0.016  1.016  244   Product B  0.314  1.369  115   Product C  0.146  1.157  39  Radar −0.162  0.851  31  Radio beam (Product B)  0.118  1.125  56  Induction loops −0.034  0.967  165   Product A −0.046  0.955  136      Product B  0.032  1.032  29  Piezoelectric strips  0.042  1.042  120   Product A  0.035  1.035  81   Product B  0.060  1.061  39  Pneumatic tubes  0.160  1.173  279   Product A (bike‐specific tubes)  0.117  1.124  172   Product B (bike‐specific tubes)  0.543  1.721  47   Product C (bike‐specific tubes)  0.078  1.081  43   Standard tubes  0.197  1.217  17  Note:  *Factor is based on a single sensor at one site; use caution when applying.