Freeway Detector Assessment- Aggregate Data from the Remote Traffic Microwave Sensor (RTMS)

Transcription

1 Freeway Detector Assessment- Aggregate Data from the Remote Traffic Microwave Sensor () Benjamin Coifman, PhD Assistant Professor, Civil and Environmental Engineering and Geodetic Science Assistant Professor, Electrical and Computer Engineering Ohio State University Hitchcock Hall Neil Ave Columbus, OH

2 ABSTRACT Loop detectors have been the preeminent detection technology for several decades, but they require closing the right of way during installation and potentially undermine the integrity of the pavement surface if they are not installed prior to paving. As a result there is great interest in emerging technologies that promise traffic detection without the liabilities of loop detectors, many of which have already been deployed in large numbers. The Remote Traffic Microwave Sensor () is among the widest deployed non-invasive traffic detector. This study evaluates the performance of the in side-fire mode relative to loop detectors in freeway applications. First by comparing the aggregated data reported by the using its internal controller emulation in comparison to data from nearby dual loop detectors. It is shown that the measures are noisier than loop detectors for occupancy (percentage time the detector is occupied by vehicles) and flow (number of vehicles per unit time), though the velocity estimates are almost as good as those from single loop detectors (while being inferior to direct measurement from dual loop detectors). Secondly, the study considers aggregate measurements from contact closure data, comparing the against the dual loop detectors. For reference, the work also compared one loop against another in a dual loop detector, with the spacing between loops being greater than the spacing between the reference loops and the detection zone.

3 Coifman INTRODUCTION Loop detectors have been the preeminent detection technology for several decades, but they require closing the right of way during installation and potentially undermine the integrity of the pavement surface if they are not installed prior to paving. As a result there is great interest in emerging technologies that promise traffic detection without the liabilities of loop detectors, many of which have already been deployed in large numbers. The Remote Traffic Microwave Sensor () manufactured by Electronic Integrated Systems (EIS) [1] is among the widest deployed non-invasive traffic detector. But performance varies from one detector technology to the next and it is important that a given detector performs as expected. To this end, there have been numerous studies comparing emerging detector technologies against loop detectors or manual validation, e.g., [2-5]. The present study continues in this approach, examining the in side-fire mode, with particular care to synchronize the data across the different detectors. Using Station 7 in the Berkeley Highway Laboratory (BHL) on I-8, north of Oakland, CA [6], (Figure 1) the study collected contact closure data from the two detector systems, recording the state at 6 Hz, using a model 17 controller running software developed by Caltrans and previously deployed in [6-7]. For this study these stations were equipped with Peek sensors that had been meticulously calibrated using the tools presented in [8-9]. The research also collected the data recorded by the using its internal controller emulation, this mode allows the to operate independently, providing aggregate traffic data without the need of a conventional traffic controller (the exact details of the controller emulation are proprietary). The unit was mounted on a CCTV pole in late 1999 in accordance with the manufacturer's specifications and the was hardwired to the controller input file. It is worth noting that the calibration software was difficult to control for novice users; however, it is believed that this deficiency could be overcome with training. To ensure optimal performance, representatives of EIS aligned and calibrated the unit. The EIS representative in California conducted the initial calibration on October 21, The EIS representative strongly recommended having a trained professional calibrate the units, stating that he believes he could calibrate approximately four devices per day. While at the site, he commented on two sitespecific features that were likely to reduce the unit's performance. First, the mounting angle of the unit was such that it would reduce performance in the closest lane (lane 5). Second, the site does not have any shoulders in the median, which he expected would reduce the performance on the inside lane (lane 1) in each direction. This degradation is due to echoes off of the concrete barrier on the near side of the median and due to the microwave "shadow" of the barrier on the far side. The EIS representative said that the shadow would impact performance in the first two lanes on the far side of the roadway and upon his suggestion, it was decided to limit the to monitoring traffic on the near side of the median. Earlier studies came to a similar conclusion about this shadowing and [2] noted the need for one unit for each direction in most situations. Although the median may impact performance on the near side, many sites in California and elsewhere do not have median shoulders and the results for lane 1 should be representative of these locations. The analysis includes lane 1 for completeness, but the reader may choose to ignore it based on these comments. The president of EIS (who lead the original development of the ) conducted the final calibration on November 3, 1999 and realigned the unit to eliminate the problems in lane 5. Finally, according to the EIS representative 1

4 Coifman who did the initial calibration, the delays the end of each pulse in the contact closure output by a fixed.15 sec to prevent erroneous dropouts. The installation did not use the EIS Interface Card, which reportedly corrects for this extension to replicate the detection zone of a loop. The data used in this study were collected in late 1999 and mid 2. The remainder of this paper examines the performance of the relative to loop detectors, closing with a discussion and conclusions. PERFORMANCE The analysis compares data from the against the downstream loop in the given lane for the detector station shown in Figure 1. This choice to use loops for the reference was made for several reasons, first, loop detectors are the de facto standard in most states. Second, the performance of the loops used in this study has been validated using microscopic data analysis tools [8-1]. Third, other research has demonstrated the reproducibility of the loop measurements from this detector station with those at neighboring stations in the process of reidentifying vehicle measurements between stations [11-12]. The reidentification work requires accurate measurement of individual vehicle passages, on times, velocities and lengths in each lane, independently at two different detector stations. Fourth, as noted above, earlier studies found that properly installed loop detectors provided very accurate count measurements [2-4]. After estimating the detector spacing, the analysis uses the aggregated data reported by the using its internal controller emulation, then the focus shifts to aggregate measurements from the contact closure data. For reference, the work also compared one loop against another in the dual loop detectors. Detector Spacing Throughout this study the performance is compared to the downstream loop detector in the given lane and the two loops in the dual loop detector are compared against one another. Of course the detection zones are spatially separated, e.g., Figure 2A shows an example in the timespace plane of a vehicle passing through the three detection zones of a single lane from Figure 1. The contact closure data records a pulse from each detector, consisting of a rising edge and falling edge, the difference between the two is simply the on-time, denoted in the figure as OT 1, OT 2, OT R, respectively for the, downstream loop and zone. Based on the spatial constraints shown, it is clear that a vehicle must pass the, then the downstream loop, then the. The leading edge of the loops in each lane is a known 2 ft apart but the exact location of the detection zone varied from lane to lane. To deduce the distance between the downstream loop and the, relative to the, all pulses at the are matched to the subsequent pulse at the downstream loop and all pulses at the are matched to the preceding pulse at the downstream loop. In the event that a pulse at the downstream loop is matched to exactly one pulse at the and one pulse at the, we can measure the traversal time between the three detectors, denoted TT L and TT R in Figure 2. For these measurements the spacing between the downstream loop and the detection zone, D R [ft], can be estimated using the following equation: D R = 2 TT R TT L (1) 2

5 Coifman As shown in [1], there are some lane change maneuvers over these detectors, but their numbers are small. Applying Equation 1 to a large number of uniquely matched pulses across the three detectors, ranging between 31, and 99, cases across the lanes, the median value for D R for the rising edge ranged between 6 and 11 ft across the five lanes while D R for the falling edge ranges between 13 and 2 ft, with the effective detection zone being 3 to 13.5 ft larger than the loops', increasing as one moves from the lane closest to the towards the median. These distributions show that the detection zone ends within 2 feet downstream of the end of the downstream loop detection zone in each lane. The actual size of the loop detection zone should be approximately 6 feet, but it was not measured (see [1] for more details). Note that this analysis did not correct for the fact that the delays the end of each pulse in the contact closure output by a fixed.15 sec. So the trailing edge of the detection zone should be even closer to the downstream loop than value reported above. In other words, the detection zone is closer to the downstream loop than the two loops are to one another. Aggregated Data Reported by the This first effort examines the aggregated data reported by the using its internal controller emulation. Four data collection runs were conducted, though the first two are omitted from this paper due to the aforementioned potential for calibration errors prior to the final alignment. The remaining two runs, presented below, occurred after the final calibration. The specific timing of the runs were arbitrary, selected ahead of time based on convenience of accessing the site and such that each run would likely include both free flow and congested traffic conditions. Run 3 This run collected data from the using conventional 3 sec sample periods, the study period lasted 1 hrs, four of which were congested. The unit was calibrated by the president of EIS and as a result, these data should be representative of the best possible calibration for the location. According to the EIS representative, the unit estimates velocity in manner similar to what is conventionally used at single loop detectors, estimated velocity = flow vehicle length occupancy (2) However, the unit excludes all vehicles that occupy the detector for more than three times the current average on-time. Presumably, this exclusion will eliminate most long vehicles from the sample (see [13-16] for further discussion of these impacts). Another difference is that the unit continually updates the vehicle length estimate in each zone, to adjust to changing traffic conditions. Figure 3 shows the scatter plots comparing the velocity reported by the internal controller against the measured velocity from the dual loop detectors in the same lane. All of the points would fall on the diagonal axis if the two detectors provided identical data, while any difference will cause the points to fall off of the axis. According to the dual loop detector data, the two inside lanes are characterized by few vehicles over 25 feet while approximately 1 percent of the vehicles in the other lanes are over 25 feet. The presence of long vehicles may have skewed the estimated "vehicle length" estimate used by the in the outer three lanes. In contrast, Figure 4 shows the estimated velocity using flow and occupancy from the downstream loop detector via Equation 2. So this figure shows estimated versus measured velocity from loop data. The average effective vehicle length was assumed to be 21 feet in all lanes for the loop estimate. Comparing Figures 3 and 4, the appears to be 3

6 Coifman noisier than the single loop estimates in lane 1 and 2, tighter in lanes 3 and 4 (presumably due to excluding long on-times), and biased in lane 5. To quantify the magnitude of errors the work used the root mean squared error (RMSE) and bias of as given parameter over n samples as follows: RMSE = n i=1 * ( x ˆ i x i ) 2 n (3) Bias = n i=1 * ( x ˆ i x i ) n (4) where x i * is the measured value from the dual loop detector for the i-th sample and ˆ x i is the corresponding measurement or estimate from the (or in the case of Figure 4, from the single loops). The resulting RMSE and Bias for each lane are reported in the first five columns of Tables 1 and 2, respectively. Note that the measures are reported for velocity, flow and occupancy. The latter two will be discussed shortly. The table also includes a comparison against the detector in the given lane. For reference, the next five columns show the auto-correlation between successive measurements from the dual loop detector and the final five columns show the auto-correlation between successive measurements. For flow and occupancy, Table 1 shows that the RMSE of versus Loop is less than either of the autocorrelations in this run. Run 4L With the sampling every 1 sec, this run attempted to analyze the performance of the unit running at its fastest sampling period. Once more the study period lasted 1 hrs, four of which were congested. The results are noisy and given the separation between detectors, comparison might not be merited at such a short sampling period. Instead, after collecting the data, the flow and occupancy data were aggregated up to 3 sec samples and the analysis applied to Run 3 is now used for longer samples. Figure 5 presents scatter plots of flow, clearly showing that the longer sampling period brought most of the observations to the central axis. Moving to occupancy in Figure 6, relative to the loop detector, the overestimates occupancy for high values in lane 2 and underestimates in lanes 4 and 5. Overestimation in lane 2 may be due to occlusion from lanes 3-5, but it is not clear why the unit is underestimating this parameter in the outside lanes unless the change in detection zone size impacts the performance. The RMSE and Bias statistics for this run are reported in Tables 1-2. Two approaches were used to aggregate the velocity data up to the 3 sec sample period. First, 3 successive 1 sec samples of average velocity were averaged together. This approach does not replicate precisely what the would report for a 3 sec sample, rather, it should be taken as an lower bound on performance when comparing the to velocity measured by the dual loops over the 3 sec sample. Nevertheless, this estimate was comparable to or better than the conventional single loop estimate for lanes 3-5. The second approach used Equation 2 applied directly to the flow and occupancy after aggregating up to the 3 sec sample period. It is clear that these results are worse than the other approach; however, the estimates from 4

7 Coifman Equation 2 are presented because they are an indication of how accurately the measures flow and occupancy. Contact Closure Data Reported by the This section considers aggregate measurements from the contact closure data from the and dual loop detector collected over a five day period starting on May 25, 2. Figure 7 compares the cumulative distribution function (CDF) of the flow difference relative to the downstream loop using 5 min samples, q 2 -q R and q 2 -q 1, (continuing the use of subscripts, 1, 2, R, denoting respectively the, downstream loop, and ). The mean and median for these distributions are close to zero in all lanes, but the has a larger variance compared to that of the, as evident in the larger tails in the distributions. The first few columns of Table 3 show the average absolute percent error for all samples that had a non-zero flow at the downstream loop, (q 2 -q R )/q 2 and (q 2 -q 1 )/q 2. Except for lane 1, the flow is within 1 percent of the downstream loop, while the is within three percent for all lanes. There were two periods of congestion during the data collection and one period where lanes 1-3 were closed for about six hours due to maintenance activity downstream of the detector location. The data from these lane closures were excluded from the results in Table 3 because the downstream loop reported zero values during these periods. The remaining columns of Table 3 subsample the data by a flow threshold of 1 vph and then separately by the velocity threshold of 45 mph. The error increases at low velocity in lanes 2 and 5 while dropping in lanes 3 and 4 at lower velocities. The RMTS error increases in all lanes at low flow. Similar trends are evident in the loop data, though the relative change is much smaller for the loops. Repeating this analysis using occupancy (occ), Figure 8 shows the CDF of the difference relative to the downstream loop, occ 2 -occ R and occ 2 -occ 1. Before measuring occ R, the recorded OT R was shortened by.15 sec to remove the extension delay on the falling transition (the results from the degrade in all but lane 5 if the extension delay is retained). The mean and median in lanes 1-2 are similar for the and, but again the variance is larger for the. As one moves to the outside lanes, closer to the unit, there is a clear bias in the measurement. The underestimates occupancy relative to the downstream loop. In contrast, the s only show a slight increase in variance as one moves to the outer lanes. Some of this error is due to the spatial separation between the detection zones, as is evident in the distribution. Recall that from the downstream loop, the upstream loop is further away than the detection zones. Again, the first few columns of Table 3 show the average absolute percent error for all samples that had a non-zero occupancy at the downstream loop, (occ 2 -occ R )/occ 2 and (occ 2 -occ 1 )/occ 2. The average error for loop 1 ranges between 1.5 percent and 2.7 percent across the lanes. The average error from the is best in lane 3, at 8.8 percent and degrades as one moves away from this lane, reaching a maximum of 57.4 percent error in lane 1. As shown in the table, after subsampling the data by the flow threshold and separately by the velocity threshold, at low velocities the performance improves in the lanes closest to the sensor unit but degrades in lanes further from the sensor. In contrast, the s show slightly higher error at low velocities compared to their performance at higher velocities, with a maximum of 6. percent. Returning to the, one sees a slight increase in the occupancy percent error at lower flows for the near lanes, with significant degradation in lanes 4 and 5. These results are important, not only did the sensor show large errors relative to the downstream loop, but the magnitude of those errors 5

8 Coifman depend on the lane. These results suggest that it would be difficult to derive a simple correction factor to map occupancy to loop occupancy. This analysis was repeated using 3 sec samples with similar results, as shown at the bottom of Table 3. Highlighting the differences compared to the 3 sec data, 5 min flow from the upstream loop improves by roughly a factor of three. The performance improves slightly in the lanes closest to the, but remains roughly the same in lane 2 and has degrades in lane 1. The 5 min loop occupancy performance improves by roughly a factor of two while the performance does not change significantly. The analysis in this section has focused on aggregating the contact closure output from the collected through a traffic controller, concurrent 3 sec aggregate data were also reported by the unit. We were not able to perfectly synchronize the and controller clocks over this extended study because the sample period was slightly longer than 3 sec, making one-to-one comparisons difficult between the two outputs at a microscopic scale. Presumably the flow and occupancy measurements reported directly from the should be identical to what was derived from the contact closure output. The loop and flows are already close, but significant discrepancies were found between the loop and occupancy measured by the controller. To verify that these results are representative of the, the analysis compared the occupancy measured directly by the unit and the occupancy measured from the contact closure data via the controller. Comparing the two time series over 24 hours, the two data sets appear to fall on top of one another [1]. Although the sample periods are not identical, taking the difference between the two series, an error in one sample due to being slightly out of phase should be offset by an opposite error in the next sample. The mean difference, or bias, is roughly.5 percent occupancy in each lane (ranging between.33 and.57 across the lanes), exactly what one would expect if the data reported directly from the truncates occupancy to integer percent. Figure 9 shows three different hour long periods from lane 2 (note that the vertical scales change from plot to plot). The top plot shows the period between hrs, which exhibited high free flow occupancies. The middle plot shows the period between hrs, which exhibited high congestion occupancies. The bottom plot shows the period between hrs, which exhibited low free flow occupancies, and the truncated occupancies are clearly evident. These plots are typical of all five lanes. The strong correlation between the two time series occupancies, direct from the and via the controller, verifies the EIS representative's statement that the delays the end of each pulse in the contact closure output by a fixed.15 sec. Reportedly, the most recent revision of the does not truncate occupancy to integer percent. CONCLUSIONS This study examined the performance of the sensor deployed in side-fire mode and the following conclusions only apply to this mode of operation. First, the manufacturer's general performance claims are over-ambitious for the sensor, most measures are not as clean as loop detectors; still, the promises significant cost savings and could be used to (relatively) cheaply provide information where none was previously available. Although the study site was reportedly less than ideal from the manufacturer's standpoint, it is representative of common freeway geometry and the was not able to monitor the opposing lanes at this location. While the occupancy and flow measures are noisier than loops, the velocity estimates are 6

9 Coifman almost as good as those from single loops (while being inferior to direct measurement from dual loops). Although many operating agencies use single loop detectors estimate velocity, our research has shown that there are several ways to improve upon the quality of these conventional estimates (and presumably the by extension) [13-16]. In the mean time, it may be advisable not to use the to estimate velocity for samples less than 5 minutes long (the author would make a similar argument for conventional velocity estimation from single loop detectors). Of course, if a higher sampling rate is desired, a moving average or exponential filter could be used on the flow and occupancy measurements. For reference, the work also compared one loop against another in a dual loop detector. The analysis consisted of aggregate data comparisons. The accuracy of the subject loop detector station has been verified elsewhere in the literature, e.g., [1]. The work used the redundancies of the dual loop detector to verify the performance of each loop in a given lane. It also showed that the two loops were further apart than the detection zone was from the closest loop. Hence, the loop comparison should provide an upper bound on the expected errors in the data due to the spatial separation. Consistent with earlier studies, the flow measurements are within 1 percent of loops, though the two loops were within 3 percent of one another. These flow measurements should be good enough for many real-time operations applications but may not be good enough for calculating average annual daily travel (AADT). The occupancy measurements showed significant discrepancies relative to the loops, ranging between 13 and 4 percent across lanes 2-5, showing an apparent lane dependency. These results suggest that it would be difficult to derive a simple correction factor. Even if one excludes the results from both lane 1 and lane 5, the occupancy measurements may not be good enough for applications such as traffic responsive ramp metering. These findings are consistent with those of the Caltrans Traffic Operations Program Data Collection Functional Requirements Task Force. The task force noted that in areas without existing loops, the can be deployed in side fire mode to, "quickly fill in gaps where detection is poor, non existent, or out due to construction. It must be accepted that this data is significantly less than perfect, but some data is always better than none." [17] Of course all of the figures are presented in this document, allowing the reader to assess the performance of the and reach their own conclusions. After all, the appropriateness of the detector depends on the applications, detection network and a host of other factors that are far beyond the scope of this work. ACKNOWLEDGEMENTS This work was performed as part of the California PATH (Partners for Advanced Highways and Transit) Program of the University of California, in cooperation with the State of California Business, Transportation and Housing Agency, Department of Transportation. The Contents of this report reflect the views of the authors who are responsible for the facts and accuracy of the data presented herein. The contents do not necessarily reflect the official views or policies of the State of California. This report does not constitute a standard, specification or regulation. 7

10 Coifman REFERENCES [1] Wang, J., Case, E., Manor, D., "The Road Traffic Microwave Sensor ()", Proc. of the 3rd Vehicle Navigation and Information Systems Conference, IEEE, 1992, pp [2] Klein, L, Kelley, M [Hughes Aircraft Company], Detection Technology for IVHS: Final Report, FHWA, 1996, FHWA-RD [3] MNDOT Field Test of Monitoring of Urban Vehicle Operations Using Non-Intrusive Technologies, FHWA, 1997,FHWA-PL [4] Middleton, D, Jasek, D., Parker, R., Evaluation of Some Existing Technologies for Vehicle Detection, Texas Transportation Institute, 1999, FHWA/TX-/1715-S. [5] Wald, W., Microwave Vehicle Detection, Final Report, Caltrans, 24, [6] Coifman, B., Lyddy, D., Skabardonis, A., "The Berkeley Highway Laboratory- Building on the I-88 Field Experiment", Proc. IEEE ITS Council Annual Meeting, 2, pp 5-1, [7] Skabardonis, A., Petty, K., Noeimi, H., Rydzewski, D., Varaiya, P., "I-88 Field Experiment: Data-Base Development and Incident Delay Estimation Procedures", Transportation Research Record 1554, TRB, 1996, pp [8] Coifman, B., "Using Dual Loop Speed Traps to Identify Detector Errors", Transportation Research Record no. 1683, Transportation Research Board, 1999, pp [9] Coifman, B., Dhoorjaty, S. "Event Data Based Traffic Detector Validation Tests", ASCE Journal of Transportation Engineering, Vol 13, No 3, 24, pp [1] Coifman, B., An Assessment of Loop Detector and Performance, Automated Diagnostics of Loop Detectors and the Data Collection System in the Berkeley Highway Laboratory- Part II, PATH research report, University of California, 24. [11] Coifman, B., Cassidy, M. "Vehicle Reidentification and Travel Time Measurement on Congested Freeways", Transportation Research: Part A, vol 36, no 1, 22, pp [12] Coifman, B. "Identifying the Onset of Congestion Rapidly with Existing Traffic Detectors", Transportation Research: Part A, vol 37, no 3, 23, pp [13] Coifman, B., Dhoorjaty, S., Lee, Z. "Estimating Median Velocity Instead of Mean Velocity at Single Loop Detectors", Transportation Research: Part C, vol 11, no 3-4, 23, pp [14] Neelisetty, S., Coifman, B., "Improved Single Loop Velocity Estimation in the Presence of Heavy Truck Traffic" Proc. of the 83rd Annual Meeting of the Transportation Research Board, 24. [15] Coifman, B. "Improved Velocity Estimation Using Single Loop Detectors", Transportation Research: Part A, vol 35, no 1, 21, pp [16] Jain, M., Coifman, B., "Improved Speed Estimates from Freeway Traffic Detectors" Proc. of the 83rd Annual Meeting of the Transportation Research Board, 24. [17] Caltrans Traffic Operations Program Data Collection Functional Requirements Task Force, Draft Data Collection Functional Requirements, 21. 8

11 Coifman LIST OF FIGURES Figure 1, (A) Photo showing the configuration of Station 7 in the BHL. (B) Schematic of the data collection site at the station. Figure 2, One vehicle passing over a dual-loop-detector and the from the previous figure, (A) the three detection zones and the vehicle trajectory as shown in the time space plane. The height of the vehicle's trajectory reflects the non-zero vehicle length. (B) The recorded response at each detector. Figure 3, Lane by lane comparison of velocity reported by the and measured by the dual loop detectors, T=3 sec, November 3, Figure 4, Lane by lane comparison of velocity estimated by the downstream loop and measured by the dual loop detectors, T=3 sec, November 3, Figure 5, Lane by lane comparison of flow reported by the and measured by the dual loop detectors, T=5 min, November 9, Figure 6, Lane by lane comparison of occupancy reported by the and measured by the dual loop detectors, T=5 min, November 9, Figure 7, Cumulative distribution by lane of the flow difference relative to the downstream loop for the and for each 5 min sample. Figure 8, Cumulative distribution by lane of the occupancy difference relative to the downstream loop for the and for each 5 min sample. Figure 9, Details of 3 sec occupancy reported directly by the, 'x', and as calculated from the contact closure output reported to the controller, 'o' in lane 2. These results were typical of all five lanes. Note the three plots are at different scales. 9

12 Coifman LIST OF TABLES Table 1, RMSE for the various runs. First five columns are or versus the downstream loop, the next five columns are loop(n+1) versus loop(n) and the final five columns are (n+1) versus (n). These last ten columns are included for reference. Table 2, Bias for the various runs. The columns correspond to those of the previous table. For the first five columns, a positive value indicates the, on average, overestimated the given parameter and a negative value indicates the underestimated it. Table 3, Average absolute percent error in aggregated flow and occupancy from the contact closure data for the and relative to the downstream loop. 1

13 A) upstream loops downstream loops eastbound CCTV Controller cabinet B) Westbound median barrier lane 1 lane 2 lane 3 lane 4 lane 5 Eastbound Figure 1, downstream loops upstream loops detection zone CCTV pole with unit (A) Photo showing the configuration of Station 7 in the BHL. (B) Schematic of the data collection site at the station.

14 (A) distance Detection Zone Downstream Loop's Detection Zone 2 ft Upstream Loop's Detection Zone time Vehicle Trajectory TT R TT L (B) OT R Response on off OT 2 Second Loop's Response on off OT 1 First Loop's Response on off time Figure 2, One vehicle passing over a dual-loop-detector and the from the previous figure, (A) the three detection zones and the vehicle trajectory as shown in the time space plane. The height of the vehicle's trajectory reflects the non-zero vehicle length. (B) The recorded response at each detector.

15 8 lane 1 8 lane 2 velocity (mph) velocity (mph) dual loop velocity (mph) dual loop velocity (mph) 8 lane 3 8 lane 4 velocity (mph) velocity (mph) dual loop velocity (mph) dual loop velocity (mph) 8 lane 5 velocity (mph) dual loop velocity (mph) Figure 3, Lane by lane comparison of velocity reported by the and measured by the dual loop detectors, T=3 sec, November 3, 1999.

16 estimated velocity single loop (mph) lane dual loop velocity (mph) estimated velocity single loop (mph) lane dual loop velocity (mph) estimated velocity single loop (mph) lane dual loop velocity (mph) estimated velocity single loop (mph) lane dual loop velocity (mph) estimated velocity single loop (mph) lane dual loop velocity (mph) Figure 4, Lane by lane comparison of velocity estimated by the downstream loop and measured by the dual loop detectors, T=3 sec, November 3, 1999.

17 2 lane 1 2 lane 2 flow (veh/5min) flow (veh/5min) downstream loop flow (veh/5min) downstream loop flow (veh/5min) 2 lane 3 2 lane 4 flow (veh/5min) flow (veh/5min) downstream loop flow (veh/5min) downstream loop flow (veh/5min) 2 lane 5 flow (veh/5min) downstream loop flow (veh/5min) Figure 5, Lane by lane comparison of flow reported by the and measured by the dual loop detectors, T=5 min, November 9, 1999.

18 5 lane 1 5 lane 2 occupancy (%) occupancy (%) downstream loop occupancy (%) 2 4 downstream loop occupancy (%) 5 lane 3 5 lane 4 occupancy (%) occupancy (%) downstream loop occupancy (%) 2 4 downstream loop occupancy (%) 5 lane 5 occupancy (%) downstream loop occupancy (%) Figure 6, Lane by lane comparison of occupancy reported by the and measured by the dual loop detectors, T=5 min, November 9, 1999.

19 1.8 1 lane 1 lane 2.8 CDF.6.4 CDF count (# vehicles) count (# vehicles) 1 1 lane 3 lane CDF.4 CDF count (# vehicles) count (# vehicles) 1 lane 5.8 CDF count (# vehicles) Figure 7, Cumulative distribution by lane of the flow difference relative to the downstream loop for the and for each 5 min sample.

20 1.8 1 lane 1 lane 2.8 CDF.6.4 CDF occupancy (percent) occupancy (percent) 1 lane 3 lane CDF.6 CDF CDF occupancy (percent) occupancy (percent) lane occupancy (percent) Figure 8, Cumulative distribution by lane of the occupancy difference relative to the downstream loop for the and for each 5 min sample.

21 3 25 occupancy (%) time of day (h) 7 6 occupancy (%) time of day (h) 1 8 occupancy (%) Figure 9, time of day (h) Details of 3 sec occupancy reported directly by the, 'x', and as calculated from the contact closure output reported to the controller, 'o' in lane 2. These results were typical of all five lanes. Note the three plots are at different scales.

22 Table 1, RMSE for the various runs. First five columns are or versus the downstream loop, the next five columns are loop(n+1) versus loop(n) and the final five columns are (n+1) versus (n). These last ten columns are included for reference. RMSE versus downstream Loop (or RMSE downstream Loop(n+1) versus RMSE (n+1) vs. (n), lane dual loop for velocity), lane # downstream Loop(n), lane # # units run T (sec) measured flow (vph) flow (vph) measured occupancy occupancy reported velocity loop estimated velocity (a) loop measured velocity (b) measured flow (%) (%) (mph) (mph) (mph) (vph) 4 L flow (vph) 4 L measured occupancy (%) 4 L occupancy (%) 4 L estimated (mph) 4 L velocity (c) average reported velocity (d) (mph) 4 L loop estimated (mph) 4 L velocity (a) loop measured velocity (b) (mph) 4 L (a) Loop estimated velocity via Equation 2. (b) The loop measured velocity is from dual loops, not just the downstream loop. (c) estimated velocity via Equation 2 after aggregating 1 sec flow and occupancy over 3 successive samples. (d) The average of 3 successive 1 sec samples of reported velocity from the.

23 Table 2, Bias for the various runs. The columns correspond to those of the previous table. For the first five columns, a positive value indicates the, on average, overestimated the given parameter and a negative value indicates the underestimated it. Bias versus downstream Loop (or dual Bias downstream Loop(n+1) versus loop for velocity), lane # downstream Loop(n), lane # Bias (n+1) vs. (n), lane # units run T (sec) measured flow (vph) flow (vph) measured occupancy occupancy reported velocity loop estimated velocity (a) loop measured velocity (b) measured flow (%) (%) (mph) (mph) (mph) (vph) 4 L flow (vph) 4 L measured occupancy (%) 4 L occupancy (%) 4 L estimated (mph) 4 L velocity (c) average reported velocity (d) (mph) 4 L loop estimated (mph) 4 L velocity (a) loop measured velocity (b) (mph) 4 L (a) Loop estimated velocity via Equation 2. (b) The loop measured velocity is from dual loops, not just the downstream loop. (c) estimated velocity via Equation 2 after aggregating 1 sec flow and occupancy over 3 successive samples. (d) The average of 3 successive 1 sec samples of reported velocity from the.

24 Table 3, Average absolute percent error in aggregated flow and occupancy from the contact closure data for the and relative to the downstream loop all v > 45 mph v < 45 mph q > 1 vph q < 1 vph # of upstream # of upstream # of upstream # of upstream # of upstream lane samples loop samples loop samples loop samples loop samples loop Flow 5 min % 31.5% 988.7% 3.1% 69.3% 27.5% 174.5% 4.8% 998.8% 36.8% Flow 5 min % 8.9% 13.6% 4.3% % 12.7% 715.7% 4.2% 471.8% 17.4% Flow 5 min % 5.1% 127.9% 4.6% 18.8% 4.3% 793.8% 4.4% % 6.6% Flow 5 min % 5.9% % 6.1% 19.9% 3.6% 852.9% 4.6% % 9.% Flow 5 min % 6.5% % 6.3% 116.7% 8.6% 824.9% 4.9% % 9.9% Occupancy 5 min % 57.4% % 59.1% 69 2.% 22.3% % 9.% % 66.9% Occupancy 5 min % 12.9% 13.8% 6.8% % 19.6% % 7.1% % 23.3% Occupancy 5 min % 8.8% % 9.4% % 3.4% % 8.6% % 9.4% Occupancy 5 min % 22.1% % 22.8% % 14.7% % 2.7% % 25.4% Occupancy 5 min % 4.% % 41.1% % 29.8% % 41.3% % 37.3% Flow 3 sec % 17.4% % 17.6% % 13.6% % 9.2% % 19.9% Flow 3 sec % 8.1% % 7.1% % 17.4% % 7.1% % 9.7% Flow 3 sec % 7.% % 6.4% % 12.5% % 6.1% % 8.8% Flow 3 sec % 8.% % 7.9% % 8.1% % 6.6% % 11.% Flow 3 sec % 1.% % 9.6% % 12.1% % 6.8% % 16.1% Occupancy 3 sec % 35.4% % 36.5% % 16.% % 17.% % 4.8% Occupancy 3 sec % 15.6% % 14.8% % 23.2% % 13.7% % 18.7% Occupancy 3 sec % 13.3% % 13.7% % 8.9% % 11.6% % 16.6% Occupancy 3 sec % 22.8% % 23.6% % 15.1% % 21.2% % 26.4% Occupancy 3 sec % 41.7% % 42.9% % 31.2% % 41.7% % 41.6%