Intersection Signal Data for Traffic Monitoring 20150801 final docxdocs.trb.org/prp/16-4355.pdf ·...

Guin, Hunter, Rodgers, Anderson, Susten, and Weigand

Integrating Intersection Traffic Signal Data into a Traffic 1

Monitoring Program 2

3

Authors: 4

5

Angshuman Guin, Ph.D. (Corresponding author) 6 Senior Research Engineer, Georgia Institute of Technology, [email protected] 7 School of Civil and Environmental Engineering 8 Georgia Institute of Technology 9 Atlanta, Georgia 30332-0355 10 Phone: 404.894.5830 11 Fax: 404.894.2278 12

13

Michael Hunter, Ph.D. 14 Associate Professor, Georgia Institute of Technology, [email protected] 15 School of Civil and Environmental Engineering 16 Georgia Institute of Technology 17 Atlanta, Georgia 30332-0355 18 Phone: 404.385.1243 19 20

Michael Rodgers, Ph.D. 21 Principal Research Scientist, Georgia Institute of Technology, [email protected] 22 School of Civil and Environmental Engineering 23 Georgia Institute of Technology 24 Atlanta, Georgia 30332-0355 25 Phone: 404.385.0569 26 27

James Anderson, Civil Designer, AECOM, [email protected] 28 400 Northpark Town Center, Suite 900 29 1000 Abernathy Road NE 30 Atlanta, Georgia 30328 31 Phone: 678-808-8952 32

33

Scott Susten 34 Transportation Data Group Leader, Georgia Department of Transportation, [email protected] 35 Office of Transportation Data – 20th Floor 36 600 West Peachtree St NW 37 Atlanta, GA 30308 38 Phone: (404) 347-0695 39

40

Kiisa Wiegand 41 Business Analyst, Georgia Department of Transportation, [email protected] 42

Guin, Hunter, Rodgers, Anderson, Susten, and Weigand 1

Office of Transportation Data – 20th Floor 1 600 West Peachtree St NW 2 Atlanta, GA 30308 3

4

Word Count: 7403 (text: 5653; tables: 1; figures: 6) 5

Submission Date: August 1, 2015 6


ABSTRACT 1

There are ongoing efforts towards leveraging the large volume of data collected to support real-2 time traffic operations for various other non-real time uses. However, adapting data streams to 3 a different purpose can be challenging especially if different applications use the data in 4 different ways. While both the traffic operations and traffic monitoring programs often use 5 similar technology for vehicle detection, they may have different sensitivities to potential errors 6 in the process. While small errors in traffic counts may not be critical in traffic operations where 7 data is refreshed every few seconds, the same error levels could become much more significant 8 in traffic monitoring where data, and corresponding errors, are typically aggregated over longer 9 periods of time. This study investigates if and how traffic volume data from detectors at 10 signalized intersections could be appropriately used in a traffic monitoring program that 11 supports the Highway Performance Monitoring System (HPMS). For this purpose, this study 12 evaluates both the accuracy and representativeness of these traffic signal detector data by 13 comparison with to standard HPMS-type traffic count data obtained by portable pneumatic 14 counters under different traffic flow conditions and different intersection geometries. The 15 results indicate that, although the results can vary by site, these traffic signal detectors can 16 produce 15-minute aggregate traffic counts of comparable quality to portable pneumatic tube 17 counters. That is, having a 90% accuracy at a 95% level of confidence for the 15-minute 18 aggregate counts. By examining the representativeness of data under different conditions, the 19 study develops a set of eligibility criteria that can be used to identify intersections that are 20 suitable for performance monitoring data collection. 21

22 23


INTRODUCTION 1

The Highway Performance Monitoring System (HPMS) is designed to monitor traffic and to 2 provide estimates for representative average annual daily traffic (AADT) throughout the 3 national highway systems and those of individual states. Historically, most of the data used to 4 produce these estimates have been derived from traffic counts collected at relatively infrequent 5 intervals and aggregated over relatively long periods, typically fifteen minutes to several hours, 6 using portable pneumatic counters or similar devices. However, many signalized intersections 7 operate under some form of actuated control, in that the intersection approaches have some type 8 of vehicle detection sensor used in the control process. The most common such sensor is an 9 inductive loop that can be configured to provide traffic counts. These inductive loops are 10 typically installed on high volume corridors for the purpose of improving traffic flow and 11 reducing vehicle emissions by optimization of signal timing rather than making estimates of 12 traffic volume. This is a significant difference in that traffic monitoring (e.g. current HPMS 13 measurements) and traffic operations data (e.g. the detectors at signalized intersections) are 14 often used for different purposes and errors of little significance to one application may be 15 critical to another. For example, while exact traffic counts may not be critical in traffic 16 operations, where data are refreshed every few seconds, these same errors may be critical for 17 traffic monitoring if these errors are not random and accumulate over the longer time intervals 18 used for traffic monitoring and produce a biased estimate. This study investigates if and how the 19 traffic volume data from these detectors at signalized intersections could be appropriately 20 utilized in a traffic monitoring program that supports the Highway Performance Monitoring 21 System (HPMS). 22

23

BACKGROUND 24

The existing literature on the subject of intersection traffic signal data pertains to a wide range 25 of topics including: placement and calibration of the embedded loops; how the signal system 26 data should be archived; the appropriate schemas; and the development of scripts to transfer 27 data, process and upload the data. One of the studies (1) focused on equipment design and 28 installation practices, reviewing various loop configurations in terms of ability to provide traffic 29 counts, and selecting instances around the country that are using the exiting loops at traffic 30 signals for counts. However, it appears that the traffic counts used by the different agencies 31 were used only for real-time traffic monitoring; not for federal reporting purposes or as part of 32 HPMS. 33

Another pilot study (2) demonstrated the feasibility of real-time monitoring of traffic operations 34 at a signalized intersection. Using traffic volumes from existing loops and signal timing 35 information, the control delay (uniform delay plus incremental delay) was estimated on a cycle-36 by-cycle basis using equations provided in the Highway Capacity Manual. 37

There are a few other studies (3,4) that investigated the use of data from signalized intersections 38 for real-time traffic monitoring. There have been some significant advances made in using 39 high-resolution event data at signals for monitoring real-time corridor performance (5, 6). 40 However there is limited amount of research on the accuracy and robustness of the data as it 41 pertains to the use of the data for long term performance monitoring. The current study aspires 42 to bridge this gap and provide an objective evaluation of the accuracy of the data and feasibility 43 of use as it pertains to federal reporting standards. 44

Accuracy of Vehicle Counting Technology 45


Inductive loop based vehicle sensors have been in use for a long time in the transportation field 1 and is one of the most well established intrusive detection technologies (1,7). Texas 2 Transportation Institute has reported the count accuracy to be 98 percent for properly designed 3 and installed, preformed and standard saw-cut loops (4). The Minnesota Department of 4 Transportation, based on field tests, have reported inaccuracies of 0.1 % - 3% for hourly counts 5 on freeways and 2.8% to 8.6% at intersections (2,3). The accuracy of inductive loop varies with 6 environmental conditions and is also dependent on the installation and calibration. 7

The survey conducted by Sherry (1) indicated that inductive loops are used in traffic monitoring 8 programs in 50 states and pneumatic tubes are used in 48 states. Pneumatic tubes have been 9 used for vehicle counts since the early 1900s and their accuracy is widely accepted for 10 temporary counts. There is minimal recent literature regarding any accuracy studies for 11 pneumatic tube counters. A South Dakota study reported a wide -0.92% to 30% error range for 12 pneumatic tubes (5). McGowen and Sanderson (9) also reported a wide error range, 6.9% to 13 27.6%, for 15 minute intervals. However over a two day period the error was in the range of 14 0.0% to 0.5%. 15

Use of Intersection Signal Data for Vehicle Counts 16

Using traffic operations data for traffic monitoring programs is an attractive idea and naturally, 17 there have been previous efforts towards achieving this goal. An extensive study was 18 performed by the Institute of Traffic Engineers (ITE) Traffic Engineering Council to study the 19 use of existing loops at signalized intersections for traffic counts (10). The study examined the 20 use of data from different types of loop configurations including exit loops, queue loops, 21 advance loops and stop line loops for single loops; stop line loops and extension loops for 22 multiple loop configurations; three and four loop configurations; as well as continuous long 23 loops. The study also provided guidance on intersection geometry considerations such as 24 presence of center median or island, loops beyond stop line, road transitions, adjacent lane 25 detection in same and opposite direction of travel, pavement surface, loop crosstalk, use of 26 multiple loops etc. Based on the information provided on the case studies in this report it 27 appeared that as of 2007, the following cities and state DOT have attempted to use signal loop 28 data for vehicle counts: 29

• City of Nashua, NH 30

• City of Fremont, CA 31

• City of Bellevue, WA 32

• Minnesota Department of Transportation 33

• North Carolina Department of Transportation 34

The survey in the report indicated that these entities used offline polling and retrieval of data at 35 the time the survey was conducted. The City of Fremont was the most advanced in this area and 36 they had decided to exclusively use far side system loops for the counts (where far side system 37 loops refer to induction loop detectors configured for vehicle counts and are installed 38 immediately outside the bounds of the intersection, downstream of the intersection). The ITE 39 study provides an appropriate pre-cursor to the current study in that the ITE study provides 40 guidance on best practices for the design and placement of the loops. The current study builds 41 on this effort and investigates the use of data from pre-installed offset (or advance) loops or 42 other technologies used for vehicle detection at intersections (such as video detection systems) 43 in a traffic monitoring program. 44


This study evaluates whether the traffic signal detector data is of comparable quality when 1 compared to portable traffic count data under different traffic flow conditions and different 2 intersection geometries. Based on these results, the study develops a set of eligibility criteria 3 that can be used to identify intersections that are suitable for performance monitoring data 4 collection. 5

6

METHODOLOGY 7

To test the feasibility of use of the operations data in long term performance monitoring, the 8 following criteria need to be met: 9

1. Availability of data 10

2. Accuracy of data 11

3. Representativeness of data 12

It is assumed that any proposal for use of these data comes with the implicit guarantee of its 13 availability. It is assumed that data communications work and the data are available, at least on 14 a monthly basis. In the Georgia Department of Transportation (GDOT) case study reported 15 here, the data were available with a maximum of 1 day of latency. The accuracy of data was 16 evaluated by spot measurement and comparison with data generated by other technologies as 17 well as manually collected ground truth data. To evaluate the representativeness of the data, 18 typical geometric and operations scenarios were identified and data were collected for 19 evaluation across sites with the different scenarios. 20

Site Selection 21

The list of potential sites to study was based on criteria such as geometric features, area 22 characteristics (urban, suburban, rural), current traffic demands and location. Rural sites could 23 not be included in the study because the traffic operations program collected data mostly on 24 high volume and congested corridors which naturally precluded rural intersections. The study 25 site selection was also driven by other criteria such as the availability of appropriate anchor 26 points for the pneumatic tubes near the proposed data collection location. Figure 1 shows the 27 typical setup of the different kinds of detection zones at an intersection on study corridors. The 28 final list of study sites shown in Figure 2 covered the following scenarios: 29

• High Volume Intersections (4 pairs) 30

• Multiple cross-streets between intersections (1 pair) 31

• No cross-streets between intersections (2 pairs) 32

• Medium Volume Intersections (1 pair) 33

• Video based detection (1 intersection) 34

35

36

Data Collection 37

Due to the resource intensive nature of video based data collection, the baseline data was 38 collected using pneumatic tubes similar to a standard portable data collection effort conducted 39 by DOTs for their traffic monitoring programs. To ensure that the pneumatic tube data is 40


reliable for use as baseline data for comparing with traffic signal detector data (also referred to 1 as signal detector data or simply signal count data in this paper), video based manually collected 2 data was used to perform quality checks on this data. 3

4

5 Figure 1: Typical Placement of Detectors on Pavement. Background Image Courtesy: Google 6 Maps 7

8

Offset Count Detectors

Offset Count Detectors

Presence Detectors

Beginning of Turn Lane

Stop Bar


1 Figure 2: Study Site Locations. Image Courtesy: Google Maps 2

3


1

Pneumatic Tubes 2

Portable count data were collected using pneumatic tube counters. The tubes were laid across 3 the roadway as close as possible to the location of the in-pavement inductive loops detectors of 4 the signal detectors. In several cases it was not possible to match the exact location because of 5 constraints regarding the availability of anchor points for the pneumatic tube equipment. In 6 such cases, the analysis has to take into account that differences in counts between the 7 pneumatic tubes and the inductive loops could arise due to lane changes of the vehicles and the 8 time offset between the arrivals of vehicles at the two detectors. The time-offset is typically 9 minor during medium to light volume conditions but could become a significant factor under 10 high volume conditions that lead to vehicle queuing over the detectors. 11

A total of six hundred and seventy two hours of directional data over nine study sites around 12 Metro Atlanta were collected for analysis in this study. 13

Video Processing 14

For validating the pneumatic tube counts, manual vehicle count data were collected. To ensure 15 repeatability and high quality of the data, counts were performed using pre-recorded video at the 16 count locations. The resulting video was processed by data collectors using a tablet application 17 that was developed at Georgia Tech (14). The data was obtained at a per-vehicle record level 18 and was aggregated up to 15 minute bins to match the data from the pneumatic tubes. The 19 video data were also used for validation of the data at intersections that used Video Detection 20 Systems for counts since the extremely high volume of traffic at these sites and the lack of 21 anchor points precluded using pneumatic tube counters. About 21 hours of per-lane manual 22 vehicle count data were collected and analyzed for this study. 23

24

Analysis steps 25

Validation of Baseline data (Pneumatic Tube data) 26

To ensure that the pneumatic tube counter data were of a quality sufficient to serve as a 27 baseline, the pneumatic tube data were validated against manually collected data. To ensure 28 repeatability and high quality of the data, counts were performed using pre-recorded video at the 29 count locations. 30

Comparison of Signal Detector Data to Baseline Data 31

Essential to the evaluation of the feasibility of integrating intersection traffic signal data into a 32 traffic monitoring program is an evaluation of whether the traffic signal detector data is of a 33 quality comparable to portable traffic count data under different traffic flow conditions and 34 different intersection geometries. In addition, it is necessary to evaluate the accuracy of both 35 systems to ensure that, apart from the comparative differences, the actual errors in the new 36 system are not excessive. Multiple paired comparisons were performed across the different 37 datasets to evaluate both criteria. 38

Visual analysis of the data was initially used to check for large differences or issues in the 39 underlying dataset. Time-series charts helped identify any systematic divergence between the 40 values or characteristic attributes of the divergence, e.g. whether the differences are higher 41 during certain flow conditions, time-of-day etc. Y-Y plots between the test data and the 42


baseline data showed the relationship of the magnitude of the error values to the magnitude of 1 the actual values. . 2

The following statistics were used to compare the differences between the observed counts and 3 the designated baseline counts. 4

• Mean Percentage Error (MPE) 5

• Mean Absolute Percentage Error (MAPE) 6

• Root Mean Squared Percentage Error (RMSPE) 7

8

For the inductive loop detector based signal detector counts, the pneumatic tube counts served 9 as the baseline. For the pneumatic tube counts as well as the video detection system based 10 signal detector counts, the manual counts served as the baseline. 11

Apart from the above measures of central tendency, the variability in the differences was 12 studied by using the 95th percent confidence intervals on the mean percentage error values. 13

14

RESULTS AND DISCUSSION 15

The results of the comparative tests at all the study sites are provided in Table 1. Table 1 16 provides, at a lane-by-lane level, the mean percentage errors, the 95% confidence interval 17 around the mean percentage errors, the mean absolute percentage errors and root mean square 18 percentage errors. 19

Quality Checks on Baseline Data 20

Based on the results of initial comparisons between the signal detector data and the pneumatic 21 tube data obtained in the pilot study, the quality of the portable traffic data collection devices 22 was deemed to be acceptable for purposes of this study. In view of concerns about pneumatic 23 tubes undercounting traffic under slow speed and high volume conditions, a detailed 24 comparison of video based counts collected by the research team and the pneumatic tube counts 25 was performed. 26

Y-Y plots for the pneumatic tube counter data and the manual count data at the SR 92 at 27 Woodstock / King Road intersection are provided in Figure 3. While most of the data are 28 within the 10% bounds (indicated by the red lines), there are some points in the lane 3 data that 29 fall outside these bounds. Overall, when aggregated across the lanes the data show a slight 30 positive bias. The mean percentage errors vary within a range of -6% to 8.7% for the individual 31 lane aggregates while the aggregate over all lanes has a mean error of 3%. For the SR 140 at 32 River Exchange Parkway site, a second site for manual data validation, the results were found to 33 be better at the individual lane level, but similar at the aggregate-over-lanes level. 34

A part of these errors can be explained by the possibility of lane changes between the location 35 of the tube count and the location of the screen-line for the manual count. However, the 36 magnitude of the errors is too high to attribute completely to lane changes and differences in the 37 arrival times at the two detection points. Due to the high degree of confidence in the manual 38 data, it can be concluded that there is a certain amount of error in the pneumatic tube counts. 39

40

41


1

2

Table 1: Summary of Percentage Errors 3

4

Location Direction Lane Evaluated Count MPE

Lower 95%

Upper 95% MAPE RMSPE

SR 140 @ River Exchange Dr EB Fast Signal 1.9% 4.0% -0.2% 8.0% 14.8%

SR 140 @ River Exchange Dr EB Slow Signal -5.2% -4.0% -6.4% 5.4% 10.1%

SR 140 @ River Exchange Dr EB All Signal -4.3% -1.3% -7.3% 3.6% 9.9%

SR 140 @ River Exchange Dr WB Fast Signal 4.4% 5.4% 3.4% 5.6% 8.3%

SR 140 @ River Exchange Dr WB Slow Signal -9.7% -7.8% -11.6% 9.8% 16.6%

SR 140 @ River Exchange Dr WB All Signal -7.6% -5.4% -9.7% 3.7% 7.6%

SR 140 @ River Exchange Dr EB Fast Tube 3.1% -2.1% 8.4% 4.9% 7.3%

SR 140 @ River Exchange Dr EB Slow Tube 2.3% 0.1% 4.4% 2.7% 3.5%

SR 140 @ River Exchange Dr EB All Tube 2.7% 0.7% 4.7% 9.7% 3.7%

SR 92 @ Woodlands Rd EB Fast Signal 7.4% 8.8% 6.1% 7.5% 10.0%

SR 92 @ Woodlands Rd EB Slow Signal -3.6% -2.8% -4.5% 3.9% 5.5%

SR 92 @ Woodlands Rd EB All Signal 2.0% 2.6% 1.5% 2.2% 3.3%

SR 92 @ Woodlands Rd WB Fast Signal -1.5% 0.0% -3.0% 4.1% 7.5%

SR 92 @ Woodlands Rd WB Slow Signal -1.1% 0.2% -2.3% 4.0% 6.2%

SR 92 @ Woodlands Rd WB All Signal -1.3% -0.2% -2.3% 2.3% 5.3%

SR 140 @ Steeplechase Rd EB Fast Signal -10.0% -6.2% -13.8% 11.1% 21.3%

SR 140 @ Steeplechase Rd EB Slow Signal -0.5% 0.7% -1.7% 4.1% 8.4%

SR 140 @ Steeplechase Rd EB All Signal -5.7% -3.2% -8.3% 7.6% 13.9%

SR 140 @ Steeplechase Rd WB Fast Signal 2.9% 3.9% 1.9% 3.7% 7.7%

SR 140 @ Steeplechase Rd WB Slow Signal -2.7% -1.9% -3.4% 3.1% 6.0%

SR 140 @ Steeplechase Rd WB All Signal 0.7% 1.4% 0.0% 2.4% 3.7%

SR141 @ E. Jones Bridge Rd* NB Fast Signal -23.3% -17.8% -28.8% 23.5% 36.0%

SR141 @ E. Jones Bridge Rd* NB Slow Signal -24.5% -15.4% -33.7% 24.6% 51.7%

SR141 @ E. Jones Bridge Rd* NB All Signal -6.2% -5.1% -7.3% 6.2% 8.2%

SR141 @ E. Jones Bridge Rd* SB Fast Signal 24.9% 30.6% 19.1% 30.0% 38.0%

SR141 @ E. Jones Bridge Rd* SB Slow Signal -27.2% -20.9% -33.5% 30.4% 41.4%

SR141 @ E. Jones Bridge Rd* SB All Signal -20.1% -15.3% -24.8% 21.6% 31.0%

SR 314 @ Banks Rd NB Fast Signal -1.4% 0.7% -3.6% 6.2% 15.4%

SR 314 @ Banks Rd NB Slow Signal -3.4% -2.1% -4.7% 4.3% 9.7%

SR 314 @ Banks Rd NB All Signal 0.5% 0.8% 0.2% 1.4% 2.1%

SR 314 @ Banks Rd SB Fast Signal 6.2% 8.1% 4.3% 7.3% 11.8%

SR 314 @ Banks Rd SB Slow Signal -2.0% -1.5% -2.6% 2.1% 3.5%

SR 314 @ Banks Rd SB All Signal -0.1% 0.4% -0.6% 1.6% 2.5%


Location Direction Lane Evaluated Count MPE

Lower 95%

Upper 95% MAPE RMSPE

SR 141 @ Old Alabama Rd NB All Signal -2.6% -1.6% -3.7% 2.1% 4.1%

SR 141 @ Old Alabama Rd SB All Signal 0.7% 1.4% 0.0% 2.1% 4.1%

SR 92@ Roswell Crossing EB Fast Signal 1.3% 2.3% 0.4% 4.2% 6.6%

SR 92@ Roswell Crossing EB Middle Signal -9.6% -8.4% -10.8% 9.7% 12.9%

SR 92@ Roswell Crossing EB Slow Signal 24.3% 27.9% 20.6% 24.3% 35.4%

SR 92@ Roswell Crossing EB All Signal 4.3% 5.3% 3.3% 5.0% 8.3%

SR 92@ Roswell Crossing EB Fast Signal D/S -2.1% -0.1% -4.1% 11.0% 13.9%

SR 92@ Roswell Crossing EB Middle Signal D/S -10.1% -7.5% -12.6% 15.3% 20.5%

SR 92@ Roswell Crossing EB Slow Signal D/S 21.7% 25.2% 18.2% 22.1% 32.6%

SR 92@ Roswell Crossing EB All Signal D/S 2.1% 3.2% 1.1% 4.9% 7.4%

SR 92@ Roswell Crossing WB Fast Signal -4.1% 0.0% -8.2% 9.3% 28.7%

SR 92@ Roswell Crossing WB Middle Signal 2.5% 3.2% 1.9% 4.0% 5.2%

SR 92@ Roswell Crossing WB Slow Signal 0.3% 0.8% -0.1% 2.1% 3.3%

SR 92@ Roswell Crossing WB All Signal -0.5% 0.9% -2.0% 4.0% 10.0%

SR 141 @ SR 237 EB Slow Signal/VDS 11.8% 0.7% 22.9% 17.9% 19.9%

SR 141 @ SR 237 EB Middle Signal/VDS 9.6% 0.2% 18.9% 11.1% 18.5%

SR 141 @ SR 237 EB Fast Signal/VDS 2.6% 1.0% 4.2% 3.1% 3.8%

SR 141 @ SR 237 EB All Signal/VDS 6.1% 3.9% 8.3% 6.1% 7.1%

SR 92 @ Woodstock WB Fast Tube 4.5% 2.6% 6.4% 4.5% 5.5%

SR 92 @ Woodstock WB Middle Tube -6.0% -8.4% -3.7% 6.5% 7.1%

SR 92 @ Woodstock WB Slow Tube 8.7% 4.3% 13.2% 8.9% 11.3%

SR 92 @ Woodstock WB All Tube 3.0% 1.8% 4.2% 3.0% 3.5%

SR 92 @ Woodstock EB Fast Signal 12.9% 19.2% 6.6% 19.1% 46.0%

SR 92 @ Woodstock EB Middle Signal 58.1% 72.7% 43.6% 58.5% 118.0%

SR 92 @ Woodstock EB Slow Signal -8.0% -3.9% -12.1% 20.5% 30.0%

SR 92 @ Woodstock EB All Signal 19.6% 25.1% 14.0% 19.6% 21.8%

SR 92@ Bowen Road EB Fast Signal 6.4% 8.9% 3.9% 12.6% 18.7%

SR 92@ Bowen Road EB Middle Signal 19.1% 22.0% 16.2% 20.0% 27.9%

SR 92@ Bowen Road EB Slow Signal 4.4% 6.6% 2.1% 11.1% 16.3%

SR 92@ Bowen Road EB All Signal 11.9% 14.0% 9.8% 13.4% 18.9%

1

* Possibly faulty baseline counts 2

3


1

2 Figure 3: Y-Y Plots for Manual Counts (X-axis) Vs Tube Counts (Y - Axis) at Intersection of SR 92 3 WB and Roswell Crossing for Lane 1 (Top-Left), Lane 2 (Top-Right), Lane 3 (Bottom-Left) and 4 Aggregates over All Lanes (Bottom Right) 5

6

Effect of Presence of Turn Lanes and Lane Changes 7

A potential impact of the presence of high turning volumes (usually evidenced by the presence 8 of turn lanes or bays) is the associated high volume of lane changing. If a significant portion of 9 the lane changes occur right over the detection zone, it could lead to faulty counts (over- or 10 under-counts depending on the wheel path). 11

The absence of turn lanes did not show a direct relationship with the degree of match between 12 the signal count data with the baseline data. SR 140 at Steeple Chase Drive was the site with no 13 turn lanes and showed a -5.7% error on eastbound lanes, and 0.7% on the westbound lanes. At 14 the SR141 site at Old Alabama Road, there is a significant amount of traffic on the turn lane, 15 but when the volumes on the through lanes are considered, the error levels are low (-2.6% NB 16 and 0.7% SB). However it was observed that a significant amount (close to 100%) of the 17

050

100150200250300350400450500

0 50 100

150

200

250

300

350

400

450

500 0

50100150200250300350400450500

0 50 100

150

200

250

300

350

400

450

500

050

100150200250300350400450500

0 50 100

150

200

250

300

350

400

450

500 0

100200300400500600700800900

1000

010

020

030

040

050

060

070

080

090

010

00


turning traffic was missed by the intersection signal detectors. While the signal detectors counts 1 aggregated over the through lanes match very well with the counts on those two lanes from the 2 pneumatic tube counts, if the counts from the signal detectors on the two through lanes were 3 used to represent the entire volume on that section of the road, the counts will be low by more 4 than 15%. 5

At the study site on SR 92 between Woodlands Parkway and Dials Drive, some amount of lane 6 changing was observed between the Pneumatic Tube data collection site and the intersection 7 signal detector’s detection zone. There is a slight positive bias of the signal detector’s count, as 8 compared to the tube count, observed in the Lane 1 counts while there is a slight negative bias 9 observed in the Lane 2 counts. However, the total counts match quite well (-1.3% - 2.0%). 10 While lane changes produce some differences in the counts between the validation stations and 11 the signal detectors, since the counts match when aggregated over stations, there is not 12 sufficient evidence to conclude that lane changing produces a significant effect on the error 13 levels. 14

15

Effect of Volume 16

Intersections without Queuing 17

At the low volume site, the errors were typically low, under 6.4%. At high volume sites where 18 there was no observed queueing, the errors were typically low as well. For example the mean 19 percentage error was under 4.3% in either direction when aggregated across lanes at the SR 92 20 intersection at Roswell Crossing site. However the results at another high volume site, SR 140 21 at River Exchange Parkway, showed some undercounting under high volume conditions in a 22 few instances (Figure 4a). But the overall aggregate mean percentage error was in a range of -23 9.7% to 4.4%. 24

Intersections with Queuing 25

The spikes in the data under congested conditions that can be observed in Figure 4 are observed 26 in all three lanes at this site. Figure 5 shows a three way comparison of manual counts, 27 pneumatic tube counts, and signal detector counts and shows that the errors in the signal 28 detector counts here are much larger than the error bounds of the pneumatic tube count data. 29

If a per-vehicle-record based data stream is available, it might be possible to design a filter to 30 completely eliminate the erroneous volumes, but it is difficult to eliminate such faulty data from 31 the 15-minute aggregates. 32

Several other high volumes sites such as SR 92 at River Exchange Parkway were also studied 33 and volumes above 250 vehicles per 15-minute period per lane were observed, but evidence of 34 these spikes in volumes were not seen. It is hypothesized that the erroneous volumes are a 35 result of queuing over the induction loop detectors and compounded by some site specific 36 calibration issues. 37

Effect of Detection Technology 38

Most of the intersections in the study region had induction loop based detection. There are a 39 limited number of intersections with Video Detection System based detection. Typically, 40 intrusive technologies such as induction loops have a higher level of accuracy of vehicle count 41 data than video detection systems that are easily affected by external factors such as shadows, 42 precipitation, wind, and vehicle induced vibration of the camera (7,11,12,13). There are studies 43


that have documented sub-optimal performances of VDS for stop bar detection (7). In this 1 study, the accuracy of counts from a video detection system installed at the intersection of SR 2 237 and SR 141 was tested against manual counts. The manual counts were extracted using the 3 same video feed that was used by the video detection system. The camera is installed on the 4 signal mast across the intersection. The detection zones are set up past the stop bar, which 5 ensures the vehicles are moving (except in case of a gridlock) when they cross the detection 6 zone, and thereby provides more accurate detection rates. 7

The accuracy test had two major limitations: 8

• Given the resource intensive nature of the manual data reduction of the videos that were 9 of low quality, the test was limited to a single approach (SR 141 EB) and a single 10 morning peak period. 11

• The only adverse condition that was tested was the transition in lighting conditions 12 (dawn). The other condition that might have been tested is camera vibration due to 13 heavy vehicle traffic, but any movement of the camera was not noticeable in the 14 observed video. Effects of precipitation, high wind, shadows etc. were not tested in this 15 study. 16

The video detection system was configured to detect vehicle counts on only three of the lanes. 17 The left turning vehicles were not counted and are therefore unaccounted for. Discounting the 18 left turn lanes, the mean percentage errors on the other three lanes were in the ranged from 2.6 19 to 11.8%. 20


1

2 Figure 4: Time Series Plot for Pneumatic Tube Counts and Signal Detector Counts at: 3

(a) Top: SR 140 EB at River Exchange Parkway for Aggregates over All Lanes 4

(b) Bottom: SR 92 EB at King Road for Lane 1 5

6

0

100

200

300

400

500

600

700

800

900

5/5/

2014

12:00

5/5/

2014

16:48

5/5/

2014

21:36

5/6/

2014

2:24

5/6/

2014

7:12

5/6/

2014

12:00

5/6/

2014

16:48

5/6/

2014

21:36

5/7/

2014

2:24

5/7/

2014

7:12

5/7/

2014

12:00

Tube Count EB All LnSignal Count EB All Ln

0

200

400

600

800

1000

1200

1400

10/2

9/20

13 12

:00

10/2

9/20

13 18

:00

10/3

0/20

13 0:

00

10/3

0/20

13 6:

00

10/3

0/20

13 12

:00

10/3

0/20

13 18

:00

10/3

1/20

13 0:

00

10/3

1/20

13 6:

00

10/3

1/20

13 12

:00

Tube + LTL92 @ King WB1


1 Figure 5: Time Series Plot for Manual Counts, Tube Counts and Signal Detector Counts at 2 Intersection of SR 92 EB and King Road for Aggregates over All Lanes 3

4

Summary of Results 5

The following are the major observations: 6

• The pneumatic tube counts, that were used as the baseline for the rest of the 7 comparisons, were within 10% of the manual counts at lane-by-lane level comparison of 8 15-minute aggregates. The manual counts were treated as ground truth in this study. 9

• Intersection signal detector counts matched closely with the tube counts, with the 95% 10 confidence intervals of the mean percentage errors within 10%. The errors increased 11 significantly under high volume conditions where is it likely that the vehicle queues 12 extended over the induction loop detection zone, causing over-counting under these 13 conditions. 14

• One of the sites, SR 141 at East Jones Bridge, showed high differences between the 15 intersection signal counts and the pneumatic tube counts. Comparison with the upstream 16 and downstream detectors indicated that the adjacent intersection signal counts matched 17 quite well between themselves but were all different from the pneumatic tube counts. 18 Since the validity of the pneumatic tube counts at this site could not be verified after the 19 fact, conclusions were not drawn based on data at this site. 20

• Higher level of variability in agreement was observed in the lane level counts than in the 21 aggregate over-all-lanes counts. Part of this difference may be attributed to possible 22 lane changes between the spatially separated detection zones of the pneumatic tube 23 counters and the intersection signal detectors. 24

0

500

1000

1500

2000

2500

10/2

9/20

13 16

:30

10/2

9/20

13 16

:45

10/2

9/20

13 17

:00

10/2

9/20

13 17

:15

10/2

9/20

13 17

:30

10/2

9/20

13 17

:45

10/2

9/20

13 18

:00

10/2

9/20

13 18

:15

10/2

9/20

13 18

:30

10/2

9/20

13 18

:45

10/2

9/20

13 19

:00

10/2

9/20

13 19

:15

10/2

9/20

13 19

:30

Tube WB All LnSignal WB All LnManual WB All Ln


• It was observed that in several cases, the detection zones of the intersection signal 1 detectors were downstream of the beginning of the turn lanes. It was confirmed in the 2 comparisons that this causes severe undercounting in these scenarios because the turning 3 vehicles are not accounted for. 4

• Data from intersections using video based detection are of comparable quality as other 5 intersections using inductive loop detectors. Like other technologies, video based 6 detection is sensitive to calibration. At the video detection test site two out of three 7 lanes reported data with less than 10% errors (MPE) while one of the lanes reported a 8 mean percentage error of 12%. While it was not verified, it is likely that the detection 9 accuracy could be improved by better placement of the detection zone in the camera 10 view. 11

12

Ideal Intersection for Data Use 13

Based on the above findings the following are the characteristics of a site that would be ideal for 14 use in traffic monitoring: 15

• Detector data does not display any sudden spikes and there is no known occurrence of 16 queuing of vehicles over the detector. Here a spike is defined as the existence of a few 17 data points within a series that deviate significantly from the other data points. At least 18 one week of data should be used to generate time series plots of the data to visually 19 check for data spikes. 20

• The signal cabinet communicates with the central server via a stable wired connection. 21 Wireless connections have been found to be unreliable, and may not guarantee 22 completeness of data. 23

• Data plots agree with expected increases and decreases in traffic demand during peak 24 and off peak travel periods. At least one typical week of data should be used to generate 25 time series plots of the data to visually check for expected data patterns. Here a typical 26 week is defined as a week that does not have any holidays, weather events etc. and is not 27 adjacent to long weekends. 28

• Data is available from two adjacent intersections. Data from signal detector downstream 29 of midblock location will be used for traffic monitoring. 30

• The intersection signal’s inductive loop detectors are physically located upstream of the 31 beginning of turn lanes (if any) (see Figure 6) 32


1

Figure 6: Proposed Detection Zone Placement Upstream of Beginning of Turn Lanes 2 (Image Courtesy: Google Maps)Recommendations 3

4

CONCLUSIONS AND RECOMMENDATIONS 5

Findings 6

The objective of this study was to perform an evaluation of the feasibility of integrating 7 intersection traffic signal data into a traffic monitoring program. About six hundred and eighty 8 hours of directional data across nine study sites around Metro Atlanta was analyzed at a lane-9 by-lane level in this study. 10

The comparison results in this study indicate that in majority of the cases, the intersection signal 11 detector data is similar in quality to pneumatic tube count data in terms of both the mean and 12 variability of the errors. While there is variability in the level of error from site to site, in 13 general following the intersection eligibility criteria guidelines should help ensure high quality 14 of data for use in the traffic monitoring program. If the data from a particular site is expected to 15 be used extensively on a long term basis, validation of the data via short term counts is 16 recommended. 17

The intersection signal detectors do not provide vehicle classification information. Therefore 18 the intersection signal detectors cannot be expected to replace the permanent traffic counting 19 stations used in the traffic monitoring program. However, these detectors provide continuous 20 data throughout the year and can possibly be used for other purposes such as determining 21 seasonality or diurnal factors. 22

To improve the usability of intersection signal detector data for traffic monitoring it is 23 recommended that for future installations and maintenance on existing detectors, detection 24 zones are moved further upstream beyond the maximum queue length of a typical peak hour 25 queue and beyond the beginning of the turn lanes. If it is not possible to move detections zones 26 upstream of turn lanes, detectors should be installed on turn lanes as well. 27

Beginnning of turn lane

Existing Detection Zone

Proposed Detection Zone


The results of this study indicate that the quality of data varies from site to site. Deployment 1 characteristics, quality of calibration, communication faults, location of loops with respect to the 2 beginning of turn lanes, presence of queues extending over the detection zone etc. are some of 3 the factors that affect the quality of the data as it pertains to the use of this data in the traffic 4 monitoring program. However several of these factors are either theoretically difficult to 5 quantify or is too resource intensive to measure for a large number of sites. 6

From an organization perspective, the intersection signal detectors would need an initial 7 validation and a software interface development would likely be needed to enable smooth 8 automated transfer of the data for use in a traffic monitoring program. The intersection signal 9 detectors could provide traffic data every day of the year, which would be useful to data 10 customers. However, if pneumatic tube counts need to be periodically collected and technicians 11 need to perform site visits to calibrate detectors, the cost value compared to the customer need 12 has to be considered. 13

Under these constraints, it would be helpful to have a mechanism for estimating, in lieu of 14 measuring, the quality of the data without performing extensive field measurements. Further 15 research needs to be conducted to develop a methodology for estimating data quality based on 16 readily available information, and without requiring additional data collection. It is envisioned 17 that a system for assigning "Quality Ratings" for signal detector data at each intersection with 18 offset detectors can be developed. The quality rating will be based on comparisons of data from 19 roads with similar volumes and demands. The comparison data can be obtained from nearby 20 permanent count stations, or data from signal detectors at nearby intersections. At the lane 21 level, comparisons can be done with data from adjacent lanes to detect anomalies. Rational rule 22 based checks such as variation of volumes with time of day, location of detection loops with 23 respect to turn lanes etc. can be used along with the data comparison statistics to determine the 24 quality ratings. The quality ratings should be available at the approach level at a minimum. 25 The quality ratings will provide a ready reference for choosing intersection approaches with 26 acceptable data quality for use in the traffic monitoring program. In addition, this methodology 27 provides a scalable architecture where new installations can be easily integrated in the list of 28 rated intersections, without requiring additional short-term data collection for every new 29 installation site. Once the base data patterns are established, this framework can be leveraged 30 for monitoring data for real-time operations as well, by developing correlation factors and 31 ranges of expected variation that will help identify deviation of data at individual detectors 32 which could be linked to potential degradation in data quality. 33

34

ACKNOWLEDGEMENTS 35

This work was sponsored by the Georgia Department of Transportation research project 13-10 36 Integrating Intersection Traffic Signal Data into a Traffic Monitoring Program. The findings and 37 conclusions presented herein represent the opinion of the authors and not necessarily that of the 38 Georgia Department of Transportation. 39

40

REFERENCES 41

1. Sherry L. Skszek, “State-of-the-Art” Report on Non-Traditional Traffic Counting 42 Methods, FHWA-AZ-01-503, October 2001. 43


2. NIT Phase II Evaluation of Non-Intrusive Technologies for Traffic Detection, Final 1 Report, Minnesota Department of Transportation, St. Paul, MN, September 2002. 2 http://www.dot.state.mn.us/guidestar/2001_2005/nit2/finalreport.pdf. Accessed July 31, 3 2015. 4

3. Field Test of Monitoring of Urban Vehicle Operations Using Non-intrusive 5 Technologies Final Report, Minnesota Department of Transportation, SRF Consulting 6 Group, Inc., MN, May 1997 http://ntl.bts.gov/lib/jpodocs/repts_te/6665.pdf. Accessed 7 July 31, 2015. 8

4. Middleton D., Jasek D., and Parker R. Evaluation of Some Existing Technologies for 9 Vehicle Detection. Project Summary Report 1715-S. Texas Transportation Institute, 10 September 1999. 11

5. Zheng, J.; X. Ma, Y. Wang, and P. Yi. Measuring Signalized Intersection Performance 12 in Real Time with Traffic Sensors. Transportation Research Board 88th Annual Meeting, 13 Transportation Research Board, 2009, 18p 14

6. Smaglik, E., A. Sharma, D. Bullock, J. Sturdevant, and G. Duncan. Event-Based Data 15 Collection for Generating Actuated Controller Performance Measures. Transportation 16 Research Record: Journal of the Transportation Research Board 2007 2035:, 97-106 17

7. Smaglik, E.J.; S. Vanjari, V. Totten, E. Rusli, M. Ndoye, A. Jacobs, D. Bullock, and J. 18 Krogmeier. Performance of Modern Stop Bar Loop Count Detectors over Various 19 Traffic Regimes. Transportation Research Board 86th Annual Meeting, Transportation 20 Research Board, 2007, 20p 21

8. Nathan A. Weber, Verification of Radar Vehicle Detection Equipment, Report SD98-15-22 F, March 1999 23

9. McGowen, P. and Sanderson, M. Accuracy of Pneumatic Road Tubes. Institute of 24 Transportation Engineers Western District Annual Meeting. Anchorage, AK. July 2011. 25

10. Using Existing Loops at Signalized Intersections for Traffic Counts: An ITE 26 Informational Report, ITE Journal, Vol: 78, Issue: 2, February 2008 27

11. Rhodes, A., D. Bullock, J. Sturdevant, Z. Clark, and D. Candey. Evaluation of the 28 Accuracy of Stop Bar Video Vehicle Detection at Signalized Intersections. 29 Transportation Research Record: Journal of the Transportation Research Board, Issue 30 1925, 2005, pp 134-145 31

12. Medina, J.C., R. Benekohal, and M. Chitturi. Changes in Video Detection Performance 32 at Signalized Intersections Under Different Illumination Conditions. Transportation 33 Research Record: Journal of the Transportation Research Board, Issue 2129, 2009, pp 34 111-120 35

13. Bonneson, J., and M. Abbas. Video Detection for Intersection and Interchange Control 36 FHWA/TX-03/4285-1, tti.tamu.edu/documents/4285-1.pdf. Accessed July 31, 2015. 37

14. Toth, C., W. Suh, V. Elango, R. Sadana, A. Guin, M. Hunter, and R. Guensler. Tablet-38 Based Traffic Counting Application Designed to Minimize Human Error. Transportation 39 Research Record: Journal of the Transportation Research Board, Issue 2339, 2013, pp 40 39–46 41

Intersection Signal Data for Traffic Monitoring 20150801 final docxdocs.trb.org/prp/16-4355.pdf ·...

Documents

Transcript of Intersection Signal Data for Traffic Monitoring 20150801 final docxdocs.trb.org/prp/16-4355.pdf ·...