Validating Dynamic Message Sign Freeway Travel Time Messages ...

Validating Dynamic Message Sign Freeway Travel Time Messages Using Ground Truth Geospatial Data Christopher M. Monsere Department of Civil & Environmental Engineering Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-9746, Fax: 503-725-5950 E-mail: [email protected] Aaron Breakstone School of Urban Studies and Planning Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-4285, Fax: 503-725-5950 E-mail: [email protected] Robert L. Bertini Department of Civil & Environmental Engineering Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-4249, Fax: 503-725-5950 E-mail: [email protected] Dean Deeter Castle Rock Consultants 6222 SW Virginia Avenue, Suite 2 Portland, OR 97239-3618 Phone: 503-892-2598, Fax: 503-892-2603 E-mail: [email protected] Galen McGill Oregon Department of Transportation 355 Capitol Street NE, 5th Floor Salem, OR 97301-3871 Phone: 503-986-4486, Fax: 503-986-4063 E-mail: [email protected] Submitted for presentation and publication to the 85th Annual Meeting of the Transportation Research Board January 22–26, 2006 Word Count = 7,255 [5,255 words + 2 Tables and 6 Figures]

Monsere, Breakstone, Bertini, Deeter, and McGill

1

ABSTRACT U.S. transportation agencies have invested more than $300 million in dynamic message sign (DMS) systems for communicating important messages to travelers, including weather conditions, incidents, construction, homeland security and AMBER alerts. The Federal Highway Administration (FHWA) is encouraging states to use their DMS infrastructure more effectively by displaying reliable travel time information along freeway corridors. The Oregon Department of Transportation (ODOT) maintains a traveler information system in the Portland metropolitan area and currently calculates and reports travel time estimates from inductive loop detectors. These estimates can be reported via 18 DMS, and are currently displayed on a limited basis at key junction points during peak periods. ODOT has also supported the development of the Regional Transportation Archive Listing (PORTAL), a transportation data archive that was designed to automatically produce travel time estimates using historical loop detector data. The objective of this paper is to describe the results of an evaluation of ODOT’s travel time reporting capabilities by comparing the ODOT travel time estimates with probe vehicle (ground truth) travel times. The evaluation included the development of a statistical bound on the validity of the current travel time algorithm. Loop detector density and location were found to be critical to the accuracy of ODOT’s estimates. Travel times on a majority of links were found to be reasonably accurate, yet the potential for considerable improvement exists.


2

INTRODUCTION Advanced Traveler Information Systems (ATIS) are a critical component of regional intelligent transportation systems (ITS). They have numerous applications and are deployed in an effort to improve the efficiency and safety of transportation systems. ATIS can include systems focusing on pre-trip and en-route, and can involve numerous delivery systems, including a desktop PC, an in-vehicle display, a handheld device, highway advisory radio and dynamic message signs (DMS). U.S. transportation agencies have invested more than $300 million in DMS systems for communicating important messages to travelers, including weather conditions, incidents, construction, homeland security and AMBER alerts (1). The Federal Highway Administration (FHWA) is encouraging states to use their DMS infrastructure more effectively by displaying reliable travel time information along freeway corridors. Such an application of ATIS is aimed at providing the public with estimates of current travel conditions between strategic points and along critical freeway segments and arterials. Real-time travel time estimates provided by most ATIS are focused on freeway networks and rely on existing surveillance infrastructure. Real-time estimates can be made from data gathered by a variety of technologies such as inductive loop detectors, microwave radar, automatic vehicle tag matching, video detection, license plate matching, cell phone matching, and other techniques (2). Travel conditions (current and/or forecasted) can then be processed and reported to the public in the form of average speeds, travel times, images (video or still) via internet, radio, 511 traveler information systems, or DMS. A July 2004 memorandum by the FHWA has encouraged traffic management agencies to display travel time messages on existing DMS (3). Data cited from the ITS deployment tracking database indicate that approximately 12 metropolitan areas were providing travel time messages on DMS in 2002 but another 25 other metropolitan areas have the capability and available infrastructure (1). There is a strong desire to obtain a greater return on investments in DMS by using them to display automatically generated travel times in major cities and routes with recurring congestion. This would also provide the public with valuable information and aid in improved route decision making. However, many traffic management agencies have limited confidence in data reported by surveillance systems and may be reluctant to display these travel time messages to the public (3). Most surveillance systems on these freeway networks measure traffic conditions at a point location and are typically inductive loop sensors. Past work has shown that there is a certain degree of error inherent to the process of automatically generating travel time estimates from loop detector data, especially under congested or incident conditions (4,5,6). Difficulties can arise when communications are disrupted, individual detectors or stations periodically malfunction, report inaccurate data, or are placed in a way that masks actual traffic conditions on key links or provides limited coverage of the freeway system. Various algorithms have been proposed or used to estimate travel times using inductive loops (e.g., 5,6,7,8,9,10,11). When using inductive loop data for incident detection, for example, control over errors is important in order to avoid false alarms. However, in the arena of traffic information, research has shown that error rates of up 20% can still produce useful travel estimates for the public (12). Despite this, it is important to ensure that travel time estimates displayed to the public are reliable so that travelers do not lose confidence in the overall information system. The Oregon Department of Transportation (ODOT), together with regional partners, has developed an extensive advanced traffic management system (ATMS). A wealth of traveler information is currently available but ODOT is experimenting with providing real-time traveler information en-route via DMS as well as enhancing 511 and internet traveler information. The


3

objective of this paper is to describe the results of an evaluation of ODOT’s travel time reporting capabilities by comparing the ODOT travel time estimates with “ground truth” travel times collected by probe vehicles equipped with global positioning system (GPS) receivers. STUDY AREA The Portland region ATMS consists of a Traffic Management Operations Center (TMOC), an extensive fiber optic communication system, and a freeway surveillance system with 485 inductive loop detectors, 135 ramp meters and 98 closed circuit television cameras, as of April 2005. There are currently 18 DMS on the freeway network which are shown in Figure 1. The primary data used in freeway management are from the loop detectors which generate count, occupancy and speed measures every 20 seconds. All loop detector stations are dual loop configuration placed in the mainline lanes and are located downstream of on-ramps in the freeway network. On-ramp count loops are also present. Figure 1 shows the lengths of segments covered by loop detectors — the delineations of each segment shown are dictated by the midpoint between adjacent detector stations. As shown, some segments are very long and the coverage of the entire freeway system is not yet complete. The ODOT ITS Plan for Portland calls for full detector coverage in the future. Currently, ODOT provides pre-trip travel conditions to the public by a variety of methods. The main dissemination occurs via the TripCheck website (www.tripcheck.com) which offers a wealth of traveler information. All 98 CCTV freeway surveillance cameras are available online as static images that are updated regularly. The updating cycle varies by camera, but most cameras update several times an hour. The moving video images are also available on local cable television in the morning peak hours. A recent enhancement to the system produces a thematic color-coded speed map that displays average speed for individual highway segments. This map is generated automatically and is updated every 20 seconds. Once en-route, travelers can obtain information via the 511 Traveler Information System or from a number of DMS within the Portland region. These DMS display a variety of messages pertaining to freeway conditions and three of them have been configured to report estimated travel times on key corridors. One issue with the reporting of travel times calculated at a particular instant to a driver entering a segment is that the travel time actually experienced by that driver may be shorter or longer if the onset of congestion, an incident, or the dissipation occurs while the driver is traveling over the segment. These transition periods are of great interest in an evaluation of freeway travel time estimation. The three VMS that currently report times are shown as diamonds while the others are shown as circles. Rather than report exact estimates of travel time, ODOT reports estimated travel times in ranges of 2 to 3 minute increments on key corridors. The first range is essentially the free-flow travel time for the link. When travel times substantially longer than free flow are estimated by the ATMS, the travel time is reported as exceeding some value (e.g. “15+ minutes”).


!

4

´

DMS Location

ÖDMS w/ Travel Time

§ ¦5 ¨ ! ! 5

4

!!

14

! 13

!

!

84 § ¦ ¨

17

!

16

!

!

26 t u

!

Ö

8 9

6 217

3

11

Ö

12

§ ¦ ¨

2 7

§ ¦5 ¨

Ö

!

205

10

0

1

2

3

4 mi

NAD 1983 State Plane HARN, Oregon North

FIGURE 1 Freeway links, detector influence areas, and DMS locations.


5

ARCHIVED ITS DATA In an effort to encourage the retention and subsequent utilization of ITS-generated data, the U.S. Department of Transportation (USDOT) revised the National ITS Architecture in 1999 to include the Archived Data User Service (ADUS). ADUS is designed to promote “the unambiguous interchange and reuse of data and information throughout all functional areas.” ADUS stipulates that ITS data be collected and archived for historical and secondary uses, as well as being made readily available (13). In cooperation with ODOT and other regional partners, PORTAL was inaugurated in July 2004 by way of a direct fiber-optic connection between ODOT and servers located at Portland State University (PSU). This data archive gathers and stores 20-second data provided by ODOT’s network of loop detectors. The travel time calculation component of the powerful PORTAL data archive (14) was customized especially for this project, was available at all stages of this analysis, and made this evaluation project possible. Travel time estimates are generated from speed measurements in individual lanes. The speed estimate for each detector station is an average speed weighted by count across all lanes measured by the inductive loop detectors at 20-second intervals. The loop stations are currently installed in dual loop (speed trap) configurations whereby vehicle speeds (and lengths) can be directly measured. However, at present ODOT’s system essentially treats the pair of loops as two single loop stations and uses an average vehicle length to calculate speed from measured occupancy. In the near future, there is a project that will modify the speed measurement system to take advantage of the speed trap. Point speeds are then extrapolated over an influence area. The detector influence areas are calculated by the traditional midpoint method. As shown in Figure 2, an influence area is assigned to an individual detector based on the locations of the midpoints between that detector and the next stations upstream and downstream. The boundaries of each influence area can be seen in Figure 1. For the purposes of this study, therefore, traffic conditions attributed to a particular detector are assumed to represent the conditions on the section of freeway extending from halfway to the next downstream station to halfway to the next upstream station. In the event that no ‘next’ station exists, the detector’s influence is extended to the end of the segment in question. Speed estimates from each detector station are converted to travel time estimates by dividing the influence area length by the weighted speed. This procedure has been automated within PORTAL, and travel time data are available for download at the raw 20-second level for the influence areas and links defined for this study.

FIGURE 2 Loop detector influence areas as defined by the midpoint method. DATA COLLECTION The objective of this project was to compare estimates of travel times made from inductive loop data to ground truth probe data on 15 identified segments of the Portland freeway network. These 15 segments (referred to as “links” hereafter) are shown in Figure 1 (the boundary of each link is distinguished by the gray and white shading). Each of the links is made up of the individual detector influence areas described in the previous section. The limits of these individual detector areas can be seen as the individual blocks in Figure 1. Due to concerns about transition periods mentioned above, one goal of the project was to attempt to collect data during the transitions


6

from freely-flowing to congested, and from congested to freely-flowing traffic conditions whenever feasible. In order to minimize costs associated with the data collection effort, an analysis was undertaken to determine the minimum number of probe vehicle runs required on each freeway link that would result in statistically significant results when compared to the archived loop data. The statistical estimation of the sample size n is typically based on specifying levels of confidence and acceptable error. For this evaluation, estimates of the mean travel speed (analogous to travel time) and standard deviation were calculated using the PORTAL archive using one week’s data (April 4 - April 8, 2005). This was done for one hour in the morning (7:30-8:30) and afternoon (16:30-17:30) peak periods. The allowable error recommended by the Institute of Transportation Engineers (ITE) for travel time studies is between 3 and 5 mi/h (15). It was determined that most links required a minimum of 5 runs with a few requiring near 10. For links where the analysis required additional runs, it was decided that the number of needed runs would be recalculated following the collection of probe vehicle samples. The PORTAL archived data was also used to determine the appropriate peak period and approximate departure times for the probe vehicle platoons to best capture changing traffic conditions. Four loop routes were designed to collect the required number of probe vehicle samples on the freeway network while minimizing wasted travel. In accordance with the FHWA’s Travel Time Data Collection Handbook (16), drivers would be instructed to follow standard probe vehicle instructions such that the data would represent the behavior of a “typical” driver. In most cases, probe vehicles were separated by 5-7 minute headways but on some links only a single probe was used, resulting in longer headways between samples. Probe Vehicle Data Probe vehicles were deployed along routes to collect data using ITS-GPS, a custom data collection application designed for use with Palm handheld computers equipped with GPS devices (17). ITS-GPS (available free at www.its.pdx.edu) continuously records the user’s position at 3-second intervals as well as calculating speed and distance traveled between readings. All of the data collected by probe vehicles were exported to local desktop computers and prepared for analysis. Each probe vehicle data collection run was imported into a geographic information system (GIS) containing a map of the ODOT freeway network. As with any GPS data collection procedure, any data that were corrupted or contained location errors were discarded. The remaining data files were appended to add a unique identification number to each GPS record in each file. Each probe vehicle run was then plotted as a series of points on a map of the freeway network displaying the start and end of each link. The data point where the probe vehicle entered and exited the link was identified by visual inspection of the plotted data and the entering time was noted. A representation of this technique is shown in Figure 3 for the interchange between OR-217 and US-26 where the probe was traveling from Link 17 to Link 9. The shaded gray area is the influence area for the start of the next westbound link. In this manner, the probe vehicle travel times were matched exactly to the freeway links and were then extracted and placed in a separate database for further analysis.


7

First point on Link 9

Last point on Link 17

FIGURE 3 Assigning probe vehicle paths to freeway sections. Travel Time Estimates from ITS Archived Data The PORTAL data archive was configured to generate the travel time estimates from the historical archived data using the same algorithm that is currently used by ODOT in real time. In this manner, probe vehicle travel times on each link could be compared to the estimate that would have been displayed to the public. PORTAL was configured to output travel time estimates either for each detector influence area, or for the identified link in the evaluation (a set of influence areas). In this evaluation, the aggregated link estimates are evaluated. The time that the probe vehicle entered the link was matched to the travel time estimates for the nearest 20-second PORTAL data archive interval. In all cases, the matching record time was matched to the closest previous 20 second interval. The time from the probe vehicle records was obtained from the GPS satellites, while the time in the ODOT ITS data archive is based on the ATMS system which updates the Type 170 ramp meter controller clocks once every 24 hours using the time from an internet web site such as www.time.gov. The error between these sources is unknown, likely to be small, and was not considered when matching records. In addition, the travel time estimate for each 20-second interval up to 3 minutes prior to the probe vehicle’s entry was also included for analysis. This was included in order to enable subsequent assessment of whether inclusion of short-term historical data could improve the accuracy of the travel time estimates generated by the ATMS. DATA ANALYSIS A total of 87 probe vehicle runs totaling 516 miles by 12 drivers over 15 hours on Wednesday (70 runs), Thursday (7 runs), and Friday (10 runs) were collected and analyzed. Some problems were encountered with the data collection software which resulted in fewer probe runs being available than planned in the experimental design. Link 11 was dropped from further analysis because only 1 probe run was valid and the project schedule did not allow for additional data


8

collection. The data collected includes both the planned data collection routes with probe vehicles at 4-5 minute headways and a number of individual runs that were collected as preliminary data. On most links, the probe data were only collected for the peak direction and time of travel. A few links (6, 13, and 16) had both morning and afternoon peaks collected. The probe vehicles encountered a variety of traffic conditions including free flow, recurring congestion, and two notable incidents. On Link 8, all probe vehicles on May 4, 2005 encountered an incident on northbound OR-217 that was the cause of the unusually long travel times on that day. A multi-vehicle crash was reported, but there were no longer any lane blockages by the time the probe vehicles arrived. Only one of the vehicles involved remained, and it had been moved to the shoulder. Further, all probe vehicles on Link 17 also encountered an incident on westbound US-26 just west of downtown Portland. The nature of the incident was unclear – a probe vehicle observed two vehicles and an ambulance stopped in the left-hand lane, but there was no sign of a crash. For analysis purposes, all probe runs for each link were pooled and summary statistics are shown in Table 1. The equivalent average speed is shown for reference by dividing the link distance by the average probe travel time. This can be interpreted as an indicator of the amount of congestion encountered by the probe vehicles on each link. To further verify traffic conditions encountered by the probe vehicles, time-space plots for each of the 87 probe runs were created for visual representation of the probe trajectories. These plots, shown in Figure 4, have time on the x-axis and distance (milepost) on the y-axis. The probe trajectories are shown as the dark solid lines. These plots clearly show when the probe vehicles were traveling in congested or free flow conditions. The plots also show the estimated trajectories (i.e. travel times) for the nearest 20-second interval when the probe vehicle entered the link and 1, 2, and 3 minutes on either side of the matched time. Inspections of these plots were used to characterize much of the traffic conditions, explain the analysis results, and make recommendations for additional detection. TABLE 1 Probe Vehicle Data Summary

Link Hwy 2 I-5 N 3 I-5 N 4 I-5 N 5 I-5 S 6 I-5 S 7 I-5 S 8 OR-217 N 9 OR-217 S 10 I-205 N 12 I-205 S 13 I-84 E 14 I-84 W 16 US-26 E 17 US-26 W

Limits I-205 to OR-217 OR-217 to I-405 I-84 to Int. Bridge Int. Bridge to Going SB I-405 to OR-217 OR-217 to I-205 I-5 to US-26 US-26 to I-5 I-5 to Oregon City I-84 to Oregon City I-5 to I-205 I-205 to I-5 OR-217 to I-405 I-405 to OR-217

Milepost MileFrom post To 287.85 291.78 291.78 300.40 300.40 308.38 308.38 303.50 300.71 292.30 290.82 287.66 7.20 0.26 0.11 6.89 3.22 10.25 19.79 2.48 0.31 3.69 4.17 1.21 68.93 73.32 74.08 69.70

Average Probe Time (mins) 3.89 9.77 17.98 7.49 11.33 4.16 21.41 7.61 14.39 18.46 5.50 7.24 9.78 14.88

Standard Dev. Of Equivalent Probe Average Time Speed Number (mins) (mph) of Runs 8 0.27 61 6 0.55 53 5 6.56 27 4 2.83 39 9 2.26 45 5 1.50 46 7 11.49 19 4 0.62 53 4 0.89 29 4 0.58 56 11 2.66 37 8 0.94 25 7 2.47 27 4 3.49 18

9

Time

Time

Run 3

8:47:02

8:45:36

8:44:10

8:48:29 9:27:22

9:25:55

9:24:29

9:23:02

9:21:36

9:20:10

9:18:43

9:17:17

9:15:50

308.0 307.5 307.0 306.5 306.0 305.5 305.0 304.5 304.0 303.5 9:14:24

Milepost 9:10:05

9:08:38

9:07:12

9:05:46

9:04:19

9:02:53

9:01:26

9:00:00

8:58:34

8:57:07

8:42:43

Run 2

308.0 307.5 307.0 306.5 306.0 305.5 305.0 304.5 304.0 303.5 8:55:41

8:41:17

Time

Run 1

Milepost

8:39:50

8:38:24

8:36:58

8:35:31

8:32:38

8:34:05

308.0 307.5 307.0 306.5 306.0 305.5 305.0 304.5 304.0 303.5

8:24:00

8:21:07

8:18:14

8:15:22

8:12:29

8:09:36

8:06:43

8:03:50

Milepost

308.0 307.5 307.0 306.5 306.0 305.5 305.0 304.5 304.0 303.5 8:00:58

Milepost


Time

Run 4

FIGURE 4 Sample link trajectory time space plots. For each freeway link, estimated travel times were compared to probe vehicle travel times for three aggregations. Both of the averaging methods generally approximate how the ODOT algorithm calculates and displays travel time on the DMS. The three estimates are: • Nearest 20 seconds – the estimated travel time at the nearest prior 20-second estimated travel time for the link from the PORTAL historical ITS data archive. • Average of previous 1 minute – an average of the travel times at the nearest 20-second interval and three prior 20-second estimates. • Average of 3 previous minutes – an average of the travel times at the nearest 20-second interval and six prior 20-second estimates. In Figure 5 (a), the observed probe vehicle time and the estimated travel time from PORTAL for the nearest 20-second interval are plotted for all 87 runs. The plot shows that probe and estimated and probe times are comparable for the majority of runs although there are a number of runs where the times differ substantially. Averaging historical estimates of both 1 minute prior shown in (b) and 3 minutes prior (c) appear to slightly improve the estimates by smoothing some of the outlying errors in estimated travel times. These techniques are tested statistically in the following section.

10

35

35

30

30

25

25

Estimated Travel Time



20 15 10 5

20 15 10 5

0

0 0

5

10

15

20

25

30

35

Obs e rve d Probe V e hicle Trave l Tim e

0

5

10

15

20

25

30

35


(a) Nearest 20 second estimate

(b) Average of previous 1 minute

35


30 25 20 15 10 5 0 0

5

10

15

20

25

30

35


(c) Average of 3 previous minutes FIGURE 5 Scatter plot of observed versus estimated travel times. Statistical Comparison Each of the travel time estimates was compared to the probe vehicle travel times for statistical validity. The paired t-test was used since the probe travel times and the estimated travel times can be considered paired samples. The power of this test is greater than a simple test of means. The paired t-test considers the hypothesis that the average of the differences between each paired sample is zero (18). The t-test statistic is calculated as: Xd t* = , (1) sd nd where: X d = average of sample differences; s d = standard deviation of sample differences; and

nd = sample size. Further, the confidence interval for the average difference can also be calculated using:


11

⎞ ⎛s ⎟ (2) X d ± tα 2 ⎜⎜ d nd ⎟⎠ ⎝ and the percent error between the average difference of the travel times and the average probe vehicle times is expressed as: Percent error = ⎛⎜ X d ⎞⎟ , (3) P⎠ ⎝ where: P = average probe travel time for the link. All of these calculations are presented in Table 2. The results of the paired t-test are displayed as p-values in Table 2. All tests and confidence intervals were calculated using a significance level of 95% (α = 0.05). If the p-value is less than the test level α then the null hypothesis of equal means is rejected. In this comparison, if the p-value is greater than 0.05 than the null hypothesis is accepted and the travel times compared can be considered equal. The smaller the p-value, the more contradictory the data is to Ho, meaning smaller p-values indicate less reliability in the statistical significance of the data. Comparisons where the travel time estimates between the probe and estimated travel times are statistically different are indicated in bold with an (*). Figure 5 presents the data in Table 2 graphically for each of the estimate methods.


12

TABLE 2 Percent Error, 95% Confidence of Error, and Paired T-Test P-Values

Link

Nearest 20 sec.

2

12%

3 4 5 6 7 8 9 10 12 13 14 16 17

I-5 NB: I-205 to 217 p-value I-5 NB: OR-217 to I-405 p-value I-5 NB: I-84 to Interstate Bridge p-value I-5 SB: Intrst. Bridge to Going St p-value I-5 SB: I-405 to 217 p-value I-5 SB: OR-217 to I-205 p-value OR-217 NB: I-5 to US-26 p-value OR-217 SB: US-26 to I-5 p-value I-205 NB: I-5 to Oregon City p-value I-205 SB: I-84 to Oregon City p-value I-84 EB: I-5 to I-205 p-value I-84 WB: I-205 to I-5 p-value US 26 EB: 217 to I-405 p-value US 26 WB: I-405 to 217 p-value

+/- 14% 0.073 4% +/- 8% 0.272 4% +/- 23% 0.651 7% +/- 27% 0.470 11% +/- 29% 0.435 -33% +/- 40% 0.009 * -22% +/- 16% 0.016 * -3% +/- 8% 0.381 7% +/- 33% 0.532 2% +/- 7% 0.537 3% +/- 9% 0.273 6% +/- 41% 0.736 20% +/- 16% 0.026 * -65% +/- 38% 0.013 *

Note: * indicates statistically significant at α =0.05

Percent Error Average previous 1 min 12% +/- 11% 0.034 * 5% +/- 9% 0.214 6% +/- 23% 0.545 13% +/- 56% 0.503 5% +/- 17% 0.676 -37% +/- 32% 0.021 * -16% +/- 11% 0.012 * -5% +/- 11% 0.199 21% +/- 50% 0.277 1% +/- 4% 0.478 8% +/- 14% 0.518 7% +/- 48% 0.754 24% +/- 16% 0.009 * -63% +/- 39% 0.014 *

Average previous 3 min 9% +/- 10% 0.071 8% +/- 13% 0.202 5% +/- 12% 0.287 14% +/- 31% 0.256 8% +/- 18% 0.605 -38% +/- 26% 0.016 * -12% +/- 10% 0.024 * -6% +/- 11% 0.159 25% +/- 23% 0.040 * 0% +/- 4% 0.731 9% +/- 12% 0.402 -9% +/- 32% 0.53 26% +/- 18% 0.012 * -62% +/- 36% 0.012 *

13

100%

100%

80%

80%

60%

60%

40%

40% Percent Error

Percent Error


20% 0% -20% -40%

20% 0% -20% -40% -60%

-60% -80%

-80%

-100%

-100%

2

3

4

5

6

7

8

9

10

12

13

14

16

17

2

3

4

5

6

7

8

9

10

12

13

14

16

17

Link Num be r

Link Num be r

(a) Nearest 20 seconds

(b) Average of previous 1 minute

100% 80% 60% Percent Error

40% 20% 0% -20% -40% -60% -80% -100% 2

3

4

5

6

7

8

9

10

12

13

14

16

17

Link Num be r

(c) Average of previous 3 minutes FIGURE 6 Percent error by link. Discussion The travel time estimates generated from the inductive loop sensors in Portland appear to perform reasonably well given the existing detector spacing and simple algorithm for predicting travel times. Many of the link travel estimates were within 20% of the probe vehicle observations, a value that FHWA research indicates is acceptable and can provide benefits to the traveling public. This is encouraging for ODOT since the placement of the inductive loop detector stations was not guided by the desire to produce travel time estimates. However, under some circumstances the existing configurations perform less accurately. One limitation of this evaluation is that, despite the best efforts and study design, many of the probe vehicle runs were in free flow conditions where inductive loop estimates could be expected to perform fairly well. In conditions where traffic congestion was forming or clearing, the link travel estimates were less accurate. However, there were not enough probe runs in congested conditions to make statistical comparisons. The limited evidence indicates that incidents produced the most error of the estimates. Two of the three links where the link estimates were not statistically similar, Link 8 and Link 17, were the result of most of the probe vehicles encountering an incident. This reinforces the concept that incidents are special events


14

and systems that automatically display travel time will have to include algorithms for incident detection (19). From a practical perspective, while the actual estimates were in error, if ODOT was to display messages with its current range policy motorists would still have been informed of the highest range in travel time. Also, with ODOT’s comprehensive CCTV surveillance system and the presence of incident management teams on the freeways, it is likely that incident information would also be displayed, providing more specific information to drivers on the affected links. Clearly, these methods used by ODOT to provide the travel time estimates could be improved. The estimates could be enhanced by a combination of improvements to the algorithms used and the physical detection infrastructure. Simple improvements such as averaging prior travel times based on PORTAL’s historical data resources do appear to measurably increase the accuracy of the travel time estimates but the pattern is not consistent for the samples collected in this evaluation. Advanced algorithms that include information about downstream detectors may improve the estimates without requiring any additional detection (2,6,9). If congestion forms after a vehicle enters a link, the methods used in the current ODOT configuration will not be sufficient. An algorithm with a predictive element (either based on historical day/time/week patterns or traffic flow theory) could increase the level of accuracy of these estimates. Although not described in this paper, the final report’s evaluation also included an investigation from the archived ITS data to identify locations of recurring congestion (20). It was apparent that some of the poorer performing links have detector locations that can not accurately measure congested travel. For example, Link 6 has a nearly 3-mile section assigned to one detector that is substantially influenced by a major merge of I-5, I-405, and OR-43. Additional detection placed downstream of the known recurring congestion would likely improve the estimate. Another approach may be to adjust detector influence areas based on observations of traffic patterns. ODOT has modified some of the detector influence areas (i.e. assigning more or less of freeway link to the speed estimated by each detector) to account for detector placement. This was studied but, in the interest of clarity, not presented here. The approach did not appear to have a major effect on accuracy of travel time estimates. CONCLUSIONS The primary objective of this project was to assess the performance of the ODOT ATMS in estimating freeway travel times based on the current algorithm and configuration of loop detectors, keeping in mind that the loop detectors were installed primarily to enable the ramp metering system. In general, based on an analysis of historical loop detector data and use of a probe vehicle field experiment, the travel time estimates were found to be reasonably accurate. In those instances where the ATMS failed to accurately estimate travel times, the discrepancies could usually be attributed to low detector density, suboptimal detector location, or the occurrence of incidents. One target result of this project was a list of recommendations of those locations where additional detectors would improve travel time predictions. A methodology was followed that resulted in identifying seven locations that would likely have a notable impact on improving travel time predictions. A comprehensive list of 13 key candidate locations for added detection was recommended, with 7 locations listed as priorities. This recommendation focused on providing improved travel time information during steady state congested conditions. This study was aimed at supporting the FHWA’s guidance toward expanding the provision of travel time to drivers using the existing DMS infrastructure. This is an excellent means of leveraging both the existing loop surveillance system and the installed DMS to provide


15

expanded information en-route to drivers. It is hoped that with this information motorists will be able to make better and safer routing decisions and may feel less stress if they know the length and type of travel. It is hoped that the methodology described here will be used in the future for further validation of travel time estimation procedures, and could be used by others to quantify the reliability of their travel time estimates. ACKNOWLEDGMENT The authors gratefully acknowledge the Oregon Department of Transportation, Traffic Engineering and Operations Section, Intelligent Transportation Systems Unit for sponsoring this evaluation. The authors would like to recognize Stacy Shetler, Dennis Mitchell, and Jack Marchant, who provided information, data, and review of the project. Steve Boice, Thareth Yin, Abram Van Elswyck, Casey Nolan, Erin Wilson, Zach Horowitz, Peter Bosa, David de la Houssaye, and Rodger Gutierrez drove the probe vehicles. Finally, this work was supported by the Department of Civil & Environmental Engineering in the Maseeh College of Engineering & Computer Science. REFERENCES 1. Federal Highway Administration, Information And Action Memorandum “Dynamic Message Sign (DMS) Recommended Practice and Guidance”, July 16, 2004, Jeffery Panaiti http://www.ops.fhwa.dot.gov/travelinfo/resources/cms_rept/travtime.htm#_ftn1 2. Cortes, C. E., R. Lavanya, J. S. Oh, and R. Jayakrishnan. General-Purpose Methodology for Estimating Link Travel Time with Multiple-Point Detection of Traffic. In: Transportation Research Record: Journal of the Transportation Research Board, No. 1802, TRB, National Research Council, Washington, D.C., 2002, pp. 181-189. 3. Meehan, B. and B. Rupert. Putting Travelers in the Know. Public Roads, Vol. 68, No. 3, November/December 2004, pp. 38-43. 4. El-Geneidy, A. and Bertini, R.L. Toward Validation of Freeway Loop Detector Speed Measurements Using Transit Probe Data. Proceedings of the IEEE 7th Annual Conference on Intelligent Transportation Systems, Washington, D.C., 2004. 5. Hellinga, B. Improving Freeway Speed Estimates from Single-Loop Detectors. Journal of Transportation Engineering, Vol. 128, No. 1, 2002, pp. 58-67. 6. Coifman, B. Improved Velocity Estimation Using Single Loop Detectors. Transportation Research: Part A, Vol. 35, No. 10, 2001, pp. 863-880. 7. Zhang, X. and J. A. Rice. Short-Term Travel Time Prediction. Transportation Research. Part C: Emerging Technologies, Vol. 11, No 3, 2003, pp. 187-210. 8. van Lint, J. W. C. and van der Zjipp, N. J. Improving a Travel Time Estimation Algorithm By Using Dual Loop Detectors. In Transportation Research Record: Journal of the Transportation Research Board, No. 1855, TRB, National Research Council, Washington, D.C., 2003, pp 41-48. 9. Coifman, B. Estimating Travel Times and Vehicle Trajectories on Freeways Using Dual Loop Detectors. Transportation Research. Part A: Policy and Practice, Vol. 36, No. 4., 2002, pp 351-364. 10. Eisele, W. and L. Rilett. Travel-Time Estimates Obtained From Intelligent Transportation Systems and Instrumented Test Vehicles: Statistical Comparison. In Transportation Research Record: Journal of the Transportation Board, No. 1804, TRB, National Research Council, Washington , D.C., 2002, pp. 8-16.


16

11. Hellinga, B. R. and L. Fu. Reducing Bias in Probe-Based Arterial Link Travel Time Estimates. Transportation Research. Part C: Emerging Technologies, Vol. 10, No. 4., 2002, pp. 257-273. 12. Toppen, A., S. Jung, V. Shah, and K. Wunderlich. Towards a Strategy for Cost Effective Deployment of Advanced Traveler Information Systems. In Transportation Research Record: Journal of the Transportation Research Board, No. 1889, TRB, National Research Council, Washington, D.C., 2004. pp. 27-34 13. Archived Data User Service (ADUS): An Addendum to the ITS Program Plan. FHWA, U.S. Department of Transportation, September 1998. 14. Bertini, R. L., S. Hansen, A. Byrd, and T. Yin. PORTAL: Experience Implementing the ITS Archived Data User Service in Portland, Oregon. In Transportation Research Record: Journal of the Transportation Research Board, TRB, National Research Council, Washington, D.C., 2004. (In Press). 15. Institute of Transportation Engineers, Manual of Traffic Engineering Studies, ITE, Washington, D.C., 2000. 16. Turner, S. M., W. L. Eisele, R. J. Benz, and D. J. Holdener. Travel Time Data Collection Handbook. Report FHWA-PL-98-035. FHWA, U.S. Department of Transportation, 1998. 17. Bertini, R. L., C. Monsere, A. Byrd, M. Rose, and T. El-Seoud. Using Custom Transportation Data Collection Software with Handheld Computers for Education, Research, and Practice. In Transportation Research Record: Journal of the Transportation Research Board, TRB, National Research Council, Washington, D.C., 2004. (In press). 18. Washington, S., M. Karlaftis, F. Mannering. Statistical and Econometric Methods for Transportation Data Analysis. Chapman and Hall/CRC, Boca Raton, Florida. 2003. 19. Hobieka, A. and S. Dhuilipala. Estimation of Travel Times on Urban Freeways Under Incident Conditions. In Transportation Research Record: Journal of the Transportation Research Board, No. 1887, TRB, National Research Council, Washington, D.C., 2004. pp. 97-106. 20. Deeter. D., C. Monsere, A. Breakstone, R.L. Bertini. Evaluation of Freeway Travel Time Estimates: Final Report. Oregon Department of Transportation, Salem, Oregon. August 2005.