MEASURING PERFORMANCE ON INTERRUPTED FLOW FACILITIES WITH GPS V2
1. MEASURING PERFORMANCE
ON INTERRUPTED FLOW
FACILITIES WITH GPS PROBE
AND BLUETOOTHTRAFFIC
MONITORING DATA
Reuben M. Juster, EIT
Stanley E.Young, P.E., Ph.D.
Elham Sharifi, Ph.D.
CATTworks
2. Vehicle Probes
• Alternative source of travel time
data
• Third party vendors aggregate
highway vehicles’ travel data
• Many different devices within or
embedded in vehicles transmit the
data
• Aggregated data is usually
cleaned to get one reading per
segment of roadway per time
period
• Data available to users through
web-interface / API
Car
Manufacturers
Fleet
Operators
Phone
Manufacturers
Third PartyVendor
Cleaning
Us
4. Not All Probe Data is Created Equal
• Probe data was first used for freeway-based applications
• Probe data users became interested in arterial-based
applications
• The I-95 Corridor CoalitionVehicle Probe Project’s (VPP)
validation program accessed the accuracy of the probe
data
• Freeway data is generally more accurate than arterial
data for several reasons
5. Fundamental Facility Differences
Freeways (Uninterrupted)
• High volumes
• Continuous Flow
Arterials (Interrupted)
• Lower volumes
• Interrupted flow
• Red lights
• Driveways
• Adjacent land uses
• Not all arterials data is created equal
• Vary by volume, signalized
intersections, driveways, geometry
• MobilityVs. Accessibility
• Which arterials can have probe data to
derive performance measurements?
Driveway
Intersection
Interrupted
Uninterrupted
6. VPPValidation
• Contract requires vendors to meet certain quality metrics
• This requires frequent validation studies on representative
corridors to ensure that data meets metrics
• For freeways these metrics include Average Absolute Speed
Error (AASE) and Speed Error Bias (SEB)
• These metrics work well for a uni-modal freeway travel time
distributions, but not multi-modal arterial travel time
distributions
Picture Sources: BTS, FHWA
7. AlternateValidation Method (1/2)
09/02 09/09 09/16 09/23 09/30
0
5
10
15
Date/Time
TravelTime-Minutes
Northbound
Traversals
Outliers
09/03/12 09/05/12 09/07/12 09/09/12 09/11/12 09/13/12 09/15/12 09/17/12 09/19/12 09/21/12 09/23/12 09/25/12 09/27/12 09/29/12
0
5
10
15
Travel Time Plot - US Route 1 NB - between Telegraph Road and Fairfax County Parkway
Date & Time
TravelTime(minutes)
Score > 25
BTM ^ VPP v
11. Example 1 Corridor Description
• US-1, Mercer County, New Jersey
(Princeton)
• 6-8 lanes total
• <1 Signal per mile, 3.2 miles long
• Grade separate interchanges
• Minimal access points
• Resembles a freeway
13. Example 2 Corridor Description
• US-130, Burlington County, New
Jersey
• 6 lanes total
• 2 Signals per mile, 1.5 miles long
• Multi-cycle signal failures
Signalized Intersection
Grade-separate interchange
15. Recommendations
Arterials likely to have
accurate probe data
Arterials possibly to have
accurate probe data
Arterials unlikely to to
have accurate probe
data
• AADT >40000
• 2+ lanes each
direction
• <= 1 signals per mile
• PrincipalArterials
• Limited Curb cuts
• Confidently
characterize
congestion and
performance
measures
• AADT 20K to 40K
• 2+ lanes each direction
• 1 to 2 signals per mile
• Minor Arterials (HPMS)
• Some segments work
(likely), others fail
(unlikely)
• No cycle failures
• Should be reviewed to
see effectiveness of
probe data
• Low volume,
AADT <20K
• >=2 signals per mile
• Major collectors
(HPMS)
• Probe data not
recommended
• Frequent cycles
failures
• Use re-identification
for performance
monitoring
16. Future
• Probe data will improve with:
• Larger sample sizes
• Better processing (point pairing as opposed to instantaneous)
• Improved segmentation (already happening)
• Arterials that previously did not have accurate probe data
may have accurate probe data (check every 18 to 24
months)
• In the mean time, verify validity if unknown
• Use the whole spectrum of the travel time distribution
Notes de l'éditeur
As opposed to embedded infrastructure such as loop detectors, automatic license plate readers, or Bluetooth traffic monitoring, vehicle probe data is provided by third party vendors without the need for costly equipment. These third party vendors aggregate highway vehicles’ position, speed, and location into travel time or speed data. You and I make up some of the sources for this data. GPS devices embedded in cars, Personal Navigation Devices, fleet vehicles, trucks and even cell phones traveling in vehicles are part of the mix. This data is usually aggregated and cleaned to a final product consisting of one speed/travel time reading per roadway segment per time period. This is not the case with NPMRDS. This data is then made available through a web-interface to departments of transportation, transportation consultants, or anyone else who buys the data.
There are two main type of applications for probe data. Operations and Planning. Operations were the first to adopt probe data. The probe data is used to inform traffic management centers the state of the highways they monitoring. The staff at the traffic management centers use the data to identify incident or congestion, to which they can respond with appropriate actions such as deploying help or informing the public through dynamic message signs. The travel time on dynamic message signs can be derived from probe data. Other travel information sources such as website or GPS devices can use traffic information from probe data. GPS devices are both sources and users and probe data.
Planning adopted probe data later than operations, but none the less, planning now uses probe data extensively. Departments of transportation use probe data to see how their facilities are operating throughout the year. Unlike in operations, which usually use real-time data, planning generally uses archived from longer time periods to come up with certain metrics or performance measurements that allude to how well facilities are operating. Performance monitoring has become even more essential as it is required under the MAP-21 law. With transportation budgets being tight these days, the probe data can help planners pinpoint which facilities need investment the most.
Probe data was first used for freeway-based applications because freeways carry more traffic and there is less mileage overall. These qualities make freeways easy to focus on. Jurisdictions began wanting to also use probe data on arterials. When I say arterial, I mean any interrupted flow facility, not just a major at-grade road. The I-95 Corridor Coalition Vehicle Probe Project (VPP), a program which allows eastern US jurisdiction access and tools to analyze probe data across state lines, monitors the accuracy of probe data as part of the program. Multiple VPP validation studies showed that probe data of freeways generally reflect the actual conditions, while arterial data can sometimes be different than reality. This is due for a number of reasons.
There are fundamental differences between freeways and arterials. Freeways generally have high volumes which flow continuously. The average speed within a platoon is usually pretty close to the speed of the vehicles that make up the platoon. Arterial roads have lower volumes than freeways. This lowers the amount of probe data available for arterial facilities, making the data less reliable and more prone to outliers. Traffic flow is interrupted from a number of difference sources. Red lights completely cut off traffic flow. In extreme congestion this can cause multi-cycle signal failures where it take vehicles multiple cycles to get through an intersection. A single platoon of vehicles can be broken up, causing vehicles in the platoon to experience drastically different travel times. These multi-modal travel time distributions can be difficult to ingest for the algorithms that clean probe data to interpret as legitimate data points. to This cars pulling in and out of driveways of adjacent land uses can disrupt flow of some, but not all of through traffic. All of these issues create noise or distortions in the arterial data.
Arterials roadways differ between themselves. Some have volumes close to freeways, while some have very low volumes. Some arterials have multiple signalized intersections per mile and some have a few miles between each signalized intersections. Some have multiple wide lanes and selected set of driveways, others have a few narrow lanes with a driveway every hundred feet. There is truly a spectrum of arterials between whether the arterial is designed to get vehicles moving long distances (mobility) or whether they are designed to get vehicles to adjacent land uses.
Across this spectrum, which arterials can use probe data for certain applications? This is the subject of our paper.
This research came about indirectly through the VPP validation process. The VPP contract requires that vendors’ data meet certain quality metrics. The coalition performs frequent validation studies to ensure that these metrics are met. For freeways, the involves calculated the Average Absolute Speed Error (AASE) and the Speed Error Bias (SEB). AASE is the absolute value of the difference between the probe data and a trusted reference data source. For the I95 corridor coalition, Bluetooth Traffic Monitoring is the reference data source. SEB is similar to AASE, except the absolute value is not taken and in VPP’s case, it is done on different speed brackets. This type of methodology works well on uni-modal freeway travel time distributions, but not multi-modal arterial travel time distributions.
Central tendency values such as mean or median loses its mean when you have multiple peaks.
Instead of focusing on central tendency or reliability measures, this alternative approach graphs data from probe data and a reference data set separately. First we start by comparing the data sources on a longer time period. On your right you will see a comparison of the reference BTM data set on the top and VPP data on the bottom. This 28 day scatter plot shows that the VPP data density is lower, but generally captures the change in travel time over time.
Next, we zoom in to a smaller time scale. The top left graph shows an hourly scatter plot for a 24 hour period and the bottom left shows the corresponding histogram. In many cases, arterials don’t have enough data from a single hour of the day to analyze. With the alternative method. Multiple days are combined into one overlay scatter plot. This could be all days of the week, weekday only, or even Tuesday-Thursday. Each hour period from the hourly scatter plot, a cumulative frequency diagram is formed. The darkened hour on the hourly overlay scatter plot corresponds to the darkened distribution on the right.
You can directly calculate performance and reliability measures from these CFD. Planning time index (PTI), is a reliability measure that indicates what factor one should multiple their free-flow travel time by to ensure they reach their destination 95% of the time. PTI only uses two percentiles for its derivation. If you compare the PTI calculated from two different data sources, you miss a lot of the spectrum, and many of the speed bins. If you consider the whole scatter plot, you get the whole view. Why use performance measures or reliability measures?
This first example corridor is US-1 in Mercer County, New Jersey around Princeton. There is 6-8 through lanes, less than 1 signalized intersection per mile along this 3.2 mile long corridor. Grade separate interchanges help keep the signal density low. There are minimal drive ways through the corridor. In fact, this corridor almost resembles a freeway.
The data VPP data compares very well to BTM data. Although VPP’s data density is lower, it matches BTM’s peak hour travel times well. Both of the data sets yield a PTI of about 2.1. Based on this validation study, this section of US-1 would be able to use probe data with no problems.
The second example corridor is US-130 in Burlington County, New Jersey near Philadelphia. This 1.5 mile long corridor has 6 lanes total and about 2 signals per mile. Vehicles using this corridor are subjected to multi-cycle signal failures.
The results for example 2 are drastically different than example 1. You can notice the multi cycle signal failure in the BTM scatter plot, but not the VPP scatter plot. The multi-cycle signal failure is represented by the horizontal dashed line at around 2.5 minutes. The is a smaller cluster of vehicles below 2.5 minutes, but a dense cluster of vehicles is seen above 2.5 minutes. The BTM data also shows how the signal timing plan changes after 3 minutes with the denser region appear below 2.5 minutes. Again, the VPP data does not show this. When compared the CFD, the VPP CFD shows that most vehicles experience a travel time of around 2 minutes during 7 pm. In reality, only about 20% of vehicles are lucky to experience that travel time or less. The distortion in the CFD is evident when calculating the PTI. The PTI from VPP data is 1.4 as opposed to 2.5 from BTM data. The VPP data makes the corridor look more reliable than it actually is. Based on this anaylsis, probe data is not suitable for this corridor.