Regionwide Traffic Performance Evaluation and Performance Measure Development Using�Multi-Source Data

​

October 10, 2024

2024 ITS Arizona 31^st Annual Conference

​

Ryan Hatch | Senior Transportation Data Scientist

Hyunsoo Noh, Ph.D. | Data Science Administrator

Presentation Outline

• Background, Motivation, Purpose
• Regional Data
• Performance measures evaluated
• Mobility Measures
• Reliability Measures
• Conclusion
• Current and Future work

Soruce: https://pagregion.com/wp-content/docs/pag/2024/08/2023-PAG-UA-Performance-Measure-Project-Final-Report.pdf

Background, Motivation, Purpose

• MAP 21 and the FAST Act require performance measure updates, and so, as the MPO for the eastern Pima County planning area, PAG has been developing new performance measures and cooperating with the state effort to update existing PMs. FHWA and FTA encourage MPOs to practice TSMO strategies.
• Tracking success/failure/progress of TSMO efforts requires performance measures, which require data, which is often challenging to collect for any number of reasons.
• Sensor and probe data is often available at a cost much lower than traditional traffic data collection methods and at many more locations, but performance measures are not necessarily readily available from this data.
• Previous UA project Comparative Analysis and Integration of Regionwide Traffic Data demonstrated the value of Miovision and Max View event data from existing sensor infrastructure in the Tucson metropolitan area and the feasibility of using this data.

MAP 21: Moving Ahead for Progress in the 21st Century | FAST: Fixing America’s Surface Transportation | MPO: metropolitan planning organization | FHWA: Federal Highway Administration

FTA: Federal Transit Administration | TSMO: transportation system management and operations

Background, Motivation, Purpose

• What data is currently available from the existing regional traffic management system?
• What performance measure(s) can be developed from the Miovision sensor data, the Max View and Miovision event data, and the Wejo probe vehicle data?
• What processing steps are needed to ensure that performance measures derived from Wejo data are acceptable for use as ground truth data?
• What acceptance criteria are appropriate to ensure sufficient accuracy of the performance measures estimated from the Max View and Miovision event data and the MAML algorithm?
• How can the available data represent the regional traffic system mobility and reliability?

Regional Data: Wejo

• Trajectories (as a sequence of GPS points) of vehicles that use connected vehicle technologies and/or navigation apps
• latitude
• longitude
• speed
• moving direction (clockwise angle relative to due north)
• timestamp
• trip ID
• information about intersection(s) that the vehicle passes through
• direction (NB, EB, SB, WB) that the vehicle is moving

Regional Data: Wejo*

• Clean and process raw Wejo data
• Use GIS map to filter GPS points within 50 meters of a major arterial
• Use intersections adjacent to the intersection of interest to further filter GPS points
• Some trips (those that suddenly reverse moving direction or have a time gap >10 seconds between 2 consecutive GPS points) are split into 2 trips
• NB, EB, SB, or WB direction is determined from the moving direction (angle) as follows:
• NB: 315° <= angle < 360° or 0° <= angle < 45°
• EB: 45° <= angle < 135°
• SB: 135° <= angle < 225°
• WB: 225° <= angle < 315°
• Directions used to determine the turning movement at the intersection

* Procured in 2022 and Wejo declared bankruptcy in May 2023. However, analytical approach was considered sound in this study.

Regional Data: Miovision

Miovision detector configuration:

• The through movements for all four approaches have both presence and advance detectors configured to cover all through lanes for that approach.
• The left-turn movements for all four approaches have only presence detectors configured, with one detector covering all left-turn lanes for that approach.
• Presence detectors are long loops, and advance detectors are short loops.

​

Regional Data: Miovision

• Downloaded directly from Miovision’s TrafficLink Portal
• Simple Control Delay

​

• Calculated directly from Miovision event data downloaded from Miovision API
• Arrival on Green
• Split Failure

Regional Data: MaxView

Equipment on major roads:

• Advance detector with a bar-shaped configuration that covers all through lanes
• Presence detectors with a long arrow-shaped configurations that cover left-turn lanes

​

The MaxView system is an advanced traffic management system (ATMS) that uses various events generated by signal assets, such as detectors, signal heads, and pedestrian push buttons, to control traffic signals.

Regional Data: MaxView

Equipment on minor roads:

• Presence detectors with long arrow-shaped configurations that cover through lanes
• Presence detectors with long arrow-shaped configurations that cover left-turn lanes

​

Regional Data: MAML* Algorithm

Traditional machine learning methods

• Struggle with heterogeneity caused by different intersections with different detector locations and lengths
• Struggle when only limited data is available

​

​

MAML algorithm

• Fast and general learning of various problems
• Requires only a small amount of data
• Adapts to new tasks after fine-tuning with a small amount of training data

​

*MAML: Model-Agnostic Meta-Learning

Regional Data: MAML Input Variables

Extracted from the Max View and Miovision event data:

• signal status combinations: red green (RG), green green (GG), red red (RR)
• average occupancy time: the average of the time differences between the detector being triggered on and the detector being triggered off over all detections in each of the 3 signal status combinations
• simple waiting time: the average of the time differences between the detector being triggered on and the start of the subsequent green interval over all detections in each of the 3 signal status combinations
• detection event count: number of pairs of ‘detector on, detector off’ events for each of the 3 signal status combinations

�Other information coded as dummy input variables:

• speed limit, hour of the day, number of lanes, shared lane (right or left turn) or not

Performance Measures Evaluated

Regional performance measure estimation

• Control delay
• Arrival on green (AoG)

​

Regional reliability measure estimation

• Control delay
• Arrival on Red (AoR)

​

Some interesting results, but no acceptance criteria developed and not included in the regional performance measure estimation

• Split failure

Control Delay

• Miovision
• Computed as the time difference between the actuation of the stop bar detector during the red interval and the start of the next green interval
• Usually referred to as ‘simple delay’
• Does not include deceleration or acceleration delay
• Wejo
• Consists of deceleration delay, stop delay, and acceleration delay
• Computed as the difference between the actual travel time and the free flow travel time
• Free flow travel time is the average speed of all vehicles passing through an intersection between 10 p.m. and 3 a.m.

Control Delay: Miovision

• Miovision simple delay
• La Cholla Blvd & River Rd (same performance measure available for other intersections with Miovision sensor data)
• Jan. 6 – Jan. 9, 2021
• Similar temporal trends for both LT and TH for all four directions
• Range from 0 to 130 seconds
• LT more variable

Control Delay: Miovision

• Miovision simple delay
• 97 intersections
• January – March 2021
• 12 a.m. – 5 a.m.: roughly log normal distribution with mean < 5 sec and standard deviation < 20 sec
• 7 a.m. – 7 p.m.: mixture of 2 distributions (urban intersections with high delay and non-urban with very low delay)
• 7 p.m. – 12 a.m.: mean and standard deviation gradually decrease

Control Delay: Wejo

Through delay intersection ranking

Control Delay: Wejo

Left turn delay intersection ranking

Control Delay: MaxView (MAML)

• MAML hourly estimate
• Speedway Blvd and Euclid Ave (possible for many others)
• Sept. 14 – 16, 2021
• Clearly higher delay during peak periods and for LT for all directions

Control Delay: MaxView (MAML)

• MAML hourly estimate averaged over all days in Sept. 2021
• Speedway Blvd and Euclid Ave (possible for many others)
• Smoother trend due to averaging over whole month
• LT delay clearly higher during AM and PM peaks, except for EB
• TH delay in all directions is more constant but still noticeably higher during PM peak, except WB

Control Delay: MaxView (MAML)

• Following four slides show monthly average hourly control delay, for all 12 months of 2021 estimated by the MAML algorithm from the event data,
• Showing both TH and LT movements of each approach direction at Speedway Blvd and Euclid Ave.
• Noticeably lower delay for LT during PM peak in June and July due to UA summer break.
• This estimation is possible at many other intersections with event data.

Control Delay: MaxView (MAML)

Speedway Blvd and Euclid Ave SB

Control Delay: MaxView (MAML)

Speedway Blvd and Euclid Ave NB

Control Delay: MaxView (MAML)

Speedway Blvd and Euclid Ave EB

Control Delay: MaxView (MAML)

Speedway Blvd and Euclid Ave WB

Arrival on Green (AoG)

• Miovision
• Percentage of vehicles that arrive at an intersection (detected by advance detectors) during the green interval
• Wejo
• Computed as the ratio of the total number of sample vehicles passing through an intersection without stopping to the total number of sample vehicles passing through the intersection
• 0% means all vehicles arrive during the red interval
• 100% means all vehicles arrive during the green interval
• A vehicle is considered as “stopped” when it has a speed less than 1 mph before passing through the intersection

Arrival on Green: Miovision

• La Cholla Blvd & River Rd (same performance measure available for other intersections with Miovision sensor data)
• Uses advance detectors, which are not installed for left turn movement
• Jan. 6 – Jan. 9, 2021
• EB, NB, SB show consistent 20% - 40% during the daytime
• WB shows high AoG all 3 days during afternoon peak, likely due to deliberate signal phasing to improve WB traffic flow
• Night is inconsistent: NB/SB show low AoG on Jan. 6 at 3 a.m. but much higher AoG on Jan. 7 at 3 a.m.

Arrival on Green: Wejo

Arrival on Green: MaxView (MAML)

• MAML hourly estimate
• Speedway Blvd and Euclid Ave (possible for many others)
• Sept. 14-16, 2021
• Clearly lower AoG during peak periods (except for SB which has much lower traffic volume)

Arrival on Green: MaxView (MAML)

• MAML hourly estimate averaged over all days in Sept. 2021
• Speedway Blvd and Euclid Ave (possible for many others)
• Smoother trend due to averaging over whole month

Arrival on Green: MaxView (MAML)

• Following four slides show heatmaps of monthly average hourly arrival on green for all 12 months of 2021 estimated by the MAML algorithm from the event data
• Showing TH movement of each approach direction at Speedway Blvd and Euclid Ave.
• Consistent patterns for many hours of the day across different months
• Noticeably higher percentage during May, June and July, especially during the PM peak
• This estimation is possible at many other intersections with event data

Arrival on Green: MaxView (MAML)

Speedway Blvd and Euclid Ave SB

Arrival on Green: MaxView (MAML)

Speedway Blvd and Euclid Ave NB

Arrival on Green: MaxView (MAML)

Speedway Blvd and Euclid Ave EB

Arrival on Green: MaxView (MAML)

Speedway Blvd and Euclid Ave WB

Split Failure

• Occurs when not all vehicles in a queue at the start of a green phase are able to pass through the intersection during that green phase
• Computed as the percentage of all vehicles that pass through an intersection during a given time interval that stop more than once

Split Failure: Wejo

• Average by intersection and hour of day for January – March 2021
• Most intersections have very low (< 2.5%) split failure percentages throughout the day so current Signal timing at most intersections serves nearly all vehicles in one cycle.
• Mission Rd and Valencia Rd is an outlier with 10% split failure from 4 p.m. – 5 pm.

Performance Measures Evaluated

• Reliability Performance Measures
• 95th percentile is the value such that 95% of the values in the sampling period are at or below it
• Buffer is the difference between the 95th percentile value and the average value.
• Buffer Index is the buffer divided by the average value.
• All three of these reliability measures are presented here for the Miovision simple control delay at La Cholla Blvd and River Rd during January – March 2021.
• Buffer AoR is presented for the same intersection and time period.
• 95th percentile and buffer split failure is presented for the same intersection and time period.

Simple Control Delay Reliability

• 95th percentile simple delay at La Cholla Blvd & River Rd, January - March 2021

Simple Control Delay Reliability

• Simple delay buffer at La Cholla Blvd & River Rd, January - March 2021

Simple Control Delay Reliability

• Simple delay buffer index at La Cholla Blvd & River Rd, January - March 2021
• Index values above 0.4 are considered unreliable.

Arrival on Red (AoR) Reliability

• Buffer AoR at La Cholla Blvd & River Rd, January - March 2021

Split Failure Reliability

• 95th percentile split failure at La Cholla Blvd & River Rd, January-March 2021
• 0% for most hours, especially for TH movement
• LT is more reliable at night

Split Failure Reliability

• Split failure buffer at La Cholla Blvd & River Rd, January-March 2021

Conclusions

• Wejo data with sufficient sample size can be used to derive accurate performance measures of control delay and AoG.
• The Miovision sensors can provide valuable and reliable delay, AoG, and split failure data.
• Due to the heterogeneity of intersections, use of the MAML algorithm estimation output is limited:
• TH movement has relatively accurate performance measure estimation.
• LT movement performance measures were not sufficiently accurate, even at the monthly average hourly level.
• The use of the MAML algorithm to estimate delay and AoG is recommended for application only at the aggregate network level.

​

Conclusions

• The buffer index is a strong indicator of the reliability of a transportation system, with values ideally very close to 0 but anywhere below 0.4 indicating a reliable system.
• Delay and AoG/AoR show good reliability (< 0.2) during daytime including the AM and PM peaks but poor reliability at night and early morning.
• Split failure buffer index was evaluated but is not recommended due to the overall scarcity of split failure observed in this study.
• To maintain the developed performance measures and incorporate the latest travel patterns and conditions, it is recommended to use a similar new connected or GPS data.

​

Current and Future Work

• Use Miovision sensor data for delay and arrival on green (and possibly split failure) performance measures at the hourly level on a monthly average basis and possibly at the hourly level on a daily basis
• This project and the previous project focused on development of performance measures for motorized traffic. UA project 3 is now exploring other existing data and developing multimodal performance measures for other travel modes, such as bike sharing, e-scooter, and transit.
• PAG develops, maintains, and updates regional performance measures (like those presented in this report) as on-going work not only for the assessment of the transportation system that the performance measures provide but also for the calibration, validation, maintenance, and updates of PAG’s various models.

​

Questions?

Ryan Hatch | Rhatch@pagregion.com

Hyunsoo Noh | Hnoh@pagregion.com

​

Yao-Jan Wu | yaojan@arizona.edu

Xiaofeng Li | lixiaof@hawaii.edu

​