1. Travel time measurement using Bluetooth detectors 2. Travel times on arterials (characteristics & challenges) 3. Dealing with outliers 4.

Transcription

1

2 1. Travel time measurement using Bluetooth detectors 2. Travel times on arterials (characteristics & challenges) 3. Dealing with outliers 4. Travel time prediction

3 Travel time = :16:00 9:15:50 Travel time = :13:20 9:10:00 9:10:04 9:10:09 Travel time = 2 26 Travel time = 2 30 Travel time = :12:26 9:12:34 Measurement error Detection time error(at intersection installation) Sampling error En-route stops Non-auto observations Multiple devices in a single vehicle

4 Arterial travel time can be decomposed into: Mid-link travel time Intersection delay Stochastic arrival Signal timing Level of saturation Level of coordination A B Time

5 Outliers Sampling Error En-route stops, non-auto observation, multiple devices in a single vehicle Measurement error Detection time error(at intersection installation) Arrival time bias Lack of information (especially en-route) signal timing, level of coordination, and level of saturation Other issues Free-flow-Congestion transition error, time lag problem, segment length, and time interval length

6 Freeway Segment (2.5 km) HW 8 (from 401 to Fairway) Waterloo Region Tuesday April 10 th 2012 Arterial Segment (2.25 km) Highway 7 (from Bayview to Leslie) Toronto Tuesday Sep 2o th 2012

7 Overtaking Rule Statistical Approaches Percentile test Deviation test Adaptive filtering Other statistical approaches Shortcomings: Often unable to distinguish between outliers and sudden changes in actual travel times

8 Travel time vs. time of day (Individual vehicles) Highway 7 (from Leslie to Bayview), Toronto Tuesday Sep 20 th 2012)

9 1) How to evaluate/compare the performance of outlier detection algorithms? For the real data we don t know which observations are really outliers! Therefore, we developed a simulationframework to perform the evaluation process (i.e. we synthesized outliers and added them to travel times obtained from Vissim micro-simulation model). t outlier 2) How do we measure which algorithm is better? Percent detected outliers by type: En-Route stops Transit vehicles Multiple devices Wrong detections Relative travel time improvement (RTTI) RTTI outlier true detected Travel time improvement true t detected t t t t = t true true true t t

10 Simulated Outliers (Individual Vehicles) University Ave. (from Westmount Road to Phillip Street), Waterloo

11 % Detected Outliers: En-route = 100 Transit = 97.6 Multiple Devices = 36.9 Wrong Detection = 41.4 Relative Improvement = % Detected outliers (Method 1)

12 % Detected Outliers: En-route = 89.8 Transit = 46.3 Multiple Devices = 7.7 Wrong Detection = 8.2 Relative Improvement = % Outlier detection algorithm can have a very large impact! Detected outliers (Method 2)

13 Outlier detection algorithm performance is affected by a number of factors including: Bluetooth penetration (%) Proportion and magnitude of en-route stops Buses in traffic stream(%) Vehicles with multiple devices(%) Location of the bluetooth detector(mid-block vs. stop-bar) Traffic conditions(v/c ratio; coordination) Our approach permits detailed evaluation of the importance of these factors for a given outlier detector algorithm

14 Travel time characteristics are different under free-flow and congested conditions Therefore, use different models for these two regimes Estimation Module (by Polling Interval) Travel Time Matched Bluetooth detections Traffic State Determination Prediction Module t=t+δt Free Flow Traffic State Prediction Traffic State? Congested Historical Database Uncongested State Travel Time Prediction Model Congested State Travel Time Prediction Model Predicted Travel Time at Time t

15 Travel time variation during free-flow state is the result of traffic signal control and corridor signal coordination level. measured In free-flow, abrupt changes in average travel time are typically not experienced. a posteriori With the low travel time variation, measurement true errors value may become significant. To predict travel times during free-flow condition Kalman Filter model has been a priori developed. Travel Time Systematic trend (governed by flow) Process variation (governed by signals) Observations Present (governed by measurement error) Time k-7 k-6 k-5 k-4 k-3 k-2 k-1 k Present Time k+1 Time of day

16 Travel Time Travel Time Travel Time Present Time Second upstream segment Time of day Present Time Time of day Immediate upstream segment Present Time Investigated segment Time of day Historical Data Travel Time Similar historical patterns Infer the next time interval travel time from similar historical instances Present Time Time of day

17 12 10 Travel Time (min) Observed Predicted Time of day (hh) Travel time vs. time of day (5 minute intervals) Highway 7 (from Leslie to Bayview), Toronto Wednesday Sep 21 st 2012)

18 Conclusions: Outliers are a more significant issue for arterial travel time measurement and prediction than for freeways. Placement of detectors at the intersection (instead of mid-block) increases the problem. Our proposed model permits different outlier detection algorithms to be compared and evaluated using synthesized data. Next Steps: Incorporate our outlier detection algorithm within our travel time prediction framework and evaluate. Apply to a field site

19