Accuracy in Real-Time Estimation of Travel Times
Galen McGill, Kristin Tufte, Josh Crain, Priya Chavan, Enas Fayed
15th World Congress on Intelligent Transportation Systems
November 18, 2008
Project Summary• Initial Project Phase
– Collected 500 ground truth runs– Analyzed travel time estimation accuracy
• Second Phase– Addressing primary causes of error– Algorithmic adjustments– Analysis of actual DMS message accuracy
Study Area and Data Collection
• 544 Ground truth probe runs
• GPS-enabled vehicles (Garmin iQue ®)
• Detector data from 500 dual loop detectors on Portland-area freeways
I-5 North of Downtown
Map of Study Area
I-84
I-205
I-5 South of Downtown
OR-217
US-26
Downtown Portland
Initial Project Results• Overall average absolute percent error 11%
(SDPE 18%)– 15% of runs had absolute percent errors
larger than 20%• Accuracy varied between segments• Primary causes of error
– Malfunctioning detectors– Large detector spacing– Changing traffic conditions
Overall Estimation Accuracy
3.1%6.1%
14.0%
28.9%31.3%
11.0%
2.8% 2.9%
0.0%
5.0%
10.0%
15.0%
20.0%
25.0%
30.0%
35.0%
< -30% -30% to-20%
-20% to-10%
-10% to0%
0% to10%
10% to20%
20% to30%
> 30%
% Error
% o
f R
un
s
Average Percent Error – All Runs
Outline• Malfunctioning Detectors
– Critical detectors• Large Detector Spacing
– Prioritizing addition of detectors• How significant are changing traffic
conditions?• Is congestion correlated with error?• Historical average period• DMS message accuracy
Malfunctioning Detectors
• 50% of the runs had one or more detector stations malfunctioning
• High error when certain critical detectors failed (i.e. near recurrent bottlenecks)
• Identify set of critical detectors– Prioritize detector maintenance– May not provide travel time when a
critical detector has failed
Critical Detectors on I-5 NB
I-5 and I-405 Junction I-5 NB Columbia River Crossing Bottleneck~
Critical Detectors
Bottlenecks
Large Detector Spacing• Where should detection be added?
– Prioritize locations of additional detection– Understand implications of detector
location• Detectors simulated with ‘virtual detectors’
using probe vehicle speeds• Compared ‘real-time’ travel time estimates
Additional Detection Locations
Terwilliger Curves (mp 298)
Marquam Bridge (mp 300.5)
I-5 NB
I-5 SB/OR 217 Junction (mp 292)
Addition of Detectors
0
5
10
15
20
25
I-5 NB Terwilliger Curves I-5 SB I-5 SB/OR 217Junction
I-5 NB Marquam Bridge
Midpoint
Real-Time
Added Detector (Real-Time)
MA
PE
Additional Detection - Conclusions• Developed methodology for prioritization of
new detector locations• Detection recommended at several
locations on I-5
Changing Traffic Conditions• Travel time estimate provided at start of
segment (DMS), but traffic conditions may change as a vehicle drives through segment
Changing Traffic Conditions
Traffic flow
DMS
End of segment (DMS predicts travel time to this location)
Vehicle (60 mph)Congestion Wave
(?? mph)
Location where vehicle encounters congestion
Travel time estimation incorrect in this section
Estimation error depends on speed of congestion wave (faster wave = more error).
Congestion Wave Speed and Error
• Analyzed four bottlenecks– Average congestion wave speed ranged
from 6.5 mph to 9.7 mph• Effect on error
– 7.5 mph congestion wave; 25 mph speed during congestion gives max error of 13.5%
Traffic flow
Traffic Speed 60 mph
Congestion Wave (7.5 mph)
Traffic Speed 25 mph
Maximum Error by Wave Speed
Is Congestion Correlated with Error?
• Little to no correlation for All Runs• Some correlation on I-5 SB SoD
0
10
20
30
40
50
60
70
80
0 10 20 30 40 50 60 70
Average Loop Speed (mph)
Ab
solu
te P
erce
nt
Err
or
(%)
0
10
20
30
40
50
60
70
80
0 10 20 30 40 50 60 70
Average Loop Speed (mph)
Ab
solu
te P
erce
nt
Err
or
(%)
Loop Speed vs. Error – All Runs Loop Speed vs. Error – I-5 SB SoD
Is Congestion Correlated with Error?
• Tried to correlate variables with error– Average loop speed– Average probe speed– Standard deviation probe speed– Estimated travel time– Minimum loop speed
• No significant correlation pattern found– Some segments correlated; no pattern
across all runs
Effect of Historical Average Period• Travel time estimation algorithms use a
speed average (i.e. 3-minute, 5-minute) for travel time calculation
• 5-minute had lowest error, but was slightly biased towards underestimation
• 3-minute also had low error and was less biased
• Conclusion: 3-minute or 5-minute average is reasonable
DMS Travel Time Accuracy• Ground truth vs. posted DMS travel times• Expected to be fairly accurate, but…
Carman DMS Prediction
Ground Truth Travel Time
< 10 min
10-12 min
12-15 min
> 15 min
< 10 min 6 6 610-12 min 2 3 6 512-15 min 1 3> 15 min
Potential problem??
DMS Travel Time Accuracy• Study showed ODOT’s estimation algorithm
was fairly accurate • DMS travel time messages were much less
accurate– No messages “> 15 minutes” ever posted
• Issue reported to ODOT Staff• Configuration error in the ATMS database
was discovered and corrected
Conclusions• Current algorithm accuracy relatively good• Critical detectors and additional detection
to address high error• Effect of changing conditions may not be
significant• 3-5 minute average window is reasonable• Need to verify actual DMS messages
Acknowledgments• Oregon Department of Transportation
– Dennis Mitchell, Jack Marchant• At Portland State University
– Robert L. Bertini, Sirisha Kothuri• Oregon Transportation, Research and
Education Consortium (OTREC)
Questions?
Thank You!portal.its.pdx.eduwww.its.pdx.edu
Thank You!portal.its.pdx.eduwww.its.pdx.edu