Alex Skabardonis

1
THE I-80 FIELD EXPERIMENT

• Alex Skabardonis
• Presentation
• PATH Research Meeting
• November 6, 1998

2
OUTLINE

• Overview/Research Team

• Background/Objectives

• Infrastructure Development

• Ongoing Work

Data Visualization
Travel Time Estimation/Prediction
Incident Detection
3
OVERVIEW
PATH MOU-353 Real-Time Algorithms for Travel
Times, O-Ds, Incident Detection PATH
MOU-356 Develop/Use Data for Research PATH
MOU-352 Travel Times from Loop Detectors
NSF/KDI Grant Management of Large Scale Systems
--TMS21
4
THE TEAM
EECS V. Anantharam, J. Malik. S. Russell, A.
Sinclair, P. Varaiya B. Coifman, D. Luddy, H.
Pasula, K Petty STATISTICS P. Bickel, J. Rice,
Y. Ritov J. Kwon, M. Ostland, X. Zhang ITS/CE A.
Skabardonis PATH J. Dahlgren
5
BACKGROUND
The I-880 Study - Database - Algorithms
Roadwatch - Detection - Tracking
6
OBJECTIVES

• Develop Infrastructure

• Data Mining Processes

• Algorithms for ATMIS

7
OUTLINE

• Overview/Research Team

• Background/Objectives

• Infrastructure Development

• Ongoing Work

Data Visualization
Travel Time Estimation/Prediction
Incident Detection
8
TEST SITE
Eight Detector Stations 170 Controllers and
PC Wireless modems for data transmission
9
Video Surveillance System
Roof Pacific Park Tower, Emeryville
10
(No Transcript)
11
(No Transcript)
12
OUTLINE

• Overview/Research Team

• Background/Objectives

• Infrastructure Development

• Ongoing Work

Data Visualization
Travel Time Estimation/Prediction
Incident Detection
13
Data Visualization
Loop Data Patterns of changes in time /space
Demonstration I-880 Database
14
LOOP SPEEDS
15
INCIDENTS
16
SHOCK WAVE
17
DATA ERRORS
18
OUTLINE

• Overview/Research Team

• Background/Objectives

• Infrastructure Development

• Ongoing Work

Data Visualization
Travel Time Estimation/Prediction
Incident Detection
19
TT Estimation Loop Data
Vehicle Reidentification Algorithm
Measure vehicle lengths from speed traps
TEST SECTION
20
ALGORITHM PERFORMANCE
907 veh (60 total)
21
TT Estimation Video
Color based Veh Matching
22
(No Transcript)
23
TT Estimation Video
24
TT Estimation Loop Data/Video
Algorithms for Vehicle Matching based on
vehicle features
Data UCI TestBed, Carlos Sun
25
Principal Components
26
(No Transcript)
27
TT Prediction
Explanatory Variables Detector Data (flows,
occupancies) Departure Time Day of the
Week Approach CART Regression
28
OUTLINE

• Overview/Research Team

• Background/Objectives

• Infrastructure Development

• Ongoing Work

Data Visualization
Travel Time Estimation/Prediction
Incident Detection
29
Incident Detection

• Typical Results DR vs. FAR
• 0.10 FAR, 50 detected
• 0.25 FAR, 80 detected
• 1.00 FAR, 90 detected
• For large freeway system this is bad
• LA has around 500 loop detector stations
• Data reported every 30 seconds
• For 0.25 FAR this mean 150 FA per hour

30

• False alarm rate is confusing
• Defined as false detections divided by
  guesses
• Changing time step will change performance curve
• Hard to compare across different algorithms
• What incidents do you look at?
• Eg on I-880 we had 1210 incidents over 22 days
• Most were RHS breakdowns little effect on traffic

• Existing Algorithms
• They determine a priori what they are going to
  try to detect
• Only run on very clean data

31
Approach use benefit-cost model
Cost of times you dispatch a tow
truck Benefit reduction of congestion due to
scheme Goal Tune algorithm to minimize total
cost
32
Prop incident detection algorithms

• Investigate effect of using historical values
• Compare two algorithms
• Just look for a drop in speed at a loop detector
• Look for a drop in speed relative to historical
  25 percentile speed