Efficient CostBased Tracking of Scheduled Vehicle Journeys Dalia Tiesyte, Christian S. Jensen, Aalborg University, Denmark MDM 2008, April 2730
Outline • • • • • •
Introduction Update policies The cost function Experimental results Conclusions Q/A
Problem Setting Aim: To track bus locations with accuracy guarantees
Objective: To reduce cost of communication between bus and server
3
Problem Setting (cont.) •
Assumptions
•
The vehicles are traveling on known routes. The vehicle receives its position from GPS and forms movement f act function . f sh The vehicle and the server share the prediction function . The central tracker (server) can change its prediction to new f ct prediction .
Problem
Minimize the number of updates inbetween the vehicle and the server. Preserve the accuracy guarantees.
Predicted position Actual position within circle
Tracking and Prediction Prediction
Vehicle
position update
position
Shared Prediction
prediction update
14 18:05 Center 2 18:25 Airport
Server predicted position
Shared Prediction Server’s Prediction
Traffic Information
State of the Art • Tracking algorithms
Shared prediction function and accuracy guarantees. [Čivilis et al., Wolfson et al., etc.]
• Travel time prediction algorithms for vehicle trajectories on known routes (e.g., buses)
Kalman filter [Cathey and Dailey 2003, Shalaby and Farhan 2001, Dailey et al. 2004]
Artificial Neural Networks [Chien et al. 2002, Park et al. 2004, Hee and Rilett 2004]
Existing Tracking Approach Client
Server predict position fact
compare with GPS
receive update
[out of threshold thr]
update DB
[existing trip] [within threshold thr]
get GPS
[finish]
[initialize trip]
store update data
[continue]
send update
store settings
receive settings
new prediction fsh [initialize trip]
send new prediction
7
New Tracking Approach Client
Server predict position fact
compare with GPS
receive update
[out of threshold thr]
update DB
[existing trip] [within threshold thr]
get GPS
[finish]
[initialize trip]
store update data
[continue]
send update
store settings
receive settings
[new information]
new prediction fsh [existing trip]
[initialize trip]
send new prediction
8
Position update Policies •
Timing pointbased tracking
•
Positionbased tracking
•
The updates are issued at defined points on the route. Update when the GPS position deviates by threshold from the server’s predicted position. Provides position accuracy guarantees.
Timebased tracking
Update when the vehicle’s predicted arrival time deviates by threshold from the server’s predicted arrival time. Can be more efficient than positionbased tracking, but provides poor position accuracy guarantees.
Server’s Updates • The server and the vehicle split the threshold.
thrv + thrs = thr
• Always update when the server’s prediction changes. Problem: possibly too frequent updates from the server to the vehicle.
• Update only when the server’s threshold is reached. Problem: possibly too frequent updates from the vehicle to the server.
• Update either when:
the server’s threshold has been reached, or the update would reduce further communication costs.
Tracking Example Actual movement fact
update Shared Prediction fsh thrv Server’s Prediction fct < thrs
The COST Function • Assumptions:
The prediction can only improve. The prediction error has Gaussian distribution with mean 0.
• The cost function depends on:
The difference between the old and the new prediction functions. The reliability of the server’s prediction. The accuracy thresholds. The COST function itself.
The Cost Function (cont.) • Estimates further communication costs of the journey. • Actual costs:
costtotal = update _ number × costup • Estimated costs (costs of the expected average number of updates):
costtotal = ∑ P (update _ number = i ) × i × costup i
The Cost Function (cont.) •
Assuming that no server updates happen,
Shared prediction
P(upv = i ) = P(thrv i <| f act (t ) − f sh (t ) |< thrv (i + 1)) Actual prediction The deviation between the actual and the predicted trajectories is the prediction error:
err (t ) =| f act (t ) − f sh (t ) | Estimate how often the server’s prediction changes
t = tcur + ∆t
Assume Gaussian distribution
Experimental Settings •
Compare the policies:
• •
Update the vehicle when server’s prediction has changed. Update the vehicle when the threshold thrct is reached. Update the vehicle according to the cost function.
Evaluate the total number of updates. Default parameters: Initial prediction variance GPS delay GPS delay start GPS delay end Server prediction variance Average delay detection time Server and vehicle thresholds thrct,thrv
100 s 600 s 0.25 0.75 50 s 300 s 400m (100 s) 1 50
Single update cost Number of timing points
Experimental Results – Variance (PBT) up
up
Only slight increase
More gain with less accuracy var
var
Experimental Results – Threshold (PBT)
Small thr – update always Small thr – large increase
Conclusions • We have defined update policies:
Timingpoint based. Timebased threshold. Positionbased threshold.
• We have defined a cost function. • Communication costs can be reduced significantly compared to the timingpoint based update policy. • In the future, we need to
Develop a cost model that would compute the minimum communication costs. Reduce updates by improving prediction.
Q/A
Thank You!
Problem Setting (cont.) Minimize costs
14 18:05 Center 2 18:25 Airport
Maintain accurate state
GPS
Prediction Position
Central server Traffic Information
Problems: • • •
The algorithms that predict travel times are not accurate. The communication is costly. Accuracy guarantees are required.