A Hybrid Prediction Model for Moving Objects

1
A Hybrid Prediction Model for Moving Objects

• Hoyoung Jeung1 Qing Liu2 Heng Tao
  Shen1 Xiaofang Zhou1
• 1School of Information Technology and
  Electrical Engineering
• The University of Queensland
• 2 Tasmanian ICT Center, CSIRO
• Australia

Presented by Hoyoung Jeung
2
(No Transcript)
3
Prediction in Moving Objects Databases

• Moving Objects Databases
• Model an objects positions ? function of time
• Reduce the frequency of location updates
• More effective to express continuous movements
• All locations in MOD ? estimated (predicted)
• Prediction models
• Linear
• Non-linear

4
Limitations of Existing Prediction Models

• Only for near future
• A location at noon from 9 am movements?
• Even for near future
• Mathematic formulas cannot represent preferred
  movements

5
Motivation
work
820 am (Factory)
900 am
800 am (Home)
If the current movements are (home,800),
(factory,820) and a 900 am location is asked
3 days trajectory
We can say it is likely to be at work
6
Problem Formulation

• Given an object's trajectory l0, l1, , ln-1,
  where li denotes a
  d-dimensional location at time i ,
  discover the objects trajectory
  patterns P of the form
• Given the object's h most recent movements
• l0, l1,, lh-1, and the query time tq,
• estimate the object's future location lq
  using P

7
Overview
8
Detection of Frequent Regions

• Decomposing a whole trajectory into
  sub-trajectories
• Grouping positions at time offset k in each
  sub-traj.
• Applying a clustering method

9
Discovery of Trajectory Patterns

• Note
• Applied and modified Apriori to prune rules
• Contradicting the temporal order (e.g.,
  )
• Multiple items in the consequence (e.g.,
  )

10
Trajectory Pattern Tree

• A variant of the Signature-tree Mamoulis ICDE03
• Different leaf nodes, novel encoding of
  signatures

11
Encoding Trajectory Patterns

• Pattern key Premise key Consequence key
• Region key table

12
Encoding Trajectory Patterns (cont.)

• Consequence key
• Pattern key

13
An Example of Searching

• Given query recent movements ,
• Query pattern key q 1000011

1000011
1000011
1000011
1000011
1000011
1000011
14
The Hybrid Prediction Algorithm

• If no pattern at tq?
• If patterns at tq?
• Distance time query / non-distance time query

15
Query Processing Near Future

• Forward algorithm ( When )
• An object is likely to follow the trend of recent
  movements
• Candidate filtering
• Ranking candidates
• Sim(the current movements, the premise of each
  pattern) x confidence
• Current movements are more important

16
Premise Similarity Measure
101101

• Ranking candidates

17
Query Processing Distant Future

• Backward algorithm ( When )
• The recent movements are not so important for
  prediction
• Candidate filtering
• Ranking candidates
• ( Sim(premise) x penalty Sim(consequence) ) x
  confidence
• As the query time , the importance of recent
  movements

18
Datasets

• Generated 4 multiple traj. based on single ones
• Bike real, from GPS mounted bike
• Cow real, from GPS ear tags
• Car real, from GPS equipped car
• Airplane synthetic from real
• Performance Comparison
• The Hybrid Prediction Model (HPM)
• Recursive Motion Function (RMF)

19
Results - Prediction Accuracy

• Change along prediction length
  Change along num. sub-trajectories

20
Results - Effect of Discovery Parameters

• Eps
  MinPts (minimum support)
• Pruning power
  Minimum confidence

21
Results Query Cost

• Query response time
  Performance of TPT

22
Conclusion

• Proposed a new prediction model for moving
  objects
• Defined, discovered, and indexed trajectory
  patterns
• Introduced the hybrid prediction algorithm
• HPM is more accurate and efficient than
  state-of-the-art

23
Thank you