Ubicomp: Modelling and Inference

Transportation and location awareness
- inferring life goals using the activity compass
- GPS is not trivial
-- inaccuracies (signal blockages, atmospherics, etc)
-- resolution > 15m
-- coordinate mismatches

Dynamic Bayesian Network
- extension of Markov model
- statistical model handling

State space
- location (street, position)
- velocity
- GPS offset error
- transportation mode

Particle filters
- solves for hidden variables
- stochastic/monte carlo
<-- "snap to a street" is brittle
<-- need to reason about several dimensions

Belief changing over time
- multiply current measurements by previous to converge on accuracy
- w/cloud of particles

Learning
- learning parameters of DBN = learning about user
-- daily routes and habits
-- expectation maximization
- smoothing forward and backward in time

Data set
- 3 mos in the life of one user
- GPS pos and velocity
- at 2 and 10 sec intervals
- eval data divided into 29 "trips"

Accuracy
- need to seed with bus routes to improve
- learns over time

Results
- predicts into the future
- 50% where the user will be 5 blocks into the future
- data collected on researcher
-- recognizes usual parking spots
-- recognizes change in transportation

Q: Trevor Pering, Intel Research
Other modes of transportation?
A: No attempt for airplanes, some errors with bikes

Q: Nigel Otis, Lancaster U.
How stable were your activities? Would this work for non-routine events?
A: He sees significant changes in routine every 2 months. But you need a month and half of data to reasonably learn from behavior
Q: how much data would you need to collect from me to conclude something about me?
A: If appropriate, you could do it immediately.