Ubicomp: modelling and inference

Activity
- subdivide into zones
-- ex: geometric locations
-- but you can also define [semantic] zones
<-- learn zones by observing human behavior
--> so that zones change over time

Learning
- use existing fine-scale ps sensing techniques

Activity Zones
- represent location and motion context
- correlate with physical zones through velocity and height, say
- infer activity, trigger app behavior

1) PERCEIVE
- person tracker
-- real time stereo vision, integrates multiple views, learns color and depth
- activity categorization
-- group activities into spatio-temporal similarity classes
- activity feature estimates from this

Activity Map
-- cluster features and zones

2) REASON

Activity Detection
- estimate which zone a person is located in (not geographic but activity)

Inference rules organized into knowledge sources
- infer higher level context
- select appropriate activity

Blackboard : knowledge source rules
- aggregate zones
- define social context
- device control
- message delivery

3) ACTION
Agent-based communication using Metaglue

USER INTERACTION
- zone editor
-- picture of a zone map that the tracker learned (from monitor camera)
-- user chooses zones of interest by filling out a table, changing names, selecting colors, which then affects picture by changing colors and labelling zones
- defining rule in rule editor with simulator
-- uses iconic map to define behavior and rules by dragging icons that represent users into zones

USER STUDIES
- 5 users unfamiliar with concept
-- drew a map of the space
-- arranged the space
-- played in space
-- looked at tracker results and compared with their ordering
- results
-- activity maps different for each user
-- rough correspondence between user and tracked zones
-- understood zones better after they rearranged space [informance/acting-out in context]
-- suggestion: learn preferred settings
-- suggestion: set availability flags
-- suggestion: need to override room