Setting: indoor environment. multiple entrances, hallways, offices, conference rooms, multiple wireless base stations.

Questions they would like to answer * Where is the professor * What is the professor doing now * What is the professor's ultimate goal

Sensory Data -> Locations -> Actions -> Goals

Collect RF signal strength from base stations

Logical knowledge (environment, map topology), statistical knowledge (sensory data)

Related Work

Location estimation: RADAR, LANDMARC

Patterson applied dynamic Bayesian network to predict transportation mode from GPS

Nguyen applied abstract Hidden Markov model to recognize human activities from video data

Work

Challenge: RF signal strength is highly uncertain and noisy

Use DBN (dynamic Bayesian network) (sensor uncertainty, structured state space)

Location space: set of grid points on map. At each location build a histogram of signal-strength values for each base station, calculate Pr(ss|b, l). b: base station, l: location, ss: signal strength

Action model used to infer higher-level behavior. Use Expectation Maximization (EM) to learn DBN.

State space is large, so use two-level recognition model that trades off between model accuracy and inference efficiency.

Results

Did DBN and DBN + Bigram. Latter faster.

Recognition accuracy 90.5% at 1s interval, 83.2% at 2s interval