Inference Techniques in Sensor-Based Activity Recognition

by
Alvin Raj

The prospect of ubiquitous computing applications being able to automatically provide context-specific services makes the task of extracting a person's activity from sensor data an important goal. Data from a GPS unit and a shoulder mounted multi-sensor-board which contain audio, accelerometer, light, barometric, and humidity sensors, are used to give the best estimate of a person's environment (e.g. outside, inside a building, or inside a vehicle), and a person's activity (e.g. stopped, walking, running, going upstairs/downstairs, or driving).

Our research focuses on using graphical models to estimate the behavior of the person, while imposing temporal constraints on the state. I shall discuss two approaches used in solving this problem. The first models the problem as a variant of a Hidden Markov Model (HMM) and was implemented using the Graphical Modeling Toolkit (GMTK). The second approach uses Rao-Blackwellised particle filters with attached hierarchical HMMs to obtain a joint estimate of position and activity.

Advised by Dieter Fox

CSE 403
Wednesday
February 22, 2006
3:30 - 4:20 pm