
A Hybrid Modality for Human Sensing
Networked cameras are an integral component of many infrastructures and hold great promise for providing ubiquitous services. Camera networks however cannot persistently track and identify people and camera paths are usually fragmented due to the vision correspondence problem. To bypass the vision correspondence problem, we have developed a hybrid sensing modality comprised of cameras deployed in the infrastructure and inertial sensors (accelerometers and magnetometers) found in mobile phones. Signals from all modalities are combined to establish correspondence between the identity and observed location of the person wearing the sensors.

Formalizing Sensor Node Lifetime Prediction
Wireless devices with limited battery or intermittent harvested resources need to consider the tradeoffs between duty cycling, and application needs to guarantee reliable operation. After experiencing the design and implementation of a few generations of sensors nodes, we have developed a detailed framework for analyzing the asymptotic lifetimes of wireless sensor nodes with respect to application needs. This considers the architecture properties, choice of components and application profiles to determine how different architectures would perform in different environments.
Extracting Spatio-Temporal Activities from GPS traces
Using GPS and camera data to understand how humans spend time at different places is a key component of many automated services from reminders and triggers to safety applications, feedback on how to conserve energy in buildings and energy management of smart phones. To allow embedded devices to exploit information contained in location data, we have developed a lightweight algorithm that can automatically extract Spatio-Temporal Activities (STAs) from traces of locations and classify them according to their temporal attributes. The resulting classes of STAs provide a summary of the important places and the schedule of the user.

The Behaviorscope Architecture
The BehaviorScope architecture brings together a heterogeneous set of sensor driven data sources into a unified hierarchical framework that converts low-level sensor data to high-level semantic forms that can be used in decision making and to provide services. This processing is done using heterogeneous set of methods including Time-Augmented Probabilistic Context Free Grammars, temporal event classification, subsequence mining and other new forms of spatio-temporal data processing.

