The Probabilistic Activity Toolkit: Towards Enabling Activity-Aware Computer Interfaces

Emerging HCI techniques require the ability to recognize activities that occur in the physical world. Systems that recognize home activities have been limited in the variety of activities they recognize, their robustness to noise, and their ease of use. We present a toolkit (PROACT) for activity recognition that addresses these problems by leveraging three novel techniques: automatically mining text documents and the web for activity structure; recognizing object use via Radio Frequency Identification (RFID) technology; and combining these two inputs to infer user behavior with a flexible and scalable, Monte- Carlo based inference engine. As an initial evaluation, we successfully applied our system to a known difficult problem from health care: recognizing multiple "Activities of Daily Living" (ADLs) in a real home environment. Promising results from a user study validate PROACT's approach.