Activity Recognition Using Multiple Sensors

by
Shiaokai Wang

Human activity recognition has practical applications and provides an interesting problem for machine learning research. The ability to track a person's activities is relevant to people such as caregivers supporting those with mild cognitive disorders, while the challenge of using sensor data to understand the intricacies of human activities also provides a new area to apply the efforts of machine learning.

This project explores the combination of RFID data along with muliple channels of personal sensing data in recognizing high level human activities along with lower level physical actions. Recent work has shown that labelling objects in a home using RFID tags to classify human activities can be done with a significant degree of accuracy. The limitations of using RFID tags however lie in that they can not answer the more detailed question of how a person is using an object. We present a system that uses a wrist mounted Multi Sensor Board (MSB), that records acceleration and other streams of data, in conjunction with a personal RFID reader that approaches this new question. The system uses the combined information of the RFID data and the MSB to classify physical actions a person is performing within an activity. The action data will then used to improve the overall activity recognition of the system by helping distinguish instances where object data alone is confounding.

Advised by Matthai Philipose and Henry Kautz

CSE 403
Wednesday
May 24, 2006
3:30 - 4:20 pm