A Preliminary Study of Sensing Appliance Usage for Human Activity Recognition Using Mobile Magnetometer Mi Zhang and Alexander A. Sawchuk Department of Electrical Engineering University of Southern California mizhang@usc.edu UbiMI workshop at ACM Ubicomp Conference, September 8, 2012 1 Introduction • Background Human activity recognition is one of the most basic problems in Ubiquitous Computing. Applications: surveillance, security, health care, etc. In this work, we focus on recognizing household activities by detecting appliance usage. • Existing Technology Computer Vision Courtesy to Opportunity Project RFID Courtesy to Intel Research Smart Meter Courtesy to IBM 2 Magnetic Field Sensing • Idea When household appliance is in operation, it generates electromagnetic waves. The electric component has been demonstrated to infer appliance usage at home. How about the magnetic component? • Magnetometer The magnetometer measures the strength and the direction of the earth’s magnetic field in 3D space. It is mainly used for outdoor navigation (referred to as compass). Magnetometer can also be used for detecting magnets and ferromagnetic materials (referred to as metal detector). 3 Key Observation • More Interestingly When household appliances are in operation, the magnetic field around them presents a different pattern compared to the scenarios when these devices are turned off. Hair Dryer is OFF Hair Dryer is ON These changes exhibit different patterns for different devices and act as the signatures of the devices. 4 Our Framework Magneto meter Sliding Window at 6s Feature Extraction Overall Magnetic Field Strength M (t ) mag x (t ) 2 mag y (t ) 2 mag z (t ) 2 Classification f1 f F 2 f n n1 Standard Deviation, Mean Derivatives, Mean Crossing Rate, Dominant Frequency, Dominant Frequency Magnitude, Energy, Spectral Entropy 5 Evaluation • Sensing Hardware and Software Hardware: 3-axis magnetometer in iPhone 4GS, sampling at 15Hz. Software: techBASIC mobile application. Data analysis is performed using MATLAB. • Household Activities Laptop Hair Dryer TV Mobile Phone Microwave 6 Preliminary Results • Scatter Plots Statistical Features Energy Features • Recognition Accuracy Laptop Microwave TV Hair Dryer Mobile Phone 84.3% 100% 81.6% 92.7% 83.4% 7 Limitations and Future Work • Sampling Rate is LOW High frequency components in the magnetic field signal that may contain important information are not captured. • Test on More Appliances • Combine with Other Environmental Sensors light sensor, temperature sensor, and motion sensors 8 Any Questions? 9 Thank You 10