Light Based Activity Recognition Using Realistic Data

Abstract

In the field of Visible Light Sensing, light sensors are used to extract information from objects which do not actively communicate any information. Previous research within this field proposed the system called SolAR, and proved the possibility of using a solar cell as both a power source and an activity sensor. A wrist mounted solar cell generates more energy than it uses during operation, while achieving a high classification accuracy for different activities. While the wearer performs different activities, the power output of the solar cell fluctuates. In turn, these fluctuations are used to recognise activities. To extend on the concept of SolAR, this paper introduces a prototype to obtain data from different activities while performing day-to-day tasks. During these activities, ordinary actions are performed to emulate natural circumstances. Analysis of this data initially shows no significant drop in accuracy when compared to SolAR. Further examination shows significant differences in mislabelling rates when comparing to the results of SolAR.