Data Efficient Sensing for Daylong Light Exposure Analysis

It's important to gather information regarding our daylong light exposure to understand how different artificial and natural light sources throughout the day physiologically and psychologically affect us. However, continuous transmission of surrounding lighting environment parameters for classifying the source is energy-expensive. Moreover, when such sensors need to function within strict energy constraints, an inefficient way of operation can significantly reduce their lifespan. In this paper, we argue that instead of naively transmitting every sensed value, indoor light sensors should recognize the importance of sample points to achieve reasonable detection accuracy and only responding to particular events of interest. However, identifying the event is challenging, as natural variations and real-world scenarios can often produce fluctuations resembling those of interest. To enhance the intelligence of indoor light sensors for accurately identifying relevant events and optimizing data usage, we first develop an algorithm to find important events in long recordings. Then we prioritize samples within timeframe and implement lossless compression methods to minimize the amount of information necessary for subsequent offline classification. Following the implementation of these methods, we examine their effectiveness in realistic testbed. Based on the result, reduced raw RGB information by up to 97.94% in the testbed scenario without a significant decline in classification accuracy.