Improving Real-World Accuracy and Data Efficiency in Indoor Light Sensing for Multifunctional Applications

In recent years, there has been an expansion of sensor applications, which includes sensing indoor environments through smart indoor light sensors. These sensors can help explore human daylong indoor light exposure, energy efficient lighting adjustments, reorganize the interior based on guidelines, or personify indoor illumination. However, data collected from these sensors to date exhibit poor accuracy in realistic scenarios, like identification of indoor light sources under unknown sources, under the presence of multiple-sources, or switching from one light to another. Furthermore, the presence of surrounding noise or fluctuations in sensor readings due to irrelevant events can adversely affect classification accuracy and lead to unnecessary resource consumption. Unfortunately, current sensing and classification techniques are data and power inefficient for daylong sensing. Moreover, understanding the capabilities of indoor light sensors and analyzing potential security issues is essential for building secure and privacy-preserving indoor environments.
In this dissertation, we aim to explore several key approaches to address these challenges and enhance real-world classification accuracy. First, we propose generating synthetic and filtered datasets to replicate real-world scenarios and provide a more comprehensive set of examples beyond controlled environments. This helps the classifier become more adept at handling diverse situations encountered in real-life settings. To enable data-efficient robust daylong source identification, we propose multiple approaches. We first introduce the dimension reduction technique to eliminate unnecessary overhead information. Secondly, we develop an intelligent on-device algorithm capable of detecting light source transitions and facilitating time-specific exposure identification. Third, we present SENTREC, a platform designed to identify the most robust and accurate segment within a long sequence of sensed values, at the same time, capable of differentiating between targeted and non-targeted events. Finally, we introduce ScreenSense, a framework that utilizes basic color information from indoor light sources for identifying users’ activities on digital screens. This framework provides a low-power and enhanced privacy solution for monitoring daylong screen activity, as well as educating smart building professionals regarding potential security risks associated with improper installation of these sensors in smart indoor environments.