Internet of Things (IoT) utilizes sensors as the information source of machine intelligence. Its applications widely exist from Smart Home, Smart City to Wearable Healthcare and Smart Farming. An IoT architecture usually covers four stages: sensor data connection, data transmission, data processing and application model. On top of prediction precision, the interest of IoT research includes efficiency, economic saving and system scalability.In pursuit of these goals, we push the limit of IoT system design from the following three perspectives. (1) We exploit the potential of sensors of smart devices, including sensor fusion and possibility of new IoT applications. (2) We design Machine Learning models for IoT applications, including feature engineering and model selection. (3) We implement lightweight IoT systems for smart devices like laptops, smartphones and voice assistants, considering the constraint of computation resources. In this dissertation, we especially introduce our effort to IoT applications related to localization and security. EyeLoc is a smartphone vision enabled localization we designed for large shopping malls. The results show that the 90-percentile errors of localization and heading direction are 5.97 m and 20◦ in a 70,000 m2 mall. Patronus protects acoustic privacy from malicious secret audio recordings using the nonlinear effect of microphones. Our experiments show that only 19.7% of the words protected by Patronus can be recognized by unauthorized recorders. SoundFlower is a sound source localization system for voice assistants. It can locate a user in 3D space through the wake-up command with a median error of 0.45 m. In general, we explore the potential of diverse sensors to IoT services and build machine learning models to exploit the most information from sensor data. The applications we study are specifically about localization and security.
Read
