An Accurate and Robust Indoor Localization System Using Deep Learning and Passive Infrared Sensors
Internet of Things (IoT) has evolved significantly over the past decade, enabling a wide range of applications for indoor environments in different fields such as smart building, healthcare, and energy management. A person's location is crucial information in providing their core services. Thus, indoor localization is a fundamental component of these applications. Different technologies are available for indoor localization. However, video camera technology and wearable devices are not always practical in all situations. A video camera may cause privacy issues. People may not cooperate to hold or wear the devices because they forget or feel uncomfortable. As a result, Passive Infrared (PIR) sensors are employed for indoor localization to mitigate these issues. The advantages of PIR sensors include their low power consumption, cost-effectiveness, and low electromagnetic interference. In addition, their analog or voltage outputs can provide fine-grained information regarding a person’s location such as phase and amplitude. Thus, it is possible to leverage analog outputs to estimate a person’s location. However, utilizing PIR analog outputs is not simple due to their ambiguity and a lack of PIR sensor specification.
This research focuses on addressing the challenges of indoor localization based on PIR analog signal including the design of a PIR sensor node, single person localization, and multi-person localization. Firstly, this thesis proposes a novel design of a PIR sensor node and a prototype is developed to collect data. Secondly, this thesis proposes a deep learning framework for a single person localization. The proposed CNN-LSTM model not only extracts features from PIR analog output automatically but also learn temporal dependencies between the extracted features. Lastly, the proposed localization framework is extended to support multi-person localization. This thesis proposes a channel separation method to generate inputs for each person. Then, deep CNN-LSTM estimates a location for each person and a mean bagging method is used to integrate multiple CNN-LSTM models for improving the accuracy of multi-person localization. A set of experiments are conducted, and the proposed localization methods can achieve good results for both single person and multi-person localization.