TRANSFORMER-BASED MULTIMODEL INDOOR LOCALIZATION USING WIRELESS AND INERTIAL SENSORS

Abstract

The challenge of outdoor navigation has effectively been been met by the deploy­ment of the Global Positioning System (GPS) and the ubiquity of mobile devices equipped with GPS sensors. However, indoor user localization and navigation are less effective , due to problems of device signal attenuation and outright blockage of satellite signals, which in turn hinders critical use-cases such as emergency response and rescue operations. In response to this challenge, researchers have developed a variety of indoor localization algorithms that utilize other methods to approximate location and estimate movement; one example uses the omnipresent wireless network infrastructure, which can be used to determine approximate location by assessing the Received Signal Strength (RSS) data as the user moves about. Using this technique, researchers are steadily improving the accuracy of localization measures. It is the hypothesis of this work that indoor localization accuracy can be further improved by taking into account other sensors which sample at shorter intervals and higher frequencies. As the proof-of-concept, we utilize Inertial Measurement Unit (IMU) sensor readings in addition to the Wi-Fi Received Signal Strength (RSS). we hypothesize that the IMU measurements can be used to describe the user displacement information and thereby improve the localization accuracy. To test the hypothesis we created an Android mobile application deployed in a testbed environment constrained to a single building, such that we could then collect the necessary RSS and IMU sensor data (from more than 100 path trajectories) needed to test the hypothesis. The next stage required the development of baseline Recurrent Neural Networks (RNNs) using only the RSS data and then expanding this network to utilize the IMU measurements. Upon verifying the model using our own data, we modified the model to incorporate the use of a state-of-the-art Transformer Networks architecture. Using both the RNN and TN models, we test both RSS data and RSS+ IMU data, with the best results attained from the TN model using RSS+IMU data.