الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 73 |  |  |
| | | from | 73 | | |
Abstract
Indoor localization has been an active research area aiming to replicate the success witnessed by GPS in outdoor environments. However, opposite to GPS, room level or sub-meter accuracy is needed for indoor environments to support numerous context aware applications. Many solutions have been proposed that involve multi-lateration, angle of arrival, or triangulation techniques, of various radio, acoustic, or ultrasound wireless technologies. The two most popular approaches, however, are those related to fingerprinting and time-based techniques in WiFi environments. In this work, we propose a deep learning-based indoor localization approach that leverages WiFi signals time of flight (ToF) as environment features, to provide accurate and robust indoor localization. DeepNarl leverages the fine-time measurement (FTM) protocol in the recent IEEE 802.11-2016 standard to measure WiFi signal round trip time (RTT). Our system combines the advantages of fingerprinting and ranging-based techniques by providing a deep learning model along with a probabilistic framework that captures the complex relation between the propagation times of the WiFi signals heard by the mobile phone and its location. By leveraging the signals RTT, collected using commercial-off-the-shelf access points and mobile phones, DeepNar overcomes the different challenges of indoor environments such as the multipath interference, non-line-of-sight transmissions, signal attenuation, and interference. Moreover, DeepNar does not require clock synchronization between the transmitter and the receiver. Our system is composed of various components that handle outlier detection, avoids over-training, and accommodates heterogeneous devices. We implemented and evaluated DeepNar over two testbeds. Our results show that DeepNar has a sub-meter localization accuracy with a median error less than 0.75m. This accuracy outperforms ranging-based multi-lateration technique by at least 182% and traditional signal strength (RSS) fingerprinting techniques by more than 119% and 33% in both testbeds considered.