Localization of Wireless Terminals via Deep Learning

点击翻译按钮获取中文摘要

Localization of wireless terminals has gained momentum over the last few years. Many applications are now being developed that provide location based services. Google map, Find My iPhone, Virtual Reality Headsets, and Pokémon Go are some examples of applications that use location information. Various new applications will be developed when 5G, Internet-of-Things, and Smart Cities technologies become available. Unfortunately, the GPS service is not available in indoors, or is very inaccurate in areas blocked by tall buildings such as downtown cores in major cities. New technologies should be developed to complement GPS and provide location estimation ubiquitously. ******Our research at the University of Toronto has been on location finding in recent past. We have developed new localization technologies based on received signal strength (RSS) using advance signal processing methods such as Compressive Sensing. We have studied WiFi received signal strength based localizations and tracking, device diversity through unsupervised learning, and crowdsourcing. We have two patented technologies and a pending patent application. Currently, we are working on location finding via image and video signal processing. ******The proposed research activity will be the continuation and refinement of our previous research. New studies have shown that using channel state information (CSI) of WiFi signals can give a significant gain in location accuracy, and that a single access point is sufficient to have decimeter level localization accuracy. There are, however, a few shortcomings in the reported literature that limits the application of these methods in practice. The method is only applicable to a pair of nodes, uses the whole ISM band in 2.4 GHz and 5.8 GHz, and its range is limited. Our work will answer some of the open problems in the application of CSI. We will use Deep Learning methods for location finding on RSS and CSI. Deep learning has gained momentum in tackling difficult problems such as voice and video classification, big data, and medical imaging. A deep neural network can extract a wide range of complex features that can be used in regression and classification. Our goal is to develop effective location finding methods that can locate users at decimeter level accuracy operating on off-the-shelf phones without any specialized hardware. **