Fusion for Improved Resolution of Unreliable User-sourced Data.
详细信息
文摘
There is abundant information available online,noisy and imperfect but generated for human consumption,which can feed solutions to problems that would otherwise be difficult to study if it were necessary to generate all data centrally. This dissertation describes two modern-day problems whose solutions are enabled by the recent and widespread availability of freely shared data on the Internet contributed by independent parties. We pursue an overarching strategy of accumulating inputs and applying data fusion to produce an improved outcome over any single input. The first problem is Indoor Localization and Wayfinding. The conventional approach is to either a) attempt to replicate GPS indoors with some form of signal triangulation,or b) build a location database of fingerprints based on various ambient signal characteristics. Our approach takes advantage of user interaction,along with publicly shared online floorplans,to deduce a likely location. Location is then followed by navigation,which reveals insights about the nature of indoor environments in the context of random graphs. The resulting characterization allows us to predict which environments are likely to support successful localization and wayfinding. The second problem introduces a new diversity combining technique,Normalized Gain Combining NGC). This is a form of signal fusion which takes as its inputs multiple received versions of a source signal and returns a single output with improved SNR. This problem is motivated by a real-world scenario in which audience members at a musical performance employ personal mobile devices to capture recordings of the performance. They then upload their recordings to a shared online database. Without knowledge of channel gains or the transmitted signal,we identify the best of the original received samples,rank them by inferred relative SNR,and combine them to generate an improved composite. A new noise random variable,Rayleigh-normalized Gaussian noise,is identified and analyzed relative to its probability density and distribution. Finally,we investigate the marginal impact of increasingly large sample sets. This work demonstrates the viability of unreliable,freely shared data on the Internet as a motivator for research into a variety of useful engineering problems.