Machine Learning for Wideband Localization
Nguyen Van Thang
Centralized consistency cooperative localization Cramer-Rao lower bound (CRLB) distributed eciency Fisher information Gilbert's disk localization network IEEE 802.15.4a least-square (LS) algorithm NLOS identication NLOS mitigation position error bound (PEB) relevance vector machine (RVM) UWB variational message passing (VMP) vision information;
- 원문 URL
Location awareness is becoming essential for emerging wireless applications where most of network activities require the location information of network nodes, e.g., routing between nodes in ad-hoc sensor networks, positioning vehicles on the road, or tracking targets in underwater acoustic sensor networks--creating a new paradigm for distributing scalable multimedia data over wireless networks, enabling a variety of context-aware applications that require precise location information of network nodes. However, in hash conditions (e.g. indoor environment), it is difficult to localize an agent with high accuracy due to the radio blockage (non-line-of-sight (NLOS)) or the deficiency of anchors. In addition, challenges due to precisely determining the positions of agents are still burdened in a variety of positioning systems. In general, it is hard to determine the global solution of a localization problem within a reasonable time-consuming since most of objective functions are non-convex, even simplest networks. To overcome these drawbacks, NLOS identification and mitigation techniques are highlighted as an effective way to improve localization accuracy, especially by applying machine learning solutions or cooperate with other nodes to offer the benefits in both the localization accuracy and coverage in harsh environments. In particular, an emerging concept for robust and accurate network localization is to exploit cooperation and heterogeneous design for harnessing multimodal fusion of sensing measurements to extract location information. In this dissertation, taking into account the presence of NLOS signals in indoor environments, least-square cooperative localization for arbitrary NLOS ranging bias is studied through analyzing the fundamental limits and the assymptotic behaviors of localization accuracy as well as developing a robust and efficient algorithm for localization problems based on relevance vector machine (RVM), leading to the significant enhancement on identication, mitigation, and localization. Moreover, the benefits of RVM in both classifier and regressor are extended for a cooperative localization framework with the aid of cooperative localization algorithms in both centralized and distributed manners. At the end of this work, a brief introduction to vision- and radio-based positioning technologies is provided--openning a promissing eva for future localization systems. This dissertation consists of four main results. First, using LS cooperative localization for a general NLOS bias model, we develop the network position error bound (PEB) through deriving the Fisher information matrix (FIM) and show that a Gaussian bias due to NLOS effects is the worst case that produces the extremal FIM, whereas a constant bias or equivalently full line-of-sight is the best situation leading to the largest FIM in a sense of Lowner partial ordering. We also analyze the asymptotic performance--such as uniform convergence, consistency, and efficiency--of LS cooperative localization to quantify the deviations of localization accuracy for LS, squared-range LS, and squared-range weighted LS solutions from the fundamental limit (i.e., Cramer-Rao lower bound) due to their practical tractability. By integrating squared-range relaxation into Gaussian variational message passing on the localization network, we propose a simple distributed algorithm for LS cooperative localization. To account for stochastic natures of node locations and populations, we introduce the notion of stochastic localization network and characterize the network PEB for Gilbert's disk localization network where anchors and/or agents are randomly distributed in the network according to point processes. Second, we design a RVM classifier to identify NLOS signals based on features extracted from the received waveforms and a RVM regressor to approximate the probability distribution function of the range estimate (pair-nodes distance) by exploiting the prediction ability of RVM regressor. We also propose a new localization algorithm, called variational localization algorithm, which guarantees to converge after a finite number of iterations. Thanks to the fine sparsity of RVM, our proposed algorithm can be readily deployed in practical localization systems to attain a high localization accuracy with a reasonable cost. Third, considering a cooperation localization network, we propose an efficient centralized cooperative localization algorithm--called the centralized cooperative localization algorithm-- guaranteeing to fast converge after a finite number of iterations. For a self-localization setting, we develop a distributed cooperative algorithm based on variational Bayesian inference--called the Gaussian variational message passing (GVMP)--to simplify message representations on factor graphs and reduce the communication overheads between agents. In particular, we build a refined version of GVMP for reducing the computational complexity while maintaining the comparable performance as the original GVMP. Fourth, we presents illustrative machine-learning methodology to successfully integrate vision information and radio time-of-arrival measurements for cooperative localization of ultra-wideband visual radios in harsh indoor environments.