Development of Different Radio Tomographic Imaging Techniques by Exploiting Sparsity

Abstract

In contemporary research, target localization has become a prevalent challenge in wireless sensor networks (WSN). The localization of targets in WSN is facilitated by attaching sensors to the target. Such a localization technique is referred to as device-based target localization. The device-based localization encounters issues such as dif-ficulty with elderly monitoring, security, and inaccessibility by non-intended targets. This opens up the scope for device-free target localization. In device-free localization (DFL) techniques, no sensor is attached to the target that needs to be localized. Most importantly, the DFL system can localize targets in WSN without knowledge of the localization system, which thereby helps with privacy during the localization process. Therefore, a DFL system can be applied for rescue operations, intrusion detection, roadside surveillance, and health care systems. Radio tomographic imaging (RTI) is one such DFL technique that enables the localization of single or multiple targets by finding the attenuation information of radio waves caused by the targets. The radio wave attenuation information is provided through a radio map and is known as spatial loss fields (SLF). Hence, SLF provide the location and shape of the targets in the RTI system. The estimation of SLF can be done by acquiring the received signal strength (RSS) information of different sensor nodes in the network. A large change in RSS is observed when the targets lie in the line-of-sight (LOS) of the link. The change in RSS information, along with the weight of every pixel in the network, is used for the estimation of the SLF. Using the RSS value of every sensor node and the weight of each pixel, the underlying SLF can be found through a linear regression model. Generally, the number of observed data (RSS) is quite less than the number of predictors (pix-els) of the network, which indicates RTI is an ill-posed problem. Hence, the estimated SLF obtained from the linear least square algorithm suffers from poor localization and shape of the targets. Therefore, in the RTI system, appropriate prior information about the underlying SLF is incorporated through a regularized cost function that enables an accurate SLF estimation. The existing RTI models try to improve target localization by ensuring a low localization error. However, some major real-world RTI issues, such as estimation of SLF by removing surrounding noisy pixels (improving sparsity), esti-mation of SLF in the presence of data uncertainty, improving the convergence speed of RTI algorithms that enable fast and online estimation of SLF, and most importantly, an accurate estimation of SLF in the presence of node or link failure, need to be given insight and solved efficiently with a lower computational burden. This dissertation primarily seeks to give some robust estimators that can estimate the SLF by incorporating real-world RTI problems such as data uncertainty, slower convergence owing to batch processing of the data, and link or node failure issues in a centralized RTI framework. Measurement errors and external factors like wind and soil movement, which cause movement of the sensor nodes, are accountable for the data uncertainty. The stochastic robust approximation (SRA) and worst-case robust approx-imation (WCRA) algorithms offer exceptionally accurate vector estimates when there is uncertainty in the data. Therefore, the regularized SRA-based estimators can be used to handle the sensor location uncertainty that results in an uncertain weight matrix for the RTI system. Our main contribution is to develop regularized SRA-based esti-mators by taking into account the uncertain region surrounding the nominal position of the sensor nodes in both sparse and non-sparse circumstances. It is observed that the l2- norm-based SRA does not improve the sparsity in the estimated SLF and leads to a higher computational cost. A sparsity-promoting l1-norm based SRA (l1-SRA) that can localize the targets with sensor location uncertainty has been developed to re-duce the computational load. However, it is observed that l1-SRA does not promote the smoothness attribute of the SLF. In order to assure reliable SLF estimation, a structured sparsity-based least absolute shrinkage selection operator (lasso) is used. This estimator is otherwise termed a fused lasso (FL)-based estimator. Hence, the FL-SRA estimator is suggested. It concurrently enhances the sparsity and structural features of the under-lying SLF. Certain types of uncertainty in observed data occur due to the digitization of data, where the main source of uncertainty is the quantization of RSS data in the RTI system. Due to this quantized RSS (q-RSS), the SLF estimation becomes inaccurate. Consequently, support vector regression (SVR)-based estimators are proposed for the RTI system to handle the uncertainty caused by the quantization error. The epsilon SVR (ϵ-SVR) model is used, which eliminates the error inside the ϵ-band. The ν-SVR is used to deal with quantized data uncertainty by adaptively estimating the epsilon value to make the SLF estimation more robust against quantized data. Furthermore, a fused l2-norm based SVR (F-l2-SVR) approach is proposed that increases the correla-tion between the pixels. The proposed sparsity-improving l1-norm SVR, also referred to as linear programming-SVR (LP-SVR), is capable of eliminating the surrounding small noisy pixels by giving the noisy pixels the lowest power and overcomes the limitations of the F-l2-SVR technique. This sparse-based LP-SVR technique has a significant re-duction in computational expenses for the estimation of SLF due to the use of a lower number of support vectors. Additionally, the FL-based SVR (FL-l1-SVR) approach is used, which has a lower computational cost than F-l2-SVR but a slightly higher com-putational cost than LP-SVR. The main benefit of this sparse-based SLF estimation is that it can simultaneously retain the SLF’s sparsity as well as its structural details. Hence, such estimators can provide structured sparsity for the estimated SLF. It has been noticed that the majority of RTI systems use first-order algorithms. The first-order algorithms have a slower convergence rate as well as a higher mean square error (MSE) compared to the second-order algorithms. Again, the majority of RTI algorithms em-ploy batch processing. On-line data processing is therefore more in demand than batch data processing. Hence, a fast algorithm is proposed that can provide high target lo-calization accuracy as well as accurate estimation of the SLF. In addition to improving the structured sparsity-based SLF estimation, the proposed time-norm-weighted-fused least absolute shrinkage selection operator (lasso), i.e., the TNWFL technique in RTI, also offers a faster operation when compared to first-order FL algorithms. Although the suggested RTI algorithms offer reliable SLF estimation for various real-world RTI issues, their effectiveness is only compatible with fusion centre (FC)-based SLF estima-tion. The FC may get overburdened under a centralized RTI architecture. In centralized RTI, the SLF estimation accuracy is also affected due to node or link failure. A dis-tributed RTI system that is resilient to node or link failure is proposed to address the shortcomings of a centralized RTI system. The proposed distributed incremental RTI system is capable of sharing the image vector or SLF information with its immediate neighbour. After completing a cyclic path from the initial node to the end node, the global SLF is acquired. Due to the NP-hardness of this distributed incremental RTI, it cannot be regarded as a fully distributed technique. The alternating direction method of multipliers (ADMM)-based consensus technique thus offers a fully distributed frame-work for the RTI system. For distributed estimation of SLF, the proposed distributed consensus ADMM RTI (DCADMM-RTI) is used for non-sparse applications. However, the distributed sparse consensus ADMM RTI (DSCADMM-RTI) allows sparsity-based distributed estimation. The main contribution is to effectively share local SLF among neighbouring nodes in order to establish a global SLF of interest. In this consensus-based approach, all sensor nodes obtain SLF data comparable to the global SLF after a few convergence iterations. In the RTI system, there is a strong possibility that there will be fewer targets in a large monitored region. The SLF of interest is thus believed to be sparse. Sparsity in an estimated SLF means there are more zero pixels than non-zero pixels. The sparsity-based regularized objective function has applications in target lo-calization in RTI, spectrum cartography of radio frequency maps, and improved feature selection in 5G communication. The proposed robust algorithms are therefore simulated with their respective sparsity-based counterparts for the RTI system. The efficiency of these robust estimators for various real-world RTI circumstances is examined for both single and multiple targets in the monitored region. It has been observed that sparsity-based robust estimators perform almost similarly for single- and multi-target scenarios. However, the non-sparse robust estimator shows a little performance degradation for multi-target scenarios compared to single-target localization scenarios. The quantita-tive analysis of estimated SLF is examined through various performance metrics such as root mean square error (RMSE), pixel attenuation ratio (PAR), structural similarity (SSIM), and feature similarity (FSIM). The simulation findings indicate that under var-ious real-world RTI circumstances, the proposed FL-based robust estimators offer the most precise SLF estimation and lowest RMSE among all robust estimators.