This thesis focuses on the development of mobile tracking approaches in wireless sensor networks (WSNs) with correlated and sparse measurements. In wireless networks, devices have the ability to transfer information over the network nodes via wireless signals. The strength of a wireless signal at a receiver is referred as the received signal strength (RSS) and many wireless technologies such as Wi-Fi, ZigBee, the Global Positioning Systems (GPS), and other Satellite systems provide the RSS measurements for signal transmission. Due to the availability of RSS measurements, various tracking approaches in WSNs were developed based on the RSS measurements.
Unfortunately, the feasibility of tracking using the RSS measurements is highly dependent on the connectivity of the wireless signals. The existing connectivity may be intermittently disrupted due to the low-battery status on the sensor node or temporarily sensor malfunction. In ad-hoc networks, the number of observation of the RSS measurements rapidly changing due to the movements of network nodes and mobile user. As a result, the tracking algorithms have limited data to perform state inference and this prevents accurate tracking. Furthermore, consecutive RSS measurements obtained from nearby sensor nodes exhibit spatio-temporal correlation, which provides extra information to be exploited. Exploiting the statistical information on the measurements noise covariance matrix increases the tracking accuracy. When the number of observations is relatively large, estimating the measurement noise covariance matrix is feasible. However, when they are relatively small, the covariance matrix estimation becomes ill-conditioned and non-invertible. In situations where the RSS measurements are corrupted by outliers, state inference can be misleading. Outliers can come from the sudden environmental disturbances, temporary sensor failures or even from the intrinsic noise of the sensor device. The outliers existence should be considered accordingly to avoid false and poor estimates.
This thesis proposes first a shrinkage-based particle filter for mobile tracking in WSNs. It estimates the correlation in the RSS measurement using the shrinkage estimator. The shrinkage estimator overcomes the problems of ill-conditioned and non-invertibility of the measurement noise covariance matrix. The estimated covariance matrix is then applied to the particle filter. Secondly, it develops a robust shrinkage based particle filter for the problem of outliers in the RSS measurements. The proposed algorithm provides a non-parametric shrinkage estimate and represents a multiple model particle filter. The performances of both proposed filters are demonstrated over challenging scenarios for mobile tracking.
