Multitarget multisensor tracking
Santosh Nannuru · 2015
In this thesis we develop various multitarget tracking algorithms that can process measurements from single or multiple sensors. The filters are derived by approximate application of the recursive Bayes filter within the random finite set framework, which is used to model the multitarget state and observations. The contributions of the thesis can be organized into three main categories.To provide a motivating application for the algorithms we develop, we first study the problem of radio frequency tomography. We empirically validate a radio frequency tomography measurement model when multiple targets are present within the sensor network. We validate modelsfor both indoor and outdoor environments. These models are then used to perform multitarget tracking using various Monte Carlo filters on data gathered from field deployments of radio frequency sensor networks.Second, we develop auxiliary particle filter implementations of the Probability Hypothesis Density filter and Cardinalized Probability Hypothesis Density filter when the measurement model has a specific form, namely the superpositional sensor model. We also derive Multi-Bernoulli filter and Hybrid Multi-Bernoulli Cardinalized Probability Hypothesis Density filter for superpositional sensors and develop their auxiliary particle filter implementations. These filters are evaluated for multitarget tracking using simulated radio frequency tomography and acoustic sensor network models.Third, we derive update equations for the General Multisensor Cardinalized Probability Hypothesis Density filter when the measurement model has a specific form, namely the standard sensor model. To overcome the combinatorial computational complexity of this filter we develop a Gaussian mixture model-based greedy algorithmto implement the filter in a computationally tractable manner. The filter is evaluated using simulated multisensor measurements.