This Ph.D. thesis is concerned with the development of algorithms for the detection and tracking of multiple targets using wireless sensor networks from the Bayesian standpoint. This is achieved by calculating the probability density function (PDF) of the multitarget state given the sensor measurements (posterior PDF) as it includes all the useful information to perform these tasks. The models of the target dynamics and the sensor measurements are usually nonlinear/non-Gaussian. Therefore, the posterior PDF cannot be calculated in closed form and approximations need to be made. Particle filters' approximations to the posterior PDF are convergent if the number of particles tends to infinity. However, in a practical situation, the computer power available is limited. As a result, the number of particles is bounded and particle filter performance is not guaranteed to be high. This decrease in performance due to the limited computational power is even more acute in a multiple target situation because of the high dimension of the state.
Therefore, this thesis focuses on the development of particle filtering techniques with lower computational burden and higher performance than previously developed ones. Three different scenarios are considered: the detection and tracking of an unknown and variable number of targets using a sensor network, the tracking of targets when there is uncertainty in the sensor positions and the tracking of targets when a non-synchronised sensor network is used.
As regards the detection and tracking of an unknown and variable number of targets, a particle filter with two layers is proposed to detect targets and an efficient algorithm, called the parallel partition method, is developed to track the detected targets. Also, a technique to extract target labelling information when there are two targets is proposed. That is, the filter is able to decide which target is which and determine the probability of error.
The tracking of targets when there is uncertainty in the sensor positions is carried out by simultaneously localising the sensors and tracking the targets using simultaneous localisation and mapping (SLAM) techniques, traditionally used in the field of robotics. However, the multiple target nature of the problem implies that traditional SLAM techniques are not suitable and a new technique, which is based on the parallel partition method, is proposed to overcome the problems of conventional SLAM techniques. Additionally, the truncated Kalman filter also presented in this thesis is of great importance to estimate the positions of the sensors and is shown to be a very useful filtering technique that can be applied to a variety of filtering problems.
When the sensors are not synchronised, conventional particle filtering techniques have a large computational load. Therefore, in this thesis, the asynchronous particle filter is proposed to lower their computational burden while providing accurate estimates.