The range of applications where target tracking is useful has grown well beyond the classical military and radar-based tracking applications. With the increasing enthusiasm in autonomous solutions for vehicular and robotics navigation, much of the maneuverability can be provided based on solutions that can track multiple targets, in a computationally inexpensive and accurate manner.

This thesis is concerned with solving the problem of tracking multiple
point targets, which is a common occurrence in many tracking applications. The main challenge in multi-target tracking is to resolve measurement-to-target association uncertainties, also called the data association uncertainties. Using the Bayesian approach, these uncertainties are modeled as random variables; where, each instance of the data association variable corresponds to a unique data association hypothesis. The objective, then, is to use the measurements from sensors to choose the best data association hypothesis, from which the estimates of target trajectories can be obtained. In an ideal world, we could maintain all possible data association hypotheses from observing all measurements, and pick the best hypothesis. But, it turns out the number of data association hypotheses grows exponentially with the number of measurements over time, rendering this optimal solution intractable. Vast literature in the multi-target tracking is dedicated to solving this problem tractably.

In this thesis, a variational Bayesian approach has been used, more specifically, the main contribution is to use expectation maximization (EM) in tracking multiple point targets. EM is an iterative algorithm that can be used to approximate the best data association hypotheses, and/or target state estimates in a computationally efficient manner. Depending on, if the overall joint density is maximized over the data association variables, or over the target state variables, two EM-based algorithms for tracking multiple point targets are derived, implemented and evaluated. In the first algorithm, the data association variable is integrated out, and the target states are estimated. In the second algorithm, the data association variable is estimated, while the target states are integrated out. In the end, both the algorithms yield the desired target states. It is shown that the two proposed algorithms can be implemented using existing, simpler solution blocks like Bayesian smoothing, 2-D auction algorithm and computation of marginal data association probabilities. Performance of the two proposed algorithms are compared with existing multi-target algorithms, to highlight the overall improvement in performance, and possible future extensions to the work are presented.