A tree search approach to target tracking in clutter

A novel approach to target tracking using tree search techniques is presented. The tracking problem is framed as a generalized sequential detection problem in which every possible sequence of target states is mapped to a path through the search tree. The stack algorithm for depth-first tree search is then employed to navigate the tree and identify the most likely path, or equivalently the most likely sequence of target states, by extending a single promising path in each iteration. The tree-search tracking technique can be viewed as approximating the full Bayesian inference approach by computing the posterior distribution only in regions in which it has significant mass. Unlike approaches that build on Kalman filtering techniques, the proposed stack-based tracker suffers no performance loss in the presence of nonlinear and/or non-Gaussian motion and measurement models. Simulation results show that the stack-based tracker can achieve significant performance gains over the extended Kalman filter for both linear and nonlinear motion models.