Multitarget tracking using maximum likelihood techniques

Date of Completion

January 2007


Engineering, Electronics and Electrical




The Maximum Likelihood-Probabilistic Data Association (ML-PDA) target tracking algorithm was originally developed for tracking Very Low Observable (VLO) or "dim" targets. VLO target tracking is challenging in that traditional Kalman Filter based tracking systems experience difficulty given the large quantity of clutter typically seen in measurement data sets. While effective, ML-PDA has not received wide acceptance as a target tracking algorithm because of its high computational complexity, the need for establishing a method for track validation, and its limitation to tracking single targets. This dissertation addresses each of these issues. First, two new computational methods are compared to the original method for computing the ML-PDA track estimate (Genetic Algorithm and Directed Subspace Search). We show that the Directed Subspace Search reduces the computational complexity of ML-PDA by an order of magnitude. Second, a new methodology for deriving the statistics required for track validation is presented which relies upon Extreme Value Theory (EVT). We show that the statistics of the ML-PDA Log Likelihood Ratio at the track estimate under the "target absent" hypothesis is most closely approximated by a Gumbel distribution and not the Gaussian distribution previously ascribed to it. We present two techniques for obtaining the track validation threshold, an off-line and a real-time technique, and demonstrate improved tracking performance through use of lower track validation threshold values. Third, we derive a version of ML-PDA for use in a multi-sensor problem. Fourth, we develop a multiple-target version of ML-PDA, called MLPDA(MT). MLPDA(MT) uses a multi-target version of the ML-PDA likelihood function for cases where measurements can be associated to multiple targets. Modules for track initiation, track maintenance/update, and track termination are also described. The effectiveness of each of these improvements to ML-PDA is tested through Monte Carlo simulations of target tracking problems and comparisons are made to either the baseline ML-PDA implementations or, in the case of MLPDA(MT), to the Probabilistic Multi-Hypothesis Tracker (PMHT). Simulation results show that by incorporating these innovations into ML-PDA, for the first time real-time target tracking is achievable without parallel processing. Further, ML-PDA(MT) performs better than PMHT in high clutter environments. ^