Some statistical models and approaches to target tracking and data association

Date of Completion

January 2003


Engineering, Electronics and Electrical




Target tracking involves estimating the state of a moving object from noisy observations of uncertain origin and is a problem of significant importance to surveillance applications. In a tracking scenario the thorniest problem is of data-association; that is, how to determine which measurements come from which targets. This topic has been studied extensively and a number of solutions have been proposed. Among them, the Probabilistic Multi-Hypothesis Tracker (PMHT) developed by Luginbuhl & Streit is a relatively new one. By making a modification on the measurement model, specifically, positing the measurement/target association process as independent across measurements, the PMHT is able to render a fully-optimal (under the modified assumption) tracker. The PMHT exhibits an elegant structure of easy extensibility and flexibility; and, at the same time, it suffers from some intrinsic problems. ^ The first topic of this dissertation is to explore the PMHT and seek its improvement in practical applications: we analyze its underlying principles, study its problems and suggest some solutions; we exploit its structural flexibility and extend it to various forms to pursue the best performance, and to function as a natural overlay to a hidden Markov “maneuver” process; we compare it to some popular tracking algorithms such as the Probabilistic Data Association Filter (PDAF), Multi-Hypothesis Tracker (MHT) and S-D assignment; we investigate its consistence, scrutinize; its model and derive the performance bound. ^ Fusion is another important topic in tracking, particularly multiple sensor tracking. To date, however, relatively little literature addresses the issue of communication, which appears to be a limited or expensive resource in many systems. The main challenge, also a second topic of this dissertation, is how to reduce the required bandwidth without, or with little, degradation of tracking accuracy. We introduce intelligent quantization schemes in measurement fusion and discuss some practical issues in target tracking and suggest solutions by marrying particle filtering with techniques to work with out-of-sequence-measurements (OOSMs) and quantizers. Simulation results show that via intelligent quantization, 3 to 4 bits per dimension per measurement per transmission is enough for fairly accurate tracking. ^