Advances in target tracking: New approaches to data association, state estimation, track initiation, and data fusion.

摘要

Data association, state estimation, track initiation, and data fusion are four fundamental problems inherent in multiple target tracking. Numerous target tracking methods have been developed to address some or all of these problems. In this dissertation, novel approaches to these key problems are detailed which utilize measurements from distributed sensors.; First, an existing algorithm for multisensor, bearings-only tracking is extended to provide track initiation capabilities. In the existing algorithm, maximum likelihood estimates of the target states and data associations were calculated by iteratively using simulated annealing over the data associations and nonlinear programming on the target states. To determine the actual number of targets, Rissanen's minimum description length (MDL) criterion is applied to the maximum likelihood estimates. It is then shown that the MDL criterion, when applied to the distributed bearings-only problem, yields inconsistent estimates of the number of targets. Consequently, an asymptotically unbiased target identification (AUTI) criterion, which is shown to be consistent, is developed. The consistency of the AUTI criterion is verified through a series of simulations.; Next, the reduced sufficient statistics (RSS) estimation technique developed by Kulhavy is applied to target tracking. In the RSS method the true a posteriori density is approximated by a parameterized model density. An RSS vector dependent on the target measurements is propagated in time, and the model parameters which yield the state estimate can be derived from the RSS vector when desired. The RSS method is then applied to the nonlinear additive Gaussian noise model. To make the resulting algorithm practical, simplifications are introduced which eliminate the need for numerical integration and an exhaustive search. Finally, a simple example is explored to demonstrate the ability of the RSS method to approximate the true posterior density. In contrast, an extended Kalman filter (EKF) applied to the same example fails to capture the multimodal nature of the posterior density.; The RSS method is then applied to two nonlinear tracking problems: distributed bearings-only tracking and infrared (IR) image tracking. Application of the RSS method to distributed estimation produces a linear fusion rule which consists of adding and subtracting the local and global RSS vectors. This has a major advantage over the general fusion rule which requires transmission of the local densities. In the bearings-only problem, an EKF formulated in modified polar coordinates is developed for comparative purposes. Both the RSS and EKF methods are capable of operating using bidirectional communications, however, the EKF method diverges when unidirectional communications is used. In the IR tracking problem, exact expressions for the RSS vector can be found thus eliminating the need for simplifying approximations. The performance of the RSS method is again compared to an EKF tracker. Here, both methods can successfully track a target with a benign trajectory, but the EKF tracker fails in a typical maneuvering scenario, whereas the RSS algorithm maintains track.