Multiple Model Filtering and Data Association with Application to Ground Target Tracking

This thesis is concerned with two central parts of a tracking system, namely multiple-model filtering and data association. Multiple models are introduced to provide accurate filtering, whereas data association deals with the unknown origin of the received measurements. For multiple-model filtering in a tracking framework, the preferred method has long been the interacting multiple model (IMM) filter. The filter finds a sub-optimal solution to a problem that has the implicit assumption that immediate model shifts have the highest probability. In this thesis an alternative switching model is proposed, which forces the models to persist for at least a model-specific time. Through the assumption, a less complex problem in terms of model hypotheses arises, and to that problem a state estimation algorithm is derived. Under the model assumptions, the filtering algorithm is close to optimal. The proposed filter, called the switch-time conditioned IMM (STC-IMM) filter, is compared with the IMM filter for two different problems, and the results indicate better performance of STC-IMM for both problems. In the area of data association, the thesis is concerned with two conceptually different multi-target tracking algorithms. The first one is multiple hypothesis tracking (MHT) and the second one is the recently developed Gaussian mixture cardinalized probability hypothesis density (GM-CPHD) filter, which is based on a random finite set description of the targets. The objective in the studied multi-target tracking problem is to estimate the number of targets and their state vectors. In the thesis, a performance evaluation of MHT and GM-CPHD is performed on a ground target tracking scenario with nine closely spaced targets. To assess performance, the Optimal Subpattern Assignment (OSPA) measure for multi-target tracking is used in conjunction with a root-mean square error of the estimate of the number of targets. Differences between the filters are pointed out and discussed. In the thesis, an alternative derivation of the GM-CPHD filter is also presented. In difference to the original derivation of GM-CPHD, the approach of the thesis is not based on finite set statistics, which might enhance understanding of the derivations