DATA FUSION AND FILTERING FOR TARGET TRACKING AND IDENTIFICATION

Abstract

<p>This thesis explores two problems in target tracking and identification: (1) robust track state filtering, and (2) decision-level identity fusion. In the first part of the thesis, a novel finite-horizon, discrete-time, time-varying state estimation method based on on the robust semidefinite programming technique is proposed. The proposed method is robust to norm bounded parameter uncertainties in the system model as well as to uncertainties in the noise statistics. The robust performance of the proposed method is achieved by minimizing an upper bound on the worst case variance of the estimation error for all admissible systems. In the second part of the thesis, two decision-level identity fusion models are proposed: Similar Sensor Fusion (SSF) model and Dissimilar Sensor Fusion (DSF) model. In the SSF model, sensors provide reports on a set of common characteristics of a target, and the fusion objective is to find a fusion result which is most consistent with all the sensor reports. In comparison, sensors in the DSF model explore different characteristics of a target. Their reports are fused in a manner that leads to decreased uncertainty on teh arget identity. In other words, these reports reinforce each other to generate increased certainty on the target identity, rather than being averaged to minimize inconsistency. Furthermore, we propose several fundamental principles for identity fusion, based on which all existing and future identity fusion methods can be evaluated and compared. For the SSF model, two fusion methods are proposed: Convex quadratic fusion method and K-L fusion method. In the first method, inconsistencies between the fusion result and the sensor reports are measured by quadratic functions, and the problem is formulated as a convex quadratic programming problem. In the second method, Kullback-Leiber distance is used to measure the inconsistencies among the probabilistic sensor reports. The resulting formulation leads to a generalized analytic center problem. For the DSF model, we use a special objective function in the optimization formulation to accumulate the physical characteristics on a target explored by each sensor. The resulting fusion method involves solving an analytic center problem. Compared with the two classical decision-level identity fusion methods: Bayesian inference method and Dempster-Shafer evidential inference method, the three new fusion methods require no a priori information on the target, and enjoy small computation complexity. In addition, we show that the three new fusion methods, as well as the two classical methods, all satisfy the fundamental principles for identity fusion. The performance of the proposed fusion methods are illustrated in several numerical examples.</p>