Ranking and Optimization of Target Tracking Algorithms

Date The final copy of this thesis has been examined by the signatories, and we find that both the content and the form meet acceptable presentation standards of scholarly work in the above mentioned discipline. In this thesis the performance of target tracking algorithms, and in particular the sequential multi-sensor joint probabilistic data association algorithm is analyzed. The main contribution is the development of a tool for ranking and selecting designs where the performance metric is the standard deviation (e.g., in target tracking, root mean square position error), and not the mean (e.g., network throughput), assuming that target tracking position error has Gaussian distribution. Further improvement in variance ranking and selection is achieved when position error data is modeled as a Gaussian mixture distribution (a weighted sum of Gaus-sian distributions, defined by a set of weights, means, and variances). Parameter estimation of a k-component Gaussian mixture distribution is obtained by applying a modified version of a well known expectation maximization algorithm. In order to reduce long simulations necessary to achieve good confidence in the observed ranking of multi-sensor fusion algorithms, an optimized computing budget allocation approach is applied, using the improved Gaussian distribution models of the root mean square track position error. The developed algorithm for variance ranking and selection leads to more than an order of magnitude reduction in computational effort and produces results with high confidence levels when comparing different orders of processing sensor information in the sequential multi-sensor joint probabilistic data association tracking algorithm. Application of the new ranking technique in comparing various types of particle filters also demonstrates that ranking can be accomplished efficiently and with high confidence levels. The proposed variance ranking and selection technique can be applied for comparing other types of designs and algorithms provided that the performance metric is a variance or root mean square position error. and help throughout the research reported in this thesis. I would also like to thank my thesis defense committee members, Prof. for their interest in my work and for motivating a part of this research that produced results presented in Chapter 6 and in publications [54, 56]. Finally, I am very grateful for the support of my wonderful family whose infinite love, patience, and encouragement helped me complete this work.