Bayesian Target Enumeration and Labeling Using Radar Data of Human Gait

Abstract

Estimating the number of clusters in an observed data set poses a major challenge in cluster analysis. In the literature, the original Bayesian Information Criterion (BIC) is used as a criterion for cluster enumeration. However, the original BIC is a generic approach that does not take the data structure of the clustering problem into account. Recently, a new BIC for cluster analysis has been derived from first principles by treating the cluster enumeration problem as a maximization of the posterior probability of candidate models given data. Based on the new BIC for cluster analysis, we propose a target enumeration and labeling algorithm. The proposed algorithm is unsupervised in the sense that it requires neither knowledge on the number of clusters nor training data. Experimental results based on real radar data of human gait show that the proposed method is able to correctly estimate the number of observed persons and, at the same time, provide labels to them with high accuracy. It is shown that, in terms of cluster enumeration performance, the proposed algorithm outperforms an existing cluster enumeration method.