Uncertain Visual Target Tracking by Hierarchical Combination of Multiple Target State Space Model and Self Organizing Map

Abstract

A novel visual target tracking method, where targets to be tracked are uncertain as they are not pre-determined, has been proposed in a framework of multiple target tracking formulation with Random Finite Set (RFS) and Probability Hypothesis Density (PHD) filter with its Sequential Monte Carlo (SMC) implementation. Self Organizing Map (SOM) and its learning algorithm have been combined to the framework as a post-process of the state estimation by SMC-PHD filter in order to classify the unlabelled set of particles, i.e. state estimation result, into structured knowledge of the scene. Synthetic and real video image experiments demonstrate preliminary results of the proposed method.