A novel approach to generate multiple shape models for tracking applications

Abstract

Many proposals to generate shape models for tracking applications are based on a linear shape model, and a constraint that delimits the parameter values which generate feasible shapes. In this paper we introduce a novel approach to generate such models automatically. Given a training set, we determine the linear shape model as classical approaches, and model its associated constraint using a Gaussian Mixture Model, which is fully parameterized by a presented algorithm. Then, from this model we generate a collection of linear shape models of lower dimensionality, each one constrained by a single Gaussian model. This set of models represents better the training set, reducing the computational cost of tracking applications. To compare our proposal with the usual one, a comparison measure is defined, based on the Bayesian Information Criterion. Both modeling strategies are analyzed in a pedestrian tracking application, where our proposal claims to be more appropriate.