A Genetic Algorithm for the construction of optimized covariance descriptors

The problem of real-time tracking has been studied widely and many methods in very different fields of application have been developed manipulating image based elements. While all use features as a way to represent a tracked object in the image, naturally, depending on the method and the objects, some features are better than others. As part of the project presented in [1], the goal of this paper is to provide efficient descriptors to perform real-time tracking of children. Covariance descriptors are a common and convenient way to describe an object, since they compile in a single matrix several features and also their statistical interrelationships. This paper introduces a Genetic Algorithm as a way to seek the best combination among a list of features for describing a selected object in a video sequence. The implemented Genetic Algorithm is a Niched Pareto Genetic Algorithm (NPGA), and two different methods of selection/reproduction have been compared; a regular method and one based on a High Elitism process. Reliable results are obtained, since the features combined seem to match the tracked object characteristics, but dissimilarities between the two methods are also highlighted. In the end, this paper doesn't focus on the performances of the GAs themselves, but it proposes a Genetic Algorithm as a way of solving a dictionary learning problem.