Robotic Color Image Segmentation by Means of Finite Mixture Models

Abstract— Image segmentation for robots requires to be fast, in order to deal with ever more powerful processors. Moreover, it is assumed to be robust to environmental changes, such as light conditions. In this paper we propose the application of a couple of unsupervised learning algorithms for the estimation of the number of components and the parameters of a mixture model for image segmentation. These serve for the unsupervised identification of multiple different objects in a visual scene, such as for a subsequent localization and tracking. We compare our previous technique against the new one. The distinctive aspect of our new approach is related to a top-down hierarchical search for the number of components by means of a binary tree decision structure. This work analyzes both approaches, two previous work of ours, in terms of applicability to object detection for robotic applications. Besides, we propose the computational burden evaluation for the two algorithms. Index Terms Robotics, Computer Vision, Object Segmentation, Unsupervised Learning, Self-Adapting Expectation Maximization