Showing papers by "Alexandre Bernardino published in 2010"

Self-adaptive Gaussian mixture models for real-time video segmentation and background subtraction

Abstract: The usage of Gaussian mixture models for video segmentation has been widely adopted. However, the main difficulty arises in choosing the best model complexity. High complex models can describe the scene accurately, but they come with a high computational requirements, too. Low complex models promote segmentation speed, with the drawback of a less exhaustive description. In this paper we propose an algorithm that first learns a description mixture for the first video frames, and then it uses these results as a starting point for the analysis of the further frames. Then, we apply it to a video sequence and show its effectiveness for real-time tracking multiple moving objects. Moreover, we integrated this procedure into a foreground/background subtraction statistical framework. We compare our procedure against the state-of-the-art alternatives, and we show both its initialization efficacy and its improved segmentation performance.

Gaussian mixture models for affordance learning using Bayesian Networks

Abstract: Affordances are fundamental descriptors of relationships between actions, objects and effects. They provide the means whereby a robot can predict effects, recognize actions, select objects and plan its behavior according to desired goals. This paper approaches the problem of an embodied agent exploring the world and learning these affordances autonomously from its sensory experiences. Models exist for learning the structure and the parameters of a Bayesian Network encoding this knowledge. Although Bayesian Networks are capable of dealing with uncertainty and redundancy, previous work considered complete observability of the discrete sensory data, which may lead to hard errors in the presence of noise. In this paper we consider a probabilistic representation of the sensors by Gaussian Mixture Models (GMMs) and explicitly taking into account the probability distribution contained in each discrete affordance concept, which can lead to a more correct learning.

Sensor-based self-calibration of the iCub's head

Abstract: In this paper we propose techniques for the calibration of the iCub's stereo head using vision and inertial measurements. Given that wear and tear can change the geometrical relationship between the different elements in the kinematic chain, new calibrations must be performed periodically. We propose methods that allow automatic calibration without the need for using external sensors or specially designed calibration objects. The methods can be applied at any time during the operation of the system, thus being an alternative for systems whose calibrations are imprecise or that require frequent recalibration. Results are shown both in simulations and on the iCub's stereo head.

An Algorithm for the Least Square-Fitting of Ellipses

Abstract: In this paper we propose a new algorithm for the least square fitting of ellipses from scattered data. Originally based on the one proposed by Fitzgibbon et Al in 1999, our procedure is able to overcome the numerical instability of that algorithm. We test our approach versus the latter and another approach with different ellipses. Then, we present and discuss our results.

A computational model of object affordances

Abstract: The concept of object affordances describes the possible ways whereby an agent (either biological or artificial) can act upon an object. By observing the effects of actions on objects with certain properties, the agent can acquire an internal representation of the way the world functions with respect to its own motor and perceptual skills. Thus, affordances encode knowledge about the relationships between action and effects lying at the core of high-level cognitive skills such as planning, recognition, prediction and imitation. Humans learn and exploit object affordances through their entire lifespan, by either autonomous exploration of the world or social interaction. Building on a biological motivation and aiming at the development of adaptive robotic systems, we propose a computational model capable of encoding object affordances during exploratory learning trials. We represent this knowledge as a Bayesian network and rely on statistical learning and inference methods to generate and explore the network, efficiently dealing with uncertainty, redundancy and irrelevant information. The affordance model serves as base for an imitation learning framework, which exploits the recognition and planning capabilities to learn new tasks from demonstrations. We show the application of our model in a real-world task in which a humanoid robot interacts with objects, uses the acquired knowledge and learns from demonstrations. Results illustrate the success of our approach in learning object affordances and generating complex cognitive behavior.

Random features vs Harris Corners in Real-Time Visual Egomotion Estimation

Abstract: We compare Randomly Selected (RanSel) features with Harris Corners within a visual egomotion estimation framework. Harris corners have been extensivelly used in visual egomotion estimation systems due to a good tracking stability. However, to compute these features the whole image has to be processed. Instead, we propose the use of randomly selected points which are virtually costless to obtai n. Despite tracking individual RanSel features is not as stable as Harris corners, we show that, when integrated in a time-filtering scheme, they provide similar results at a muc h faster rate. We have performed experiments using a synthetic setup with ground-truth and discuss the advantages of using RanSel features.

Image segmentation for robots: fast self-adapting gaussian mixture model

Abstract: Image segmentation is a critical low-level visual routine for robot perception. However, most image segmentation approaches are still too slow to allow real-time robot operation. In this paper we explore a new method for image segmentation based on the expectation maximization algorithm applied to Gaussian Mixtures. Our approach is fully automatic in the choice of the number of mixture components, the initialization parameters and the stopping criterion. The rationale is to start with a single Gaussian in the mixture, covering the whole data set, and split it incrementally during expectation maximization steps until a good data likelihood is reached. Singe the method starts with a single Gaussian, it is more computationally efficient that others, especially in the initial steps. We show the effectiveness of the method in a series of simulated experiments both with synthetic and real images, including experiments with the iCub humanoid robot.

A practical method for self-adapting gaussian expectation maximization

Abstract: Split-and-merge techniques have been demonstrated to be effective in overtaking the convergence problems in classical EM. In this paper we follow a split-and-merge approach and we propose a new EM algorithm that makes use of a on-line variable number of mixture Gaussians components. We introduce a measure of the similarities to decide when to merge components. A set of adaptive thresholds keeps the number of mixture components close to optimal values. For sake of computational burden, our algorithm starts with a low initial number of Gaussians, adjusting it in runtime, if necessary. We show the effectivity of the method in a series of simulated experiments. Additionally, we illustrate the convergence rates of of the proposed algorithms with respect to the classical EM.

Unsupervised Greedy Learning of Finite Mixture Models

Abstract: This work deals with a new technique for the estimation of the parameters and number of components in a finite mixture model. The learning procedure is performed by means of a expectation maximization (EM) methodology. The key feature of our approach is related to a top-down hierarchical search for the number of components, together with the integration of the model selection criterion within a modified EM procedure, used for the learning the mixture parameters. We start with a single component covering the whole data set. Then new components are added and optimized to best cover the data. The process is recursive and builds a binary tree like structure that effectively explores the search space. We show that our approach is faster that state-of-the- art alternatives, is insensitive to initialization, and has better data fits in average. We elucidate this through a series of experiments, both with synthetic and real data.

Surveillance with Pan-Tilt Cameras: Background Modeling

Abstract: In this paper we propose a methodology for mini- mizing the variance of a cube (mosaicked) representation of a scene imaged by a pan-tilt camera. The minimization is based on the estimation of the vignetting image distortion, using the pan and tilt degrees of freedom instead of color calibrating patterns. Experiments with real images show that variance minimization is effective for improving event detection.