Catálogo de publicaciones - libros

People detection and tracking are essential capabilities in human-robot interaction. However, the development of these tasks is specially difficult in cluttered environments where it is not possible to create a background model because of the robot movement. To detect and track people in a scene the use of vision sensors is convenient in order to distinguish people from other objects with similar shapes. This paper presents a novel approach for person tracking which combines depth, color and gradient information based on stereo vision. The degree of confidence assigned to depth information in the tracking process varies according to the amount of it found in the disparity map. A novel confidence measure is defined for it. To test the validity of our proposal, it is evaluated in several color-with-depth sequences where people interact in complex situations.

We present a method for recovering facial shape using an image of a face and a reference model. The zenith angle of the surface normal is recovered directly from the intensities of the image. The azimuth angle of the reference model is then combined with the calculated zenith angle in order to get a new field of surface normals. After integration of the needle map, the recovered surface has the effect of mapped facial features over the reference model. Experiments demonstrate that for the lambertian case, surface recovery is achieved with high accuracy. For non-Lambertian cases, experiments suggest potential for face recognition applications.

This paper proposes a faces verification in which the feature extraction is carried out using the discrete Gabor function (DGF), while the Gaussian Mixture Model (GMM) is used in the face verification stage. Evaluation results using standard data bases with different parameters, such as the number of mixtures and the number of face used for training show that proposed system provides better results that other proposed systems with a correct verification rate larger than 95%. Although, as happens in must face recognition systems, the verification rate decreases when the target faces present some rotation degrees.

In this work we assume a frontal view of a face for the lips shape extraction, then the first step is locate a face inside a digital image, for this task we use techniques based in color to extract only the pixels with skin tone, a templates based in integral projections are applied to verify and locate the face, using integral projections, we locate and define a region of interest for lips. Previously a statistical model of lips (ASM) was created in the same way, local appearance patterns of landmarks are modeled using Local Binary Patterns (LBP) , in this model we try to capture a variation from a closed lips to an opened lips. For the search task Local Binary Pattern Histogram (LBPH) are used.

This paper presents a novel approach for continuous gesture recognition using depth range sensors. Our approach can be seen as an extension of Motion Templates [1] using multiple layers that register the three-dimensional nature of the human gestures. Our Multi-Layered templates are created using , the extension of binary silhouettes when depth information is available. Both the original Motion Templates and our extension have been tested using several classification approaches in order to determine the best one. These approaches include the use of Hu-moments (originally employed in [1]), PCA and Support Vector Machines. Finally, we propose a methodology for creating a continuous gesture recogniser using motion templates. The methodology is applied both to our representation approach and to the original proposal. In order to validate our proposal, several stereo-video sequences have been recorded showing eight people performing a total of ten different gestures that are prone to be confused when monocular vision is used. The conducted experiments show that our proposal performs a 20% better than the original method.

This work studies the interaction of the nonholonomic and visibility constraints of a robot to maintain visibility of a landmark. The robot is a differential drive system (nonholonomic robot) and has a sensor with limited capabilities (limited field of view). In this research, we want to determine whether or not a robot can always maintain visibility of a landmark during the execution of a path between any two locations. We present two kinematic models. First, a robot with 3 controls, where the controls correspond to the two wheels velocities plus one independent controlled sensor. Second, a model with only 2 controls, which controls both the wheels and the sensor rotation. We show that our system (with 3 or 2 controls) is small-time local controllable.

In this paper we utilize information theory to study the impact in learning performance of various motivation and environmental configurations. This study is done within the context of an evolutionary fuzzy motivation based approach used for acquiring behaviors in mobile robot exploration of complex environments. Our robot makes use of a neural network to evaluate measurements from its sensors in order to establish its next behavior. Adaptive learning, fuzzy based fitness and Action-based Environment Modeling (AEM) are integrated and applied toward training the robot. Using information theory we determine the conditions that lead the robot toward highly fit behaviors. The research performed also shows that information theory is a useful tool in analyzing robotic training methods.

Among various document clustering algorithms that have been proposed so far, the most useful are those that automatically reveal the number of clusters and assign each target document to exactly one cluster. However, in many real situations, there not exists an exact boundary between different clusters. In this work, we introduce a fuzzy version of the MajorClust algorithm. The proposed clustering method assigns documents to more than one category by taking into account a membership function for both, edges and nodes of the corresponding underlying graph. Thus, the clustering problem is formulated in terms of weighted fuzzy graphs. The fuzzy approach permits to decrease some negative effects which appear in clustering of large-sized corpora with noisy data.

A problem of supervised approaches for text classification is that they commonly require high-quality training data to construct an accurate classifier. Unfortunately, in many real-world applications the training sets are extremely small and present imbalanced class distributions. In order to confront these problems, this paper proposes a novel approach for text classification that combines under-sampling with a semi-supervised learning method. In particular, the proposed semi-supervised method is specially suited to work with very few training examples and considers the automatic extraction of untagged data from the Web. Experimental results on a subset of Reuters-21578 text collection indicate that the proposed approach can be a practical solution for dealing with the class-imbalance problem, since it allows achieving very good results using very small training sets.

The writing style of an author is a phenomenon that computer scientists and stylometrists have modeled in the past with some success. However, due to the complexity and variability of writing styles, simple models often break down when faced with real world data. Thus, current trends in stylometry often employ hundreds of features in building classifier systems. In this paper, we present a novel set of synonym-based features for author recognition. We outline a basic model of how synonyms relate to an author’s identify and then build an additional two models refined to meet real world needs. Experiments show strong correlation between the presented metric and the writing style of four authors with the second of the three models outperforming the others. As modern stylometric classifier systems demand increasingly larger feature sets, this new set of synonym-based features will serve to fill this ever-increasing need.