Catálogo de publicaciones - libros

This paper describes an ongoing project that develops a set of generic components to help humans (semi-automatic system) in surveillance and security tasks in several scenarios. These components are based in the computational model of a set of selective and Active VISual Attention mechanisms with learning capacity () and in the superposition of an “intelligence” layer that incorporates the knowledge of human experts in security tasks. The project described integrates the responses of these alert mechanisms in the synthesis of the three basic subtasks present in any surveillance and security activity: real-time monitoring, situation diagnosing, and action planning and control. In order to augment the diversity of environments and situations where system may be used, as well as its efficiency as support to surveillance tasks, knowledge components derived from situating cameras on mobile platforms are also developed.

Attention is a mechanism present in most of the more complex and developed living beings. It is responsible for much of their ability to operate in real time in unstructured and dynamic environments with a limited amount of processing resources. In this paper, an architecture for developing attentional functions in agents is presented. This architecture is based on the concept of attentor and it allows for the real time adaptation to the environment and tasks to be performed in a natural manner. One of the main requirements of the system was the ability to handle different sensorial varieties and attentional streams in a transparent manner while, at the same time, being able to progressively create more complex attentional structures. The main characteristics of the architecture are presented through its implementation in a real robot.

The computer model of visual attention derives an interest or saliency map from an input image in a process that encompasses several data combination steps. While several combination strategies are possible, not all perform equally well. This paper compares main cue combination strategies by measuring the performance of the considered models with respect to human eye movements. Six main combination methods are compared in experiments involving the viewing of 40 images by 20 observers. Similarity is evaluated qualitatively by visual tests and quantitatively by use of a similarity score. The study provides insight into the map combination mechanisms and proposes in this respect an overall optimal strategy for a computer saliency model.

Many researchers have explored the relationship between recurrent neural networks and finite state machines. Finite state machines constitute the best characterized computational model, whereas artificial neural networks have become a very successful tool for modeling and problem solving. Recently, the neurally-inspired algorithmic lateral inhibition (ALI) method and its application to the motion detection task have been introduced. The article shows how to implement the tasks directly related to ALI in motion detection by means of a formal model described as finite state machines. Automata modeling is the first step towards real-time implementation by FPGAs and programming of ”intelligent” camera processors.

This paper is intended to address three of the most common problems arising in the field of non-rigid biomedical image registration. Firstly, a regularization term based on fractional-order derivatives is proposed. It can be seen as a generalization of the bio-inspired diffusion and curvature smoothing terms but, since it incorporates features of both regularizers, with this approach it is possible to obtain better registration results from a variational point of view. Next, a frequency-domain formulation for the image registration problem is presented. It provides efficient and stable implementations for the considered registration techniques. Finally, a two-step method is proposed for obtaining the optimal values of the registration parameters, because in the literature there is no agreement about which are the optimal values for these parameters, leading the authors to arbitrarily fix them. The resulting registration scheme, after the incorporation of these three strategies, is tested on two biomedical imaging scenarios.

Artificial muscles are bio-inspired devices formed by several layers of conducting polymers. These devices have the ability of transform electrical energy into mechanical energy through an electrochemical reaction, which is produced by an oxidation or reduction of the polymer due to an electric current. Since the device have a strip shape, this reaction results in a macroscopic swelling and shrinking movement. This movement is similar to the biological muscles and it has several applications as motor prostheses and as part of complex biomaterials. In this paper we describe a computer vision system developed to analyze and characterize these devices through their cycle of life. The method includes cameras for tracking the movement of the muscle from different angles and a set of algorithms to characterize the motion of the device through its use. By means of active contours it is determined the instantaneous position of the muscle in the space. From these contours other parameters like the parametric motion and energy of curvature are calculated. These data are compared with the physical parameters of the device, like the tension and energy consumption, providing a way for performing automatic testing on the research of artificial muscles.

In this paper, an implementation of the algorithm for segmenting stereoscopic video sequences is shown. This algorithm is an essential task in the method in order to obtain a 3D characterization of artificial muscles. Image sequences are acquired by a two-cam computer vision system. Optimal and efficient segmentation of these images is our goal; information obtained from the segmented first frame of the video sequence is used for segmenting the next frame and so on. Redundancy between stereoscopic pairs of images is also used to optimize the segmentation. In this paper, the algorithm is described and our own specific implementation is addressed. Particular problems of stereoscopic video segmentation are shown and how they are solved. Finally, results yielded from simulations are presented and conclusions close the paper.

Human learning inspires a large amount of algorithms and techniques to solve problems in image understanding. Supervised learning algorithms based on support vector machines are currently one of the most effective methods in machine learning. A support vector approach is used in this paper to solve a typical problem in image registration, this is, the joint probability density function estimation needed in the image registration by maximization of mutual information. Results estimating the joint probability density function for two CT and PET images demonstrate the proposed approach advantages over the classical histogram estimation.

In real sequences, one of the factors that most negatively affects the segmentation process result is the existence of scene noise. This impairs object segmentation which has to be corrected if we wish to have some minimum guarantees of success in the following tracking or classification stages. In this work we propose a generic knowledge-based model to improve the segmentation process. Specifically, the model uses a decomposition strategy in description levels to enable the feedback of information between adjacent levels. Finally, two case studies are proposed that instantiate the model proposed for detecting humans.

This article presents a knowledge-based application to study and analyze traffic behavior on major roads, using as the main surveillance artefact a video camera mounted on a relatively high place with a significant image analysis field. The system described presents something new which is the combination of both traditional traffic monitoring systems, that is, monitoring to get information on different traffic parameters and monitoring to detect accidents automatically. Therefore, we present a system in charge of compiling information on different traffic parameters. It also has a surveillance module, which can detect a wide range of the most significant incidents on a freeway or highway.