Catálogo de publicaciones - libros

This paper proposes a method for computing the medial axis transform (MAT) or the skeleton of a general 2D shape using a technique with a high performance, based on a distance transform computation from the shape’s boundaries. The distance transform is computed propagating a wavefront from the boundary, and the skeleton is obtained detecting the points where the wavefronts collide themselves, and applying connectivity rules during the process. This method has two main advantages: the efficiency and the preservation of the skeleton properties.

This paper describes a face detection system conceived to process video streams in real-time. Cue combination allows the system to tackle the temporal restrictions achieving a notable detection rate. The system developed is able to detect simultaneously at different resolutions multiple individuals building a feature based model for each detected face.

In recent years, extensive work has been done to design algorithms that strive to mimic the robust human vision system which is able to perceive the true colors and discount the illuminant from a scene viewed under light having different spectral compositions (the feature is called “color constancy”). We propose a straightforward approach to the color constancy problem by employing an Interactive Genetic Algorithm [1] (e.g. a Genetic Algorithm [2], [3] guided by the user) that optimizes a well known and robust variant of color constancy algorithm called “gamut mapping” [4]. Results obtained on a set of test images and comparison to various color constancy algorithms, show that our method achieves a good color constancy behavior with no additional knowledge required besides the image that is to be color-corrected, and with minimal assumptions about the scene captured in the image.

Airbags have been saved thousands of lives and reduced the number of serious injuries from collisions. However, the car occupant can be often hurt, or killed, by the airbag itself. For reducing the risk caused by airbag, designing a smart airbag is an important issue. This paper presents a vision based automatic system that can control triggering and intensity of airbag deployment. The system consists of an occupant classification system and an occupant pose recognition system, by which we aim to control whether the airbag should be triggered or not, and how strongly it should be deployed when it is triggered. Results suggest that the system is feasible as a vision based airbag controller.

A mathematical characterization of the membrane potential as an instantaneous return process in the presence of random refractoriness is investigated for the Wiener neuronal model. In the case of constant refractoriness, simple closed form expressions are obtained.

As a model of Brownian motor, we consider the motion of particles in an asymmetric, single-well, periodic potential undergoing half-period shifts driven by two Poisson processes. Probability currents and stopping force are explicitly obtained as a function of the model parameters, and use of the notion of driving effective potential is made to bridge the present model with our previous works involving washboard potentials.

A network consisting of two Stein-type neuronal units is analyzed under the assumption of a complete interaction between the neurons. The firing of each neuron causes a jump of constant amplitude of the membrane potential of the other neuron. The jump is positive or negative according to whether the firing neuron is excitatory or inhibitory.

Making use of a simulation procedure designed by ourselves, we study the interspike intervals of the two neurons by means of their histograms, of some descriptive statistics and of empirical distribution functions. Furthermore, via the crosscorrelation function, we investigate the synchronization between the neurons firing activity in the special case when one neuron is excitatory and the other is inhibitory.

For a class of stationary Gaussian processes and for large correlation times, the asymptotic behavior of the upcrossing first passage time probability densities is investigated. Parallel simulations of sample paths of special stationary Gaussian processes for large correlations times provide a statistical validation of the theoretical results.

Convergence is a central problem in both computer science and in population biology.

Will a program terminate? Will a population go to an equilibrium?

In general these questions are quite difficult – even unsolvable.

In this paper we will concentrate on very simple iterations of the form

= ()

where each is simply a real number and () is a reasonable real function with a single fixed point. For such a system, we say that an iteration is “globally stable” if it approaches the fixed point for all starting points. We will show that there is a simple method which assures global stability. Our method uses bounding of () by a self-inverse function. We call this bounding “enveloping” and we show that For a number of standard population models, we show that local stability implies enveloping by a self-inverse linear fractional function and hence global stability. We close with some remarks on extensions and limitations of our method.

Ultrasonography plays a crucial role in the diagnosis of breast cancer. However, it is one of the most difficult types of images to segment and analyze. The presence of speckle noise and low contrast areas limits the success of most noise reduction filters and segmentation algorithms. In this paper, we propose a combination of different techniques which provide quite satisfactory results in the segmentation of breast tumors on ultrasonography. It is performed in a semiautomatic way, which eliminates the need for a manual delineation of the contour of the nodules. These techniques include the truncated median filter, a region-growing algorithm and active contours. Furthermore, this can be the initial phase for an exhaustive analysis of the diagnostic criteria in breast ultrasound.