Catálogo de publicaciones - libros

Local space-time features can be used to detect and characterize motion events in video. Such features are valid for recognizing motion patterns, by defining a vocabulary of , and representing each video sequence by means of a histogram, in terms of such vocabulary. In this paper, we propose a supervised vocabulary computation technique which is based on the prior classification of the training events into classes, where each class corresponds to a human action. We will compare the performance of our method with the global approach to show that not only does our method obtain better results but it is also computationally less expensive.

In this work, we present a fast stochastic context-free parsing algorithm that is based on a stochastic version of the Valiant algorithm. First, the problem of computing the string probability is reduced to a transitive closure problem. Then, the closure problem is reduced to a matrix multiplication problem of matrices of a special type. Afterwards, some fast algorithm can be used to solve the matrix multiplication problem. Preliminary experiments show that, in practice, an important time savings can be obtained.

Texture identification can be a key component in Content Based Image Recognition systems. Although formal definitions of texture vary in the literature, it is commonly accepted that textures are naturally extracted and recognized as such by the human visual system, and that this analysis is performed in the frequency domain. The method presented here employs a discrete Fourier transform in the polar space to extract features, which are then classified with a vector quantizer for supervised segmentation of images into texture regions. Experiments are conducted on a standard database of test problems that show this method compares favorably with the state-of-the-art and improves over previously proposed frequency-based methods.

The Distance Transform is a powerful tool that has been used in many computer vision tasks. In this paper, the use of relevant maxima in distance transform’s medial axis is proposed as a method for fast image data reduction. These disc-shaped maxima include morphological information from the object they belong to, and because maxima are located inside homogeneous regions, they also sum up chromatic information from the pixels they represent. Thus, maxima can be used instead of single pixels in algorithms which compute relations among pixels, effectively reducing computation times. As an example, a fast method for color image segmentation is proposed, which can also be used for textured zones detection. Comparisons with mean shift segmentation algorithm are shown.

We here compare the performance (predictive accuracy and processing time) of different neural network ensembles with that of nearest neighbor classifier ensembles. Concerning the connectionist models, the multilayer perceptron and the modular neural network are employed. Experiments on several real-problem data sets demonstrate a certain superiority of the nearest-neighbor-based schemes, in terms of both accuracy and computing time. When comparing the neural network ensembles, one can observe a better behavior of the multilayer perceptron than that of the modular networks.

Robust and accurate people tracking is a key task in many promising computer-vision applications. One must deal with non-rigid targets in open-world scenarios, whose shape and appearance evolve over time. Targets may interact, causing partial or complete occlusions. This paper improves tracking by means of particle filtering, where occlusions are handled considering the target’s predicted trajectories. Model drift is tackled by careful updating, based on the history of likelihood measures. A colour-based likelihood, computed from histogram similarity, is used. Experiments are carried out using sequences from the CAVIAR database.

The last advances on multiple kernel tracking consider the kernels as estimators of target features. The state space of the target is defined by the individual state space of these features.

The aim of this work is to construct an algorithm robust against three dimensional rotations and partial occlusions. For this purpose, we take as the state space the two dimensional position of the features and an indicator of occlusions. We extract the three dimensional structure of the target from the first tracked frames and estimate the projection of this structure on each frame. By using this information, we are able to predict the position of a feature even when the kernel provides a wrong estimation, for example during an occlusion. The experimental results showed a good performance correcting errors and in presence of partial occlusions.

Multivariate imaging technologies such as Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) have recently gained an important attention as it improves tumour detection. Modelling of contrast media uptake and washout kinetic parameters which are closely related to physiological and anatomical features helps to diagnose and detect a possible cancer. One issue that does not generally receive much attention is the process of detecting regions of interest (ROIs). An automatic region-of-interest (ROI) selection method is presented to avoid the time consuming and subjective task of manual ROI selection, which significantly affects reproducibility and accuracy of measurements.

The driver’s loss of attention is an important problem in which are spent considerable research efforts in different areas such as psychology, automobile technology, computer vision and driving assistance. We use here a simple algorithm based on rigid-body and motion detection. This scheme efficiently segments moving objects using the visual field of the driver’s rear-view mirror. The overtaking scene in the rear-view mirror is distorted due to perspective, making it difficult to detect the overtaking car. Thus we propose two alternative methods to deal with this problem and compare the results in different overtaking sequences.

Dry-cured Iberian pig products are some of the most valuables meat products in Spain. Visually discernible features of fat and lean, such as marbling, have an effect on the acceptability of these products. Marbling properties include the amount and spatial distribution of intramuscular fat streaks. Thresholding techniques are the simplest and most widely used for image segmentation, but pay no attention to all spatial information on the images. In this paper we propose another method to evaluate fat level in loin and make a three-dimensional visualization with only the isolated fat in order to give knowledge about fat distribution, and about width of fat streaks, helping the experts in food science to analyze marbling. In order for fat isolation we apply a method based on a pyramidal decomposition of images combined with region-growing techniques. 3D reconstruction is obtained by marching cubes algorithm.