Catálogo de publicaciones - libros

Knowledge Discovery from Databases (KDD) – also named Data Mining – is a growing field since 10 years which combines techniques from databases, statistics, and machine learning. Applications of KDD most often have one of the following :

– Customer relationship management: who are the best customers, which products are to be offered to which customers (direct marketing or customer acquisition), which customers are likely to end the relationship (customer churn), which customers are likely to not pay (also coined as fraud detection)?

– Decision support applies to almost all areas, ranging from medicine over marketing to logistics. KDD applications aim at a data-driven justification of decisions by relating actions and outcomes.

– Recommender systems rank objects according to user profiles. The objects can be, for instance, products as in the amazon internet shop, or documents as in learning search engines. KDD applications do not assume user profiles to be given but learns tehm from observations of user behavior.

– Plant asset management moves beyond job scheduling and quality control. The goal is to optimize the overall benefits of production.

This paper traces the history of spoken language communication with computers, from the first attempts in the 1950s, through the establishment of the theoretical foundations in the 1980s, to the incremental improvement phase of the 1990s and 2000s. Then a perspective is given on the current conversational technology market and industry, with an analysis of its business value and commercial models.

Autonomous robot guidance in dynamic environments requires, on the one hand, the study of relative motion of the objects of the environment with respect to the robot, and on the other hand, the analysis of the depth towards those objects. In this paper, a stereo vision method, which combines both topics with potential utility in robot navigation, is proposed. The goal of the stereo vision model is to calculate depth of surrounding objects by measuring the disparity between the two-dimensional imaged positions of the object points in a stereo pair of images. The simulated robot guidance algorithm proposed starts from the motion analysis that occurs in the scene and then establishes correspondences and analyzes the depth of the objects. Once these steps have been performed, the next step is to induce the robot to take the direction where objects are more distant in order to avoid obstacles.

Active contours based tracking methods have widely used for object tracking due to their following advantages. 1) effectiveness to descript complex object boundary, and 2) ability to track the dynamic object boundary. However their tracking results are very sensitive to location of the initial curve. Initial curve far form the object induces more heavy computational cost, low accuracy of results, as well as missing the highly active object. Therefore, this paper presents an object tracking method using a mean shift algorithm and active contours. The proposed method consists of two steps: object localization and object extraction. In the first step, the object location is estimated using mean shift. And the second step, at the location, evolves the initial curve using an active contour model. To assess the effectiveness of the proposed method, it is applied to synthetic sequences and real image sequences which include moving objects.

In this paper, we propose a place recognition system which recognize places from a large set of images obtained over time. A set or a sequence of images provides more information about the places and that can be used for more robust recognition. For this, the proposed system recognize places using density matching between the estimated density of the input set and density of the stored images for each place. In the proposed system, we use global texture feature vector for image representation and their density for place recognition. We use a method based on a Gaussian model of texture vector distribution and a matching criterion using the Kullback-Leibler divergence measure. In the experiment, the system successfully recognized the places in several image sequence, the success rate of place recognition was 87% on average.

Detecting, locating, and tracking people in a dynamic environment is important in many applications, ranging from security and environmental surveillance to assistance to people in domestic environments, to the analysis of human activities. To this end, several methods for tracking people have been developed using monocular cameras, stereo sensors, and radio frequency tags.

In this paper we describe a real-time People Localization and Tracking (PLT) System, based on a calibrated fixed stereo vision sensor. The system analyzes three interconnected representations of the stereo data (the left intensity image, the disparity image, and the 3-D world locations of measured points) to dynamically update a model of the background; extract foreground objects, such as people and rearranged furniture; track their positions in the world.

The system can detect and track people moving in an area approximately 3 x 8 meters in front of the sensor with high reliability and good precision.

In this paper, we introduce a new feature representation method for face recognition. The proposed method, referred as Kernel ICA, combines the strengths of the Kernel and Independent Component Analysis approaches. For performing Kernel ICA, we employ an algorithm developed by F. R. Bach and M. I. Jordan. This algorithm has proven successful for separating randomly mixed auditory signals, but it has never been applied on bidimensional signals such as images. We compare the performance of Kernel ICA with classical algorithms such as PCA and ICA within the context of appearance-based face recognition problem using the FERET database. Experimental results show that both Kernel ICA and ICA representations are superior to representations based on PCA for recognizing faces across days and changes in expressions.

Many different path planning methods have been proposed over recent years, although there are only a few that deal with computer vision techniques. In this work we implement a path planning algorithm which takes into account a vision processing system. Thus, we develop a method that uses Mathematical Morphology to provide near-optimal paths throughout an environment. The experiments show that our path planning algorithm is able to locate good solution paths after a training process, which is necessary to fix some parameters. This will make possible its adaptation to a practical robot system.

We present an example-based learning approach for detecting a partially occluded human face in a scene provided by a camera of Automated Teller Machine (ATM) in a bank. Gradient mapping in scale space is applied on an original image, providing human face representation robust to illumination variance. Detection of the partially occluded face, which can be used in characterization of suspicious ATM users, is then performed based on Support Vector Machine (SVM) method. Experimental results show that a high detection rate over 95% is achieved in image samples acquired from in-use ATM.

Colour, shape, geometry, contrast, irregularity and roughness of the visual appearance of vocal cords are the main visual features used by a physician to diagnose laryngeal diseases. This type of examination is rather subjective and to a great extent depends on physician’s experience. A decision support system for automated analysis of vocal cord images, created exploiting numerous vocal cord images can be a valuable tool enabling increased reliability of the analysis, and decreased intra- and inter-observer variability. This paper is concerned with such a system for analysis of vocal cord images. Colour, texture, and geometrical features are used to extract relevant information. A committee of artificial neural networks is then employed for performing the categorization of vocal cord images into , , and classes. A correct classification rate of over 93% was obtained when testing the system on 785 vocal cord images.