Catálogo de publicaciones - libros

The visual tracking of image regions is a research area of great interest within the computer vision community. One issue which has received quite attention in the last years has been the analysis of tracking algorithms which could be able to cope with changes in the appearance of the target region. Probably one of the most studied techniques proposed to model this appearance variability is that based on linear subspace models. Recently, efficient algorithms for fitting these models have been developed too, in many cases as an evolution of well studied approaches for the tracking of fixed appearance images.

Additionally, new methods based on second order optimizers have been proposed for the tracking of targets with no appearance changes. In this paper we study the application of such techniques in the design of tracking algorithms for linear appearance models and compare their performance with three previous approaches. The achieved results show the efficiency of the use of second-order minimization in terms of both number of iterations required for convergence and convergence frequency.

We present an algorithm for interpolating the visible portions of a point cloud that are sampled from opaque objects in the environment. Our algorithm projects point clouds onto a sphere centered at the observing locations and performs essentially non-oscillatory (ENO) interpolation to the projected data. Curvatures of the occluding objects can be approximated and used in many ways. We show how this algorithm can be incorporated into novel algorithms for mapping an unknown environment.

In adjusting for discrepancies between adjacent airborne laser scanning (ALS) data strips, previous studies generally used conjugate features such as points, lines, and surface objects; however, irrespective of the types of features employed, the adjustment process relies upon the existence of suitable conjugate features within the overlapping area and the ability of the employed method to detect and extract the features. These limitations make the process complex and sometimes limit the applicability of developed methodologies because of a lack of suitable features in overlapping areas. To address these problems, this paper presents a methodology that uses the topological characteristics of the terrain itself, which is represented by a contour tree (CT). This approach provides a robust methodology without the restrictions involved in methods that employ conjugate features. Our method also makes the overall process of adjustment generally applicable and automated.

Most existing cognitive architectures integrate computer vision and symbolic reasoning. However, there is still a gap between low-level scene representations (signals) and abstract symbols. Manually attaching, i.e. grounding, the symbols on the physical context makes it impossible to expand system capabilities by learning new concepts. This paper presents a visual bootstrapping approach for the unsupervised symbol grounding. The method is based on a recursive clustering of a perceptual category domain controlled by goal acquisition from the visual environment. The novelty of the method consists in division of goals into the classes of parameter goal, invariant goal and context goal. The proposed system exhibits incremental learning in such a manner as to allow effective transferable representation of high-level concepts.

An image-based 3D model acquisition system using projections of level curves is presented. The basic idea is similar to surface from parallel planar contours, such as 3D reconstruction from CT or laser range scanning techniques. However, our approach is implemented on a low-cost passive camera system. The object is placed in a water container and the level curves of the object’s surface are generated by raising the water level. The 3D surface is recovered by multiple 2D projections of parallel level curves and the camera parameters. Experimental results are presented for both computer simulated data and real image sequences.

A robust registration of 3D-point data and a triangle mesh of the corresponding 3D-structure is presented, where the acquired 3D-point data may be noisy, may include outliers and may have wrong scale. Furthermore, in this approach it is not required to have a good initial match so the 3D-point cloud and the according triangle mesh may be loosely positioned in space. An additional advantage is that no correspondences have to exist between the 3D-points and the triangle mesh. The problem is solved utilizing a robust cost function in combination with an evolutionary global optimizer as shown in synthetic and real data experiments.

This paper presents a problem of 3D images decomposition into spheres. The presented method is based on a fast Hough transform with an input image space division. An essential element of this method is the use of a clustering technique for partial data sets. The method simplifies the application of Hough transform to segmentation tasks as well as accelerates calculations considerably.

Scene understanding is an important problem in intelligent robotics. Since visual information is uncertain due to several reasons, we need a novel method that has robustness to the uncertainty. Bayesian probabilistic approach is robust to manage the uncertainty, and powerful to model high-level contexts like the relationship between places and objects. In this paper, we propose a context-based Bayesian method with SIFT for scene understanding. At first, image pre-processing extracts features from vision information and objects-existence information is extracted by SIFT that is rotation and scale invariant. This information is provided to Bayesian networks for robust inference in scene understanding. Experiments in complex real environments show that the proposed method is useful.

In the recent times, many computer vision supported e-learning applications have been constructed, to provide the participants with the automated and real-time camera control capabilities. In this paper, we describe a portable and single-PC based instructional video capture system, which incorporates a variety of computer vision techniques for its video directing and close-up region specification. We describe the technologies used, including the laser-pointer detections, instructor’s lip tracking and individual teaching object recognition. As the same time, we also explain how we have achieved both and property in our design.

In this paper, we present a method for producing omnidirectional image signatures that are purposed to localize a mobile robot in an office environment. To solve the problem of perceptual aliasing common to the image based recognition approaches, we choose to build signatures that greatly vary between rooms and slowly vary inside a given room. To do so, an invariant approach has been developed, based on Haar invariant integrals. It takes into account the movements the robot can do in a room and the omni image transformations thus produced. A comparison with existing methods is presented using the Fisher criterion. Our method appears to get significantly better results for place recognition and robot localization, reducing in a positive way the perceptual aliasing.