Catálogo de publicaciones - libros

High-level understanding of motion events is a critical task in any system which aims to analyse dynamic human-populated scenes. However, current tracking techniques still do not address complex interaction events among multiple targets. In this paper, a principled event-management framework is proposed, and it is included in a hierarchical and modular tracking architecture. Multiple-target interaction events, and a proper scheme for tracker instantiation and removal according to scene events, are considered. Multiple-target group management allows the system to switch among different operation modes. Robust and accurate tracking results have been obtained in both indoor and outdoor scenarios, without considering a-priori knowledge about either the scene or the targets based on a previous training period.

A novel algorithm for field programmable gate array (FPGA) realization of NN classifier is presented in this paper. The algorithm identifies first closest vectors in the design set of a NN classifier for each input vector by performing the partial distance search (PDS) in the wavelet domain. It employs subspace search, bitplane reduction and multiple-coefficient accumulation techniques for the effective reduction of the area complexity and computation latency. The proposed implementation has been embedded in a softcore CPU for physical performance measurement. Experimental results show that the implementation provides a cost-effective solution to the FPGA realization of NN classification systems where both high throughput and low area cost are desired.

We apply a general form of affine transformation model to compensate illumination variations in a series of multispectral images of a static scene and compare it to a particular affine and a diagonal transformation models. These models operate in the original multispectral space or in a lower-dimensional space obtained by Singular Value Decomposition () of the set of images. We use a system consisting of a multispectral camera and a light dome that allows the measurement of multispectral data under carefully controlled illumination conditions to generate a series of multispectral images of a static scene under varying illumination conditions. We evaluate the compensation performance using the colour difference between images. The experiments show that the first 2 models perform satisfactorily in the original and lower dimensional spaces.

The rendering of lower resolution image data on higher resolution displays has become a very common task, in particular because of the increasing popularity of webcams, camera phones, and low-bandwidth video streaming. Thus, there is a strong demand for real-time, high-quality image magnification. In this work, we suggest to exploit the high performance of programmable graphics processing units (GPUs) for an adaptive image magnification method. To this end, we propose a GPU-friendly algorithm for image up-sampling by edge-directed image interpolation, which avoids ringing artifacts, excessive blurring, and staircasing of oblique edges. At the same time it features gray-scale invariance, is applicable to color images, and allows for real-time processing of full-screen images on today’s GPUs.

We propose a coarse registration method of range images using both geometric and photometric features. The framework of existing methods using multiple features first defines a single similarity distance summing up each feature based evaluations, and then minimizes the distance between range images for registration. In contrast, we formulate registration as a graph-based optimization problem, where we independently evaluate geometric feature and photometric feature and consider only the order of point-to-point matching quality. We then find as large consistent matching as possible in the sense of the matching-quality order. This is solved as one global combinatorial optimization problem. Our method thus does not require any good initial estimation and, at the same time, guarantees that the global solution is achieved.

There is great interest in the development of automatic geometric correction systems for satellite images. A fully automatic system, based exclusively on the identification of tie points (image to image control points) by image matching needs to use efficient selection and validation methods. Four Tie Point Suitability Indices (TPSI) are proposed to select the most suitable areas in an image to search for tie points. Three tie point validation parameters are also proposed. The validation parameters make use of the various spectral bands available in hyperspectral and multispectral satellite images. The proposed TPSIs and validation parameters were tested with hyperspectral high-resolution satellite images from the CHRIS/PROBA sensor.

In this paper, we propose a pose editing and animation method for triangulated surfaces based on a user controlled partitioning of the model into deformable parts and rigid parts which are denoted handles. In our pose editing system, the user can sculpt a set of poses simply by transforming the handles for each pose. Using Laplacian editing, the deformable parts are deformed to match the handles. In our animation system the user can constrain one or several handles in order to define a new pose. New poses are interpolated from the examples poses, by solving a small non-linear optimization problem in order to obtain the interpolation weights. While the system can be used simply for building poses, it is also an animation system. The user can specify a path for a given constraint and the model is animated correspondingly.

Liveness detection is increasingly planned to be incorporated into biometric systems to reduce the risk of spoofing and impersonation. Some of the techniques used include detection of motion of the head while posing/speaking, iris size in varying illumination, fingerprint sweat, text-prompted speech, speech-to-lip motion synchronization etc. In this paper, we propose to build a biometric signal to test attack resilience of biometric systems by creating a text-driven video synthesis of faces. We synthesize new realistic looking video sequences from real image sequences representing utterance of digits. We determine the image sequences for each digit by using a GMM based speech recognizer. Then, depending on system prompt (sequence of digits) our method regenerates a video signal to test attack resilience of a biometric system that asks for random digit utterances to prevent play-back of pre-recorded data representing both audio and images. The discontinuities in the new image sequence, created at the connection of each digit, are removed by using a frame prediction algorithm that makes use of the well known block matching algorithm. Other uses of our results include web-based video communication for electronic commerce and frame interpolation for low frame rate video.

Current automatic methods based on mid-sagittal plane to segment left and right human brain hemispheres in 3D magnetic resonance (MR) images simply use a planar surface. However, the two brain hemispheres, in fact, can not be separated by just a simple plane properly. A novel automatic method to segment left and right brain hemispheres in MR images is proposed in this paper, which is based on an extended shape bottlenecks algorithm and a fast and robust partial volume estimation approach. In this method, brain tissues firstly are extracted from the MR image of human head. Then the information potential map is generated, according to which a brain hemisphere mask with the same size of the original image is created. 10 simulated and 5 real T1-weighted MR images were used to evaluate this method, and much more accurate segmentation of human brain hemispheres was achieved comparing with the segmentation with mid-sagittal plane.

We discuss a method suitable for inpainting both large scale geometric structures and stochastic texture components. We use the well-known FRAME model for inpainting. We introduce a temperature term in the learnt FRAME Gibbs distribution. By using a fast cooling scheme a MAP-like solution is found that can reconstruct the geometric structure. In a second step a heating scheme is used that reconstruct the stochastic texture. Both steps in the reconstruction process are necessary, and contribute in two very different ways to the appearance of the reconstruction.