Catálogo de publicaciones - libros

The authors consider combining correlations of different orders in kernel principal component analysis (kPCA) and kernel canonical correlation analysis (kCCA) with the correlation kernels. We apply combining methods, e.g., the sums of the correlation kernels, Cartesian spaces of the principal components or the canonical variates and the voting of kPCAs and kCCAs output and compare their performance in the classification of texture images. Further, we apply Kansei information on the images obtained through questionnaires to the public to kCCA and evaluate its effectiveness.

Visual hand gestures offer an interesting modality for Human-Computer-Interaction (HCI) applications. Gesture recognition and hand tracking, however, are not trivial tasks and real environments set a lot of challenges to algorithms performing such activities. In this paper, a novel combination of techniques is presented for tracking and recognition of hand gestures in real, cluttered environments. In addition to combining existing techniques, a method for locating a hand and segmenting it from an arm in binary silhouettes and a foreground model for color segmentation is proposed. A single hand is tracked with a single camera and the trajectory information is extracted along with recognition of five different gestures. This information is exploited for replacing the operations of a normal computer mouse. The silhouette of the hand is extracted as a combination of different segmentation methods: An adaptive colour model based segmentation is combined with intensity and chromaticity based background subtraction techniques to achieve robust performance in cluttered scenes. An affine-invariant Fourier-descriptor is derived from the silhouette, which is then classified to a hand shape class with support vector machines (SVM). Gestures are recognized as changes in the hand shape with a finite state machine (FSM).

In this work a combination of depth and silhouette information is presented to track the motion of a human from a single view. Depth data is acquired from a Photonic Mixer Device (PMD), which measures the time-of-flight of light. Correspondences between the silhouette of the projected model and the real image are established in a novel way, that can handle cluttered non-static backgrounds. Pose is estimated by Nonlinear Least Squares, which handles the underlying dynamics of the kinematic chain directly. Analytic Jacobians allow pose estimation with 5 FPS.

This paper explores how spin images can be constructed using shape-from-shading information and used for the purpose of face recognition. We commence by extracting needle maps from gray-scale images of faces, using a mean needle map to enforce the correct pattern of facial convexity and concavity. Spin images [6] are estimated from the needle maps using local spherical geometry to approximate the facial surface. Our representation is based on spin image histograms for an arrangement of image patches. Comparing to our previous spin image approach, the current one has two basic difference: Euclidean distance is replaced by geodesic distance; Irregular face region is applied to better fit face contour. We demonstrate how this representation can be used to perform face recognition across different subjects and illumination conditions. Experiments show the method to be reliable and accurate, and the recognition precision reaches 93% on CMU PIE sub-database.

Inter-layer prediction is the most important technique for improving coding performance in spatial enhancement layers in Scalable Video Coding (SVC). In this paper we discuss Adaptive Residual Interpolation (ARI), a new approach to inter-layer prediction of residual data. This prediction method yields a higher coding performance. We integrated the ARI tool in the Joint Scalable Video Model software. Special attention was paid to the CABAC context model initialization. Further, the use, complexity, and coding performance of this technology is discussed. Three filters were tested for the interpolation of lower-layer residuals: a bi-linear filter, the H.264/AVC 6-tap filter, and a median filter. Tests have shown that ARI prediction results in an average bit rate reduction of 0.40 % for the tested configurations without a loss in visual quality. In a particular test case, a maximum bit rate reduction of 10.10 % was observed for the same objective quality.

Two deformable registration methods, the Demons and the Morphon algorithms, have been used for registration of CT datasets to evaluate their usability in radiotherapy planning for prostate cancer. These methods were chosen because they can perform deformable registration in a fully automated way. The experiments show that for intrapatient registration both of the methods give useful results, although some differences exist in the way they deform the template. The Morphon method has, however, some advantageous compared to the Demons method. It is invariant to the image intensity and it does not distort the deformed data. The conclusion is therefore to recommend the Morphon method as a registration tool for this application. A more flexible regularization model is needed, though, in order to be able to catch the full range of deformations required to match the datasets.

In the area of quality control by vision, the reconstruction of 3D curves is a convenient tool to detect and quantify possible anomalies. Whereas other methods exist that allow us to describe surface elements, the contour approach will prove to be useful to reconstruct the object close to discontinuities, such as holes or edges.

We present an algorithm for the reconstruction of 3D parametric curves, based on a fixed complexity model, embedded in an iterative framework of control point insertion. The successive increase of degrees of freedom provides for a good precision while avoiding to over-parameterize the model. The curve is reconstructed by adapting the projections of a 3D NURBS snake to the observed curves in a multi-view setting.

We present a video shot boundary detection (SBD) algorithm that spots discontinuities in visual stream by monitoring video frame trajectories on Self-Organizing Maps (SOMs). The SOM mapping compensates for the probability density differences in the feature space, and consequently distances between SOM coordinates are more informative than distances between plain feature vectors.

The proposed method compares two sliding best-matching unit windows instead of just measuring distances between two trajectory points, which increases the robustness of the detector. This can be seen as a variant of the adaptive threshold SBD methods. Furthermore, the robustness is increased by using a committee machine of multiple SOM-based detectors. Experimental evaluation made by NIST in the TRECVID evaluation confirms that the SOM-based SBD method works comparatively well in news video segmentation, especially in gradual transition detection.

This paper presents a general approach for surface-to-surface registration (S2SR) with the Euclidean metric using signed distance maps. In addition, the method is symmetric such that the registration of a shape A to a shape B is identical to the registration of the shape B to the shape A.

The S2SR problem can be approximated by the image registration (IR) problem of the signed distance maps (SDMs) of the surfaces confined to some narrow band. By shrinking the narrow bands around the zero level sets the solution to the IR problem converges towards the S2SR problem. It is our hypothesis that this approach is more robust and less prone to fall into local minima than ordinary surface-to-surface registration. The IR problem is solved using the inverse compositional algorithm.

In this paper, a set of 40 pelvic bones of Duroc pigs are registered to each other w.r.t. the Euclidean transformation with both the S2SR approach and iterative closest point approach, and the results are compared.

In this paper, we propose a new multiple object tracking method via multi-layer multi-modal framework. To handle erroneous merge and labeling problem in multiple object tracking, we use a multi layer representation of dynamic Bayesian network and modified sampling method. For robust visual tracking, our dynamic Bayesian network based tracker fuses multi-modal features such as color and edge orientation histogram. The proposed method was evaluated under several real situations and promising results were obtained.