Catálogo de publicaciones - libros

Foreground blobs in a mixed stereo pair of videos (visible and infrared sensors) allow a coarse evaluation of the distances between each blob and the uncalibrated cameras. The main goals of this work are to find common feature points in each type of image and to create pairs of corresponding points in order to obtain coarse positionning of blobs in space. Feature points are found by two methods: the skeleton and the Discrete Curve Evolution (DCE) algorithm. For each method, a feature-based algorithm creates the pairs of points. Blob pairing can help to create those pairs of points. Finally, a RANSAC algorithm filters all pairs of points in order to respect the epipolar geometrical constraints. The median horizontal disparities for each pair of blobs are evaluated with two different ground truths. In most cases, the nearest blob is detected and disparities are as accurate as the background subtraction allows.

Several approaches to optical flow estimation use differential methods to model changes in image brightness over time. In computer vision it is often desirable to over constrain the problem to more precisely determine the solution and enforce robustness. In this paper, two new solutions for optical flow computation are proposed which are based on combining brightness and gradient constraints using more than one quadratic constraint embedded in a robust statistical function. Applying the same set of differential equations to different quadratic error functions produces different results. The two techniques combine the advantages of different constraints to achieve the best results. Experimental comparisons of estimation errors against those of well-known synthetic ground-truthed test sequences showed good qualitative performance.

We present a novel method for human motion recognition. A video sequence is represented with a sparse set of spatial and spatial-temporal features by extracting static and dynamic interest points. Our model learns a set of poses along with the dynamics of the sequence. Pose models and the model of motion dynamics are represented as a constellation of static and dynamic parts, respectively. On top of the layer of individual models we build a higher level model that can be described as “constellation of constellation models”. This model encodes the spatial-temporal relationships between the dynamics of the motion and the appearance of individual poses. We test the model on a publicly available action dataset and demonstrate that our new method performs well on the classification tasks. We also perform additional experiments to show how the classification performance can be improved by increasing the number of pose models in our framework.

In this paper, we introduce a new technique to estimate optical flow fields based on wavelet decomposition. In order to block error propagation between layers of multi-resolution image pyramid, we consider information of the all pyramid levels at once. We add a homogenous smoothness constraint to the system of optical flow constraints to obtain smooth motion fields. Since there are approximations on both sides of our over determined equation system, a total least square method is used as a minimization technique. The method was tested on several standard sequences in the field and megavoltage images taken by linear accelerator devices and showed promising results.

A tensor-based optical flow algorithm is presented in this paper. This algorithm uses a cost function that is an indication of tensor certainty to adaptively adjust weights for tensor computation. By incorporating a good initial value and an efficient search strategy, this algorithm is able to determine optimal weights in a small number of iterations. The weighting mask for the tensor computation is decomposed into rings to simplify a 2D weighting into 1D. The devised algorithm is well-suited for real-time implementation using a pipelined hardware structure and can thus be used to achieve real-time optical flow computation. This paper presents simulation results of the algorithm in software, and the results are compared with our previous work to show its effectiveness. It is shown that the proposed new algorithm automatically achieves equivalent accuracy to that previously achieved via manual tuning of the weights.

A two-phase segmentation mechanism is described that allows a global homogeneity-related measure to be optimized in a level-set formulation. The mechanism has uniform treatment toward texture, gray level, and color boundaries. Intensities or colors of the image are first coarsely quantized into a number of classes. Then a class map is formed by having each pixel labeled with the class identity its gray or color level is associated with. With this class map, for any segmented region, it can be determined which pixels inside the region belong to which classes, and it can even be calculated how spread-out each of such classes is inside the region. The average spread-size of the classes in the region, in comparison with the size of the region, then constitutes a good measure in evaluating how homogeneous the region is. With the measure, the segmentation problem can be formulated as the optimization of the average homogeneity of the segmented regions. This work contributes chiefly by expressing the above optimization functional in such a way that allows it to be encoded in a variational formulation and that the solution can be reached by the deformation of an active contour. In addition, to solve the problem of multiple optima, this work incorporates an additional geodesic term into the functional of the optimization to maintain the active contour’s mobility at even adverse condition of the deformation process. Experimental results on synthetic and real images are presented to demonstrate the performance of the mechanism.

A general framework for image segmentation is presented in this paper, based on the paradigm of water flow. The major water flow attributes like water pressure, surface tension and capillary force are defined in the context of force field generation and make the model adaptable to topological and geometrical changes. A flow-stopping image functional combining edge- and region-based forces is introduced to produce capability for both range and accuracy. The method is assessed qualitatively and quantitatively on synthetic and natural images. It is shown that the new approach can segment objects with complex shapes or weak-contrasted boundaries, and has good immunity to noise. The operator is also extended to 3-D, and is successfully applied to medical volume segmentation.

Complex multiple structures, high uncertainty due to the existence of moving objects, and significant disparity in the size of features are the main issues associated with processing range data of outdoor scenes. The existing range segmentation techniques have been commonly developed for laboratory sized objects or simple architectural building features. In this paper, main problems related to the geometrical segmentation of large and significant buildings are studied. A robust and accurate range segmentation approach is also devised to extract very fine geometric details of building exteriors. It uses a hierarchical model-base range segmentation strategy and employs a high breakdown point robust estimator to deal with the existing discrepancies in size and sampling rates of various features of large outdoor objects. The proposed range segmentation algorithm facilitates automatic generation of fine 3D models of environment. The computational advantages and segmentation capabilities of the proposed method are shown using real range data of large building exteriors.

This paper proposes a novel method for segmenting lips from face images or video sequences. A non-linear learning method in the form of an SVM classifier is trained to recognise lip colour over a variety of faces. The pixel-level information that the trained classifier outputs is integrated effectively by minimising an energy functional using level set methods, which yields the lip contour(s). The method works over a wide variety of face types, and can elegantly deal with both the case where the subjects’ mouths are open and the mouth contour is prominent, and with the closed mouth case where the mouth contour is not visible.

This paper describes a two level classification algorithm to discriminate the handwritten elements from the printed text in a printed document. The proposed technique is independent of size, slant, orientation, translation and other variations in handwritten text. At the first level of classification, we use two classifiers and present a comparison between the nearest neighbour classifier and Support Vector Machines(SVM) classifier to localize the handwritten text. The features that are extracted from the document are seven invariant central moments and based on these features, we classify the text as hand-written. At the second level, we use Delaunay triangulation to reclassify the misclassified elements. When Delaunay triangulation is imposed on the centroid points of the connected components, we extract features based on the triangles and reclassify the misclassified elements. We remove the noise components in the document as part of the pre-processing step.