Catálogo de publicaciones - libros

Detecting regions of interest in video sequences is one of the most important tasks in many high level video processing applications. In this paper a novel approach based on Support Vector Data Description (SVDD) is presented. The method detects foreground regions in videos with quasi-stationary backgrounds. The SVDD is a technique used in analytically describing the data from a set of population samples. The training of Support Vector Machines (SVM’s) in general, and SVDD in particular requires a Lagrange optimization which is computationally intensive. We propose to use a genetic approach to solve the Lagrange optimization problem. The Genetic Algorithm (GA) starts with the initial guess and solves the optimization problem iteratively. Moreover, we expect to get accurate results with less cost than the Sequential Minimal Optimization (SMO) technique.

In this paper, we present an approach for addressing the ‘query by example’ problem in video surveillance, where a user specifies an object of interest and would like the system to return some images (e.g. top five) of that object or its trajectory by searching a large network of overlapping or non-overlapping cameras. The approach proposed is based on defining an appearance model for every detected object or trajectory in the network of cameras. The model integrates relative position, color, and texture descriptors of each detected object. We present a ‘pseudo track’ search method for querying using a single appearance model. Moreover, the availability of tracking within every camera can further improve the accuracy of such association by incorporating information from several appearance models belonging to the object’s trajectory. For this purpose, we present an automatic clustering technique allowing us to build a multi-valued appearance model from a collection of appearance models. The proposed approach does not require any geometric or colorimetric calibration of the cameras. Experiments from a mass transportation site demonstrate some promising results.

Surveillance systems that operate continuously generate large volumes of data. One such system is described here, continuously tracking and storing observations taken from multiple stereo systems. Automated event recognition is one way of annotating track databases for faster search and retrieval. Recognition of complex events in such data sets often requires context for successful disambiguation of apparently similar activities. Conditional random fields permit straightforward incorporation of temporal context into the event recognition task. This paper describes experiments in activity learning, using conditional random fields to learn and recognize composite events that are captured by the observation stream.

Intrinsic image decomposition is useful for improving the performance of such image understanding tasks as segmentation and object recognition. We present a new intrinsic image decomposition algorithm using the Fisher Linear Discriminant based on the assumptions of Lambertian surfaces, approximately Planckian lighting, and narrowband camera sensors. The Fisher Linear Discriminant not only considers the within-sensor data as convergent as possible but also treats the between-sensor data as separate as possible. The experimental results on real-world data show good performance of this algorithm.

Shape-from-shading method for oblique light source appears most applicable with minimal effects of the convex-concave ambiguity and shadows. In this paper, first, a robust iterative relation that reconstructs shape is constructed by applying the Jacobi iterative method to the equation between the reflectance map and image for each of the four approximations of the surface normal and by combining the resulting four relations as constraints. The relation ensures convergence, but that alone is not enough to reconstruct correct shapes for bright image parts or mathematically singular points. Next, to solve the problem, the light direction is tilted in slant angle following a criterion and the average tilt of the resulting shape is compensated. A numerical study using synthetic Mozart images shows that the method works well for a wide direction of the light source and that it gives more correct shapes than any of existing methods. Results for real images are also given, showing its usefulness more convincingly.

We present a new camera based algorithm to track pedestrians from a moving host using corner points. The algorithm can handle partial shape variations and the set of point movement vectors allows to estimate not only translation but also scaling. The algorithm works as follows: Corner points are extracted within a bounding box, where the pedestrian is detected in the current frame and in a search region in the next frame. We compare the local neighbourhood of points to find point correspondences using an improved method. The point correspondences are used to estimate the object movement using a translation scale model. A fast iterative outlier removal strategy is employed to remove single false point matches. A correction step is presented to correct the position estimate. The step uses the accumulated movement of each point over time to detect outliers that can not be found using inter-frame motion vectors. First tests indicate a good performance of the presented tracking algorithm, which is improved by the presented correction step.

This paper addresses the problem of 3D reconstruction and orientation of the cutting tool on a machine tool after it is loaded onto the spindle. Considering the reconstruction efficiency and that a cutting tool is a typical object of surface of revolution (SOR), a method based on a single calibrated view is presented, which only involves simple perspective projection relationship. First, the position and the orientation of the cutting tool is determined from an image. Then the silhouette of the cutting tool on the image is used to generate the 3D model, section by section. The designed algorithm is presented. This method is applicable to various kinds of cutting tools. Simulation and actual experiments on a machine tool verify that the method is correct with an accuracy of less than 1 mm.

The ball is really hard to be detected when it is merged with field lines or players in soccer video. A trajectory based ball detection scheme together with an approach of playfield detection is proposed to solve this problem. Playfield detection plays a fundamental role in semantic analysis of soccer video. An improve Generalized Lloyd Algorithm (GLA) based method is introduced to detect the playfield. Based on the detected playfield, an improved Viterbi algorithm is utilized to detect and track the ball. A group of selected interpolation points are calculated employing least squares method to track the ball in the playfield. An occlusion reasoning procedure is used to further qualify some undetected and false ball positions. The experimental results have verified their effectiveness of the given schema.

A single-view matching constraint is described which represents a necessary condition which 6 points in an image must satisfy if they are the images of 6 known 3D points under an arbitrary projective transformation. Similar to the well-known matching constrains for two or more view, represented by fundamental matrices or trifocal tensors, single-view matching constrains are represented by tensors and when multiplied with the homogeneous image coordinates the result vanishes when the condition is satisfied. More precisely, they are represented by 6-th order tensors on ℝ which can be computed in a simple manner from the camera projection matrix and the 6 3D points. The single-view matching constraints can be used for finding correspondences between detected 2D feature points and known 3D points, e.g., on an object, which are observed from arbitrary views. Consequently, this type of constraint can be said to be a representation of 3D shape (in the form of a point set) which is invariant to projective transformations when projected onto a 2D image.

This paper proposes an approach of face recognition using 3D face data based on Principle Component Analysis (PCA) and Linear Discriminant Analysis (LDA). This approach first aligned 3D faces based on nonuniform mesh re-sampling by computing face surface curves. This step achieves aligning of 3D prototypes based on facial features, eliminates 3D face size information and preserves important 3D face shape information in the input face. Then 2D texture information and the 3D shape information are extracted from 3D face images for recognition. Experimental results for 105 persons 3D face data set obtained by Cyberware 3030RGB/PS laser scanner have demonstrated the performance of our algorithm.