Catálogo de publicaciones - libros

In this article, we propose a generalization of the Euclidean -shape, and through it we present an algorithm to retrieval shapes that are similar to a query shape from a finite set of points. Similarity means that we look for patterns identical, according to a given measure, to a query shape independently of translation, rotation and scaling transforms.

This paper presents a direct and stochastic technique for real time estimation of on board camera position and orientation—the ego-motion problem. An on board stereo vision system is used. Unlike existing works, which rely on feature extraction either in the image domain or in 3D space, our proposed approach directly estimates the unknown parameters from the brightness of a stream of stereo pairs. The pose parameters are tracked using the particle filtering framework which implicitly enforces the smoothness constraints on the dynamics. The proposed technique can be used in driving assistance applications as well as in augmented reality applications. Experimental results and comparisons on urban environments with different road geometries are presented.

A new fast motion estimation algorithm is presented. Our goal was to find the best block-matching results in a computation-limited and varying environment, and the conventional Diamond Search (DS) algorithm, though faster than most known algorithms, was found to be not very robust in terms of objective and subjective qualities for several sequences. The proposed algorithm, Enhanced Cross-Diamond Search (ECDS), employs a small cross search in the initial step. The large/small DS patterns are utilized as subsequent steps for fast block motion estimation. Experiment results show that the proposed algorithm outperforms other popular fast motion estimation algorithms in terms of PSNR and search speed, especially for sequences having random or fast motions.

When finding true motion vectors in video sequences, multiple local minima areas such as periodic pattern cause severe motion errors. There have been efforts to reduce motion errors in the region, where they use exhaustive full-search motion estimation scheme to analyze or find a solution for the region. To find robust motion vectors in the periodic pattern region using non-exhaustive motion estimator, we propose a recursive motion vector smoothing method. Recursively averaged vectors are used for periodic pattern region and the input vectors from a conventional search method are used for other regions, controlled by a weighting parameter. Properties of periodic pattern is considered in calculating the parameter to adaptively weight on the input or the mean vectors. Experimental results show motion vector improvements in the periodic pattern region with the input vectors from non-exhaustive search method.

Image mosaicing is stirring up a lot of interests in the research community for both its scientific significance and potential spin-off in real world applications. Being able to perform automatic image alignment in a common tonal and spatial reference can trigger a wide range of higher level image processing tasks such as panoramic image construction, scene depth computation, resolution enhancement, motion detection and tracking using non stationary camera. In this work we propose a fully automated real time on-line mosaicing algorithm able to build high quality seam-free panoramic images. Moreover, the whole approach does not exploit any a priori information regarding scene geometry, acquisition device properties or feedback signals, thus resulting in a fully image based solution. Extensive experiments have been accomplished to assess the quality of the attained mosaics by using them as the background to perform motion detection and tracking with a Pan Tilt Zoom camera.

We describe the implementation of a 2D optical flow algorithm published in the European Conference on Computer Vision (ECCV 2004) by Brox et al. [1] (best paper award) and a qualitative and quantitative evaluation of it for a number of synthetic and real image sequences. Their optical flow method combines three assumptions: a brightness constancy assumption, a gradient constancy assumption and a spatio-temporal smoothness constraint. A numerical scheme based on fixed point iterations is used. Their method uses a coarse-to-fine warping strategy to measure larger optical flow vectors. We have investigated the algorithm in detail and our evaluation of the method demonstrates that it produces very accurate optical flow fields from only 2 input images.

In this paper, we propose an efficient and robust multiple targets tracking method based on particle filtering and Dezert-Smarandache theory. A model of cue combination is designed with plausible and paradoxical reasoning. The proposed model can resolve the conflict and paradoxes that arise between measured cues due to the particle or total occlusion. Experimental results demonstrate the efficiency and accuracy of the model in case of tracking with multiple cues.

Because many model-based object representation appro- aches such as active appearance models (AAMs) use a fixed linear appearance model, they often fail to fit to a novel image that is captured in a different imaging condition from that of training images. To alleviate this problem, we propose to use adaptive linear appearance model that is updated by the incremental principal component analysis (PCA). Because the incremental update algorithm uses a new appearance data that is obtained in an on-line manner, a reliable method to measure the quality of the new data is required not to break the integrity of the appearance model. For this purpose, we modified the adaptive observation model (AOM), which has been used to model the varying appearance of the target object using statistical model such as Gaussian mixtures. Experiment results showed that the incremental AAM that uses adaptive linear appearance model greatly improved the robustness to the varying illumination condition when compared to the traditional AAM.

We propose an efficient real-time face tracking system that can follow fast movements. For face tracking, we use a particle filter that can handle arbitrary distributions. To track fast movements, we predict the motions using motion history and motion estimation, hence we can find the face with fewer particles. For observation model, we use active appearance model(AAM) to obtain an accurate face region, and update the model using incremental principle component analysis(IPCA). Occlusion handling scheme incorporates motion history to handle the moving face with occlusion. We present several experimental results to prove that our system shows better performance than previous works.

Mean-Shift tracking gained a lot of popularity in computer vision community. This is due to its simplicity and robustness. However, the original formulation does not estimate the orientation of the tracked object. In this paper, we extend the original mean-shift tracker for orientation estimation. We use the gradient field as an orientation signature and introduce an efficient representation of the gradient-orientation space to speed-up the estimation. No additional parameter is required and the additional processing time is insignificant. The effectiveness of our method is demonstrated on typical sequences.