Catálogo de publicaciones - libros

The recovery of a motion-blurred image is an important illposed inverse problem. But this subject has not recently received lot of attention. We propose a probabilistic method for the estimation of motion parameters based on the geometrical characteristic of the Fourier spectrum. Indeed, the Fourier spectrum of the blurred image is made by the product of the original Fourier spectrum with an oriented cardinal sine function. The estimation of the parameters reduces to the detection of the direction and of the gap between oscillations of the Fourier spectrum. Using the Helmholtz principle, the maximum meaningful parallel alignments are detected in the frequency domain, and then the direction and the extent of the blur are identified by an adapted K-means cluster algorithm. Simulation results show that the approach is very promising.

We present the results of a study to exploit a multiple colour space model () and variable kernels for object tracking in video sequences. The basis of our work is the mean shift algorithm; for a moving target, we develop a procedure to adaptively change the throughout a video sequence. The optional components are ranked using a similarity distance within an inner (representing the object) and outer (representing the surrounding region) rectangle. Rather than use the standard, Epanechnikov kernel, we have also used a kernel weighted by the normalized Chamfer distance transform to improve the accuracy of target representation and localization, minimising the distance between the two distributions of foreground and background using the Bhattacharya coefficient. To define the target shape in the rectangular window, either regional segmentation or background-difference imaging, dependent on the nature of the video sequence, has been used. Experimental results show the improved tracking capability and versatility of the algorithm in comparison with results using fixed colour models and standard kernels.

Ground vehicle tracking is an important component of Aerial Video Surveillance System (AVS). We address the problem of real-time and precise vehicle tracking from a single moving airborne camera which faces the challenges of congestion, occlusion and so on. We track a set of point features of the selected vehicle by the technique of image alignment. An edge feature-based outlier rejection criterion is proposed to eliminate the outlier caused by congestion and occlusion. Large motion and total occlusion is handled by a Kalman filter. Furthermore, a reappearance verification program is used to ensure the tracker gets back the right object. Experimental results on real aerial videos show the algorithm is reliable and robust.

This paper presents a novel and fast probabilistic method for white matter fiber tracking from diffusion weighted magnetic resonance imaging (DWI). We formulate fiber tracking on a nonlinear state space model which is able to capture both smoothness regularity of fibers and uncertainties of the local fiber orientations due to noise and partial volume effects. The global tracking model is implemented using particle filtering. This sequential Monte Carlo technique allows us to recursively compute the posterior distribution of the potential fibers, while there is no limitation on the forms of the prior and observed information. Fast and efficient sampling is realised using the von Mises-Fisher distribution on unit spheres. The fiber orientation distribution is theoretically formulated by combining the axially symmetric tensor model and the formal noise model for DWI. Given a seed point, the method is able to rapidly locate the global optimal fiber and also provide a connectivity map. The proposed method is demonstrated both on synthetic and real-world brain MRI dataset.

The complexity of visual representations is substantially limited by the compositional nature of our visual world which, therefore, renders learning structured object models feasible. During recognition, such structured models might however be disadvantageous, especially under the high computational demands of video. This contribution presents a compositional approach to video analysis that demonstrates the value of compositionality for both, learning of structured object models and recognition in near real-time. We unite category-level, multi-class object recognition, segmentation, and tracking in the same probabilistic graphical model. A model selection strategy is pursued to facilitate recognition and tracking of multiple objects that appear simultaneously in a video. Object models are learned from videos with heavy clutter and camera motion where only an overall category label for a training video is provided, but no hand-segmentation or localization of objects is required. For evaluation purposes a video categorization database is assembled and experiments convincingly demonstrate the suitability of the approach.

Video segmentation requires the partitioning of a series of images into groups that are both spatially coherent and smooth along the time axis. We formulate segmentation as a Bayesian clustering problem. Context information is propagated over time by a conjugate structure. The level of segment resolution is controlled by a Dirichlet process prior. Our contributions include a conjugate nonparametric Bayesian model for clustering in multivariate time series, a MCMC inference algorithm, and a multiscale sampling approach for Dirichlet process mixture models. The multiscale algorithm is applicable to data with a spatial structure. The method is tested on synthetic data and on videos from the MPEG4 benchmark set.

In multiple object tracking, the confusion caused by occlusion and similar appearances is an important issue to be solved. In this paper, trackability is proposed to measure how well given features can be used to find the correspondence of any given object in videos with multiple objects. Based on the analysis of trackability and computational complexity of the features under various occlusion conditions, a dynamic feature method cascade is presented to match the objects in consecutive frames. The cascade is composed of three tracking features: appearance, velocity and position. These features are enabled or disabled online to reduce computational complexity while obtaining similar trackability.

Experiments are conducted on 27062 frame occlusion objects, in the cases of good trackability, our experiments can obtain high succussful tracking rate with low computation burden, and in the cases of poor trackability, our estimation of trackability and confusion matrix can explain why they can not be tracked well.

Skeleton can be viewed as a compact shape representation in that the shape can be completely reconstructed form the skeleton. We present a novel method for skeleton pruning that is based on this fundamental skeleton property. We iteratively remove skeleton end braches with smallest relevance for shape reconstruction. The relevance of branches is measured as their contribution to shape reconstruction. The proposed pruning method allows us to overcome the instability of skeleton representation: a small boundary deformation leads to large changes in skeleton topology. Consequently, we are able to obtain very stable skeleton representation of planar shapes.

Most of the traditional methods for shape classification are based on contour. They often encounter difficulties when dealing with classes that have large nonlinear variability, especially when the variability is structural or due to articulation. It is well-known that shape representation based on skeletons is superior to contour based representation in such situations. However, approaches to shape similarity based on skeletons suffer from the instability of skeletons and matching of skeleton graphs is still an open problem. Using a skeleton pruning method, we are able to obtain stable pruned skeletons even in the presence of significant contour distortions. In contrast to most existing methods, it does not require converting of skeleton graphs to trees and it does not require any graph editing. We represent each shape as set of shortest paths in the skeleton between pairs of skeleton endpoints. Shape classification is done with Bayesian classifier. We present excellent classification results for complete shape.

This paper illustrates and extends an efficient framework, called the square-root-elastic (SRE) framework, for studying shapes of closed curves, that was first introduced in [2]. This framework combines the strengths of two important ideas - elastic shape metric and path-straightening methods - for finding geodesics in shape spaces of curves. The elastic metric allows for optimal matching of features between curves while path-straightening ensures that the algorithm results in geodesic paths. This paper extends this framework by removing two important shape preserving transformations: rotations and re-parameterizations, by forming quotient spaces and constructing geodesics on these quotient spaces. These ideas are demonstrated using experiments involving 2D and 3D curves.