Etude de contraintes spatiales bas niveau appliquees a la vision par ordinateur

The goal of this thesis is to study a series of low-level and spatial constraints to help better regularize higher-level optimization methods of computer vision. More specifically, we will address the problems of optical flow estimation, motion detection, occlusion detection and motion segmentation.; Optical flow is a mathematical concept used to describe the visual movement in a video sequence. Among the many methods used to estimate optical flow is the one proposed by Lucas and Kanade (LK) [22]. Their method is based on a sum of least-squares regression technique that is easy to implement and which quickly converges. Although their method has definite advantages, it is very sensitive to lack of texture and to the presence of outliers. As an answer to these two limitations, we propose in Chapter 1 a slightly modified version of the LK algorithm. Our modification includes two low-level constraints. The first constraint works as a Best Linear Unbiased Estimate (BLUE) low-pass filter that helps regularize the flow in noisy areas and in areas with little (or no) texture. The second constraint is based on the mean-shift algorithm [37]. The aim of this second constraint is to eliminate the presence of outliers during the optical flow estimation. The main idea of this constraint is to slightly move the local neighborhood toward regions where motion is more likely to be unimodal than multimodal.; A second problem that we address in this document is motion detection. The goal of motion detection is to segment a video sequence in mobile and static regions. Most motion detection methods implemented nowadays rely on the background subtraction principle. In Chapter 3, we propose a new probabilistic background subtraction paradigm. This paradigm allows us to train the PDFs on only one background frame instead of many frames as is usually the case. In this chapter, we show how the use of local spatial constraints can help reduce some basic limitations of most statistical background subtraction methods.; Many problems in digital imagery require the estimation of a so-called label field. A label field can be seen as an image whose pixels contain a class label. Some of the label-field estimation algorithms come from the Statistical Learning Theory. The most well known algorithms are those based on the maximum likelihood and the maximum a posteriori principles. Despite the many advantages of those approaches, they are sometimes sensitive to outliers and to noisy input data. In response to those weaknesses, we propose in Chapter 2 a way to enhance the precision of an estimated label field. In fact, the proposed method fuses label fields instead of features as is usually the case. The aim of this method is to force the label field regions to fit the shape of the most predominant object in the scene. To do so, we suggest two energy functions that we minimize with the deterministic Iterative Conditional Mode (ICM) algorithm. The first function is formed by a likelihood energy term whereas the second one is made up of a likelihood and an a priori energy term. We also show how this fusion procedure can be adapted to the context of motion segmentation and occlusion detection.; In Chapter 4, we present a way to drastically reduce the processing times related to many statistical learning algorithms applied to computer vision. In fact, we can reach acceleration factors from 4 to 200. Our approach is based on the "parallel" property some algorithms have. In fact, based on that property, we show how many Markovian density estimation and segmentation algorithms can be significantly accelerated when implemented on a Graphics Processor Unit (GPU). A GPU is a streaming processor (i.e. a processor with inherent parallel abilities) embedded in most graphics cards nowadays available on the market. In this chapter, we concentrate on Markovian algorithms applied to stereovision, motion estimation, motion segmentation and color segmentation. (Abstract shortened b...