Robust feature extraction in color and three-dimensional spaces

Distinct object description and matching is a common goal for many computer vision and medical imaging applications. Object recognition/retrieval, robotic vision, driver support systems, visual tracking, panoramic vision, camera self-calibration, camera planning, and image registration are a few out of the many applications that need robust description, matching and recognition methodologies. Successful object description requires two main conditions to be achieved: invariance and distinction. An efficient description methodology must be robust enough to accommodate different variations in imaging conditions. At the same time, it must produce a distinctive characterization of the object of interest.; Scale-space-based feature extraction approaches are the most efficient in extracting distinctive invariant features. Differential singularities of scale-space are used as the most robust points for feature extraction. However, in these approaches, scale-space differential singularities are considered to be equally stable. The first part of this dissertation is concerned with quantifying the stability of the detected interest points at scale-space differential singularities. An analysis of the behavior of these points in presence of noise is presented. Then, a novel stability measure of scale-space differential singularities is proposed. A transformation is developed in order to generate a transformed 1D function from the local neighborhood of these singularities. The curvature of the transformed 1D function is proved to be a quantitative measure that expresses the stability of the detected interest at these singularities.; To avoid the color sensitivity to illumination changes, most of the existing approaches depends on extracting local invariant features for gray images. However, color is a crucial source of distinction for many applications and it can be vital information in some systems. Despite the existence of some attempts to add color clues to local invariant features, they suffer from inefficient color invariance characteristics, limited geometric description, and/or large storage requirements. From another side, although robust point matching is an essential task for many 3D-based applications, the existing local invariant feature techniques are dedicated to 2D space. The existing point matching approaches for 3D data either deal with surface points only or use primitive matching techniques in comparison with matching using scale-space-based local invariant features.; So, two novel frameworks are proposed to extract local invariant features in both color and 3D spaces using scale-space theory. In the first framework, a novel local invariant feature descriptor that has colored invariant characteristics is developed. Scale-space theory is used to detect the most robust points at scale-space differential singularities in order to build descriptors representing local characteristics. The descriptors' entries are based on gradient orientation histograms, which are inspired from the response of the human perceptual system to visual effects. To embed color information in the built descriptors and at the same time achieve maximum robustness to illumination changes, a color invariant model is used as the working space of the feature extraction. The proposed framework achieves several advantages: including color as an important source of distinction in the extracted features, achieving illumination invariance, and disaffecting the size of the built descriptors in spite of adding color information. In the second framework, the use of 2D local invariant features is extended to 3D space. The proposed framework is summarized as follows: An input 3D volume is represented by a 4D Laplacian-of-Gaussian hyper-pyramid in scale-space. Then, interest points are detected at differential singularities of this scale-space 4D hyper-pyramid. For each of these detected points, a 3D gradient orientation histogram of the local neighborhood is constru...