Generalized least squares-based parametric motion estimation and segmentation

Abstract

This thesis proposes several techniques related with the motion estimation problem. In particular, it deals with global motion estimation for image registration and motion segmentation. In the first case, we will suppose that the majority of the pixels of the image follow the same motion model, although the possibility of a large number of outliers are also considered. In the motion segmentation problem, the presence of more than one motion model will be considered. In both cases, sequences of two consecutive grey level images will be used. A new generalized least squares-based motion estimator will be proposed. The proposed formulation of the motion estimation problem provides an additional constraint that helps to match the pixels using image gradient information. That is achieved thanks to the use of a weight for each observation, providing high weight values to the observations considered as inliers, and low values to the ones considered as outliers. To avoid falling in a local minimum, the proposed motion estimator uses a Feature-based method (SIFT-based) to obtain good initial motion parameters. Therefore, it can deal with large motions like translation, rotations, scales changes, viewpoint changes, etc. The accuracy of our approach has been tested using challenging real images using both affine and projective motion models. Two Motion Estimator techniques, which use M-Estimators to deal with outliers into a iteratively reweighted least squared-based strategy, have been selected to compare the accuracy of our approach. The results obtained have showed that the proposed motion estimator can obtain as accurate results as M-Estimator-based techniques and even better in most cases. The problem of estimating accurately the motion under non-uniform illumination changes will also be considered. A modification of the proposed global motion estimator will be proposed to deal with this kind of illumination changes. In particular, a dynamic image model where the illumination factors are functions of the localization will be used replacing the brightens constancy assumption allowing for a more general and accurate image model. Experiments using challenging images will be performed showing that the combination of both techniques is feasible and provides accurate estimates of the motion parameters even in the presence of strong illumination changes between the images. The last part of the thesis deals with the motion estimation and segmentation problem. The proposed algorithm uses temporal information, by using the proposed generalized least-squares motion estimation process and spatial information by using an iterative region growing algorithm which classifies regions of pixels into the different motion models present in the sequence. In addition, it can extract the different moving regions of the scene while estimating its motion quasi-simultaneously and without a priori information of the number of moving objects in the scene. The performance of the algorithm will be tested on synthetic and real images with multiple objects undergoing different types of motion.