3D-rigid objects motion segmentation: a study of practical limitations

Motion segmentation or recovering structure-and-motion (SaM) from images of dynamic scenes plays a significant role in many computer-vision applications ranging from local navigation of a mobile robot to image rendering for multimedia applications. Since in many applications, the exact type of motion and camera parameters are not known, a priori, the fundamental matrix is commonly used as a general motion model. Although the estimation of a fundamental matrix and its use for motion segmentation are well understood, the studies of conditions governing the feasibility of segmentation for different types of motions are largely unaddressed. In this thesis, the feasibility of motion segmentation using the fundamental matrix is analysed. The focus is on a scene including multiple SaMs viewed by an uncalibrated camera. The quantifiable measures for the degree of separation were theoretically derived for the types of motion that are usually seen in practical applications, namely, motion from background, translational motions and planar motions. Sets of condition to guarantee successful segmentation were proposed via extensive experiments, the design of which was based on the Monte Carlo statistical method, using synthetic images. Experiments using real image data were set up and executed to examine the relevance of those conditions to the problems encountered in real applications. The experimental results show the capability of the proposed conditions to correctly predict the outcome of several segmentation scenarios. In addition, they also show that the Monte Carlo experimental results are very relevant to the problems encountered in real applications. In practice, the success of motion segmentation could be predicted via the value of the degree of separation between two motions estimated from obtainable scene and motion parameters. Therefore, the proposed conditions serve as a guideline for practitioners in designing motion segmentation solutions for computer-vision applications