Optical flow estimation in the presence of fast or discontinuous motion.

This thesis focuses on the computation of optical flow, i.e., the motion perceived from a sequence of gradually changing images, as an estimate for the 2D velocity of the scene. Due to the large variety and high complexity of the motion types existing in practice, motion recovery requires the estimation process to be highly adaptive. This thesis investigates how to select and combine the reasoning rules, namely the optical flow constraints, according to the type of motion information detected. Moreover, the thesis extends optical flow computation to fast rotation, an important, frequent and challenging motion type, which has not been addressed much in the literature. The thesis starts by proposing various measures, based on theory as well as heuristics, for motion inconsistency detection. This facilitates selecting only the optical flow constraints that are valid for each pixel. While this selection benefits pixels affected by inconsistent motion, the combination of different constraints also enhances flow recovery for pixels that have consistent motion. Two frameworks are designed for the combination of flow constraints. One utilizes motion segmentation; and the other is close in spirit to Expectation-Maximization. Within these frameworks, new constraints are formulated and tested. Furthermore, the adaptive reasoning is generalized from translational motion to motion that includes fast rotation. The key concept that enables this generalization is the use of intrinsic directions in differential geometry. Experimental results on a variety of benchmark sequences have demonstrated the ability of the proposed methods to improve the performance of existing techniques in several situations, including strong motion discontinuities and fast rotational motion.Thesis (Ph.D.) -- University of Adelaide, School of Computer Science, 201