Global Optimization Methods for Full-Reference and No-Reference Motion Estimation with Applications to Atherosclerotic Plaque Motion and Strain Imaging.

Pixel-based motion estimation using optical flow models has been extensively researched during the last two decades. The driving force of this research field is the amount of applications that can be developed with the motion estimates. Image segmentation, compression, activity detection, object tracking, pattern recognition, and more recently non-invasive biomedical applications like strain imaging make the estimation of accurate velocity fields necessary. The majority of the research in this area is focused on improving the theoretical and numerical framework of the optical flow models. This effort has resulted in increased method complexity with an increasing number of motion parameters. The standard approach of heuristically setting the motion parameters has become a major source of estimation error. This dissertation is focused in the development of reliable motion estimation based on global parameter optimization methods. Two strategies have been developed. In full-reference optimization, the assumption is that a video training set of realistic motion simulations (or ground truth) are available. Global optimization is used to calculate the best motion parameters that can then be used on a separate set of testing videos. This approach helps provide bounds on what motion estimation methods can achieve. In no-reference optimization, the true displacement field is not available. By optimizing for the agreement between different motion estimation techniques, the no-reference approach closely approximates the best (optimal) motion parameters. The results obtained with the newly developed global no-reference optimization approach agree closely with those produced with the full-reference approach. Moreover, the no-reference approach calculates velocity fields of superior quality than published results for benchmark video sequences. Unreliable velocity estimates are identified using new confidence maps that are associated with the disagreement between methods. Thus, the no-reference global optimization method can provide reliable motion estimation without the need for realistic simulations or access to ground truth. The methods developed in this dissertation are applied to ultrasound videos of carotid artery plaques. The velocity estimates are used to analyze plaque motion and produce novel non-invasive elasticity maps that can help in the identification of vulnerable atherosclerotic plaques.