Abstract
The use of a robust, low-level motion estimator based on a Robust Hough Transform (RHT) in a range of tasks, such as optical flow estimation, and motion estimation for video coding and retrieval from video sequences was discussed. RHT derived not only pixels displacements, but also provided direct motion segmentation and other motion-related clues. The RHT algorithm employed an affine region-to-region transformation model and was invariant to illumination changes, in addition to being statistically robust. It was found that RHT did not base the correspondence analyses on any specific type of feature, but used textured regions in the image as non-localized features.