Abstract

This thesis presents a systematic study of motion estimation in the 3-D Gabor domain. The 3-D Gabor representation provides a joint spatiotemporal/spatiotemporal-frequency(st/stf) analysis framework to disclose the spatiotemporal localized frequency information in an image sequence in a manageable information structure. It offers the opportunity to take advantages of frequency domain approaches with spatiotemporal locality. The main contribution of this thesis is to realize an innovative motion estimation technique achieving high spatiotemporal resolution with the superior reliability of the model-based frequency domain approaches. Compared with the global nature of frequency domain approaches and the noise sensitive spatiotemporal approaches, the scheme enhances motion estimation performance both in accuracy and robustness. The mathematical forms of translational motion and rotational motion in the 3-D Gabor domain are derived. For the translational motion, an algorithm for piecewise uniform translational motion is designed. Various piecewise cases can be addressed by setting different Gabor representation parameters. For applications demanding very high spatiotemporal resolutions, a dense motion field estimation method is constructed in the 3-D Gabor domain. The performance evaluation based on artificial image sequences shows high spatiotemporal resolution motion estimation results with small errors and good noise resistance. In rotational motion studies, the 3-D Gabor representation also leads to an algorithm to identify and locate multiple rotating objects in an image sequence.