Motion estimation and motion-based segmentation in digital image sequences

This dissertation considers four independent but interconnected problems relating to motion estimation and motion-based segmentation in digital image sequences. The work is based on velocity tuned 2D + T filter method which tunes the mother function to spatial translation, temporal translation, spatial scale, spatial orientation, translational velocity and angular velocity.; The first problem is deformation estimation. The types of deformation include expansion, contraction and shear. The main contribution here is that a log sampling method was devised to convert the deformation to a translational motion, thus derive the deformation parameters and perform selective reconstruction using translational velocity tuned filter method. It is robust against noise and occlusion.; The second problem is translational and rotational motion estimation and reconstruction. The motion being studied are rotational motion and translational composed with rotational motion. The new algorithm is based on translational velocity and rotational velocity tuned 2D + T filters. It is effective to be applied to both spinning motion and orbiting motion for rotational motion estimation. The translational motion parameters and the rotational motion parameters can be derived separately for translational motion composed with rotational motion. The experiments on image sequences demonstrate their robustness against noise and occlusion.; The third problem is optical flow estimation and its application to motion-based segmentation. Simultaneously, the local edge locations and their corresponding optical flows can be obtained, which can be further used for motion-based segmentation using clustering and ridge skeleton. It is achieved by using translational velocity tuned filters. The algorithms are tested on real image sequences obtained by a moving camera and shows to be accurate and robust.; In the fourth problem, a line segment on a plane is undergoing 3D affine motion. By perspective projection, an optical flow equation can be obtained, which has eight parameters needed to be estimated. A least square error estimation method is designed to solve this problem. By analysis, a condition number problem is identified, which affects the accuracy of the parameters estimation and the solution to this problem is also provided.