There are radically some errors a on motion estimation in successive images by several reasons to make the false estimation : the change of intensities, various motion within a region, and the deformation of measured gradients. In this thesis, we analyze quantitatively these errors and present the new algorithms for the error reduction.
First, the causes of intensity change are the noises added from such sources as sensor, film grain irregularities, and atmospheric light fluctuation and the sub-pel motion following sampling since the sub-pel motion changes the sample position on spatial coordinate. Through the analysis, in order to reduce the effect of noises, the sampling rate must be decreased, while increased in the effect of the sub-pel motion. Therefore, with the trade-off between two effects we find a fact that the sampling rate should be 2 - 16 times of Nyquist rate.
Second, if the motion within a given region is not purely translation, as happens in practice especially near the motion boundaries, some errors can be generated with using the fixed region size. Therefore, using the variable block size according to the motion it can improve the performance of motion compensation and such result is validated with the analysis.
Third, if the intensity function is nonlinear along the spatial coordinate, the error on the measurement of gradients will be drastically increased. To solve this problem, we propose a new spatial gradient, which is adaptively changed according to the estimated motion, to minimize the motion constraint equation error. Through the analysis and the experimental results, we confirmed that the proposed technique has better performance than the conventional methods in the case of the severe noise and large motion.
Finally, to use efficiently the motion, the estimated motion must be checked in its correctness. Therefore, in the thesis we propose a new algorithm using the error surface, which is constructed with the candidate vectors, for the measurement of correctness. With the simulations, the proposed technique has convincible results in performance.