There are two distinct categories in the computation of motion and structure of objects in a scene from a sequence of images; the feature based approach and the optical flow based approach. In the latter approach, which is concerned with this thesis, the instantaneous changes in brightness values in the image are analyzed to yield a dense velocity (or displacement) map called optical flow. The three-dimensional motion and structure parameters are then computed. There are two main causes to make the estimated optical flow inaccurate, i.e., the erroneous measurements of spatial gradients and temporal gradient of image brightness values with which one constructs a constraint relation called the motion constraint equation, and the blurring of motion boundaries which is caused by the other constraint called smoothness constraint.
In this dissertation an accurate optical flow estimation method is proposed, which is based on the Kalman filtering and improved temporal gradient measurement technique. The measurement errors of spatial gradients are analyzed and it is shown that the best measurement technique is to take a motion compensated interframe average (MCIA) for the spatial gradients. The measurement error of temporal gradient is also analyzed and a new temporal gradient measurement technique is proposed. It is shown that this new technique, named extrapolated frame difference (EFD), make less error than the conventional technique-the frame difference(FD). Besides it is shown that EFD is closely related to the displaced frame difference (DFD) which has been utilized for reducing the optical flow estimation error caused by the nonlinearity of the image brightness function. When EFD and DFD are individually applied to the iterative optical flow estimation based on the global optimization technique, both results show almost the same performance and show much better performance than FD in the sense of the convergence speed and the steady state error. As DFD is not a temporal gradient measure, it can not be applied to recent techniques for motion boundary detection in which motion boundaries are inferred using the spatial and the temporal gradients. In this case EFD also shows better performance than FD in artificial and real image sequences. Therefore EFD should be selected as temporal gradient measure.
Another study of this thesis shows that a pixel velocity can be estimated by combining the average velocity of neighboring pixels and a velocity observation at the pixel, that is, by Kalman filter form. Hence we can regard the optical flow estimation problem as two independent parts; the prediction part and the observation part. Using the generalized least square error estimation (GLSE) in the observation part, the optical flow estimation can be improved. A fast convergent iterative method for optical flow estimation is devised with this new Kalman filter model with the observation based on GLSE and the new temporal gradient measurement method mentioned, EFD. This new method is compared with the two existing techniques in artificial and real image sequences and shows the best performance in the sense of the convergence speed and the steady state error. Finally the proposed method is applied to the analysis of the blood flow in the artificial heart.