In this paper, we propose a model-based image processing for real-time position estimation of a moving object. Based on the dynamic equation of a moving object, which is usually nonlinear, we construct an extended Kalman filter predicting the position at the next sampling step from a series of image data. The prediction of the position is effectively used to restrict the search area in the image plane, which shortens the image processing time and reduce the effect of background noise. In addition, since the effect of the quantization errors can be reflcted in the Kalman filter algorithm as observation noise, we can achieve near-optimal estimation by the appropriate choice of covariance matrices. To show the effectiveness of the proposed method, we present experimental results for the simple-pendulum motion estimation.