X-ray fluoroscopic images are widely used for evaluating the knee kinematics after the total knee Arthroplasty, TKA in short. There are some conventional methods for estimating the motion of the knee implant in vivo using single-plane fluoroscopic images and 3-D geometric models of the knee implant. However, these methods have a problem that estimation results are dependent on the initial pose and position given at the start frame of sequence for analysis because these methods analyze the motion of a series of the knee implant for every frame based on the pose and position estimated at the previous frame. This paper proposes a 3-D/4-D registration method that estimates the initial pose and position by predicting dynamic knee kinematics. To quantitatively evaluate our method, it was applied to computer-simulated images and phantom images that took the knee implant in vitro fixed with arbitrary pose and position by a jig. The experimental results showed that the proposed method could estimate the pose of within the error of 0.04 deg in the computer-simulated images and 0.82 deg in the phantom images, and the accuracy was improved in comparison with the conventional methods.