This paper presents an on-line pose measurement method of a 3D object. The proposed method utilizes an on-line evolutionary search technique of the genetic algorithm (GA) in pattern recognition and a fitness evaluation based on matching input dynamic images with stereo models whose poses are expressed by unit quaternion. To improve the dynamics of recognition, a motion-feedforward compensation method is proposed for the hand-eye system by predicting target object's motion in camera frame induced by the hand-eye motion through the kinematic calculation of robot's motion. The effectiveness of the proposed method is confirmed by simulation experiments.