A new method for calibrating stereo vision measurement system is presented. By moving a rigid optical reference bar in a pattern in and around the measurement volume, the relative orientation and position of the two cameras can be determined by using the epipolar constraint, the linear normalization eight-point algorithm and the M-estimator method. The scale factor of the translation between the two cameras can be calculated by the distances between the marks on the reference bar. With automatically controlled infrared LED marks, the imaging feature points have a uniform intensity profile and a high contrast with background. The calibration accuracy can be improved. The simulation and experiments have shown that the calibration accuracy can be comparable with the complex off-line calibration.