A new system for robot vision is proposed that integrates a 3-D object recognition task and a 3-D object tracking task, enabling real-time 6-DOF localization of a known continuously moving object. A computational time-lag between the two tasks is absorbed by a large amount of frame memory. For 3-D sensing, calibrated trinocular stereo cameras are used to employ stereo-vision-based object recognition and tracking methods that allow objects of any shape to be dealt with and to provide robust performance in an environment with partial occlusions and cluttered backgrounds. The effectiveness of the system is demonstrated by experimental results.