As traditional measuring methods and equipment can not meet the requirement of 3D measurement of workpiece dimensions on-machine in real-time, an on-machine 3D vision measuring system based on the principle of binocular stereo vision was established to work at the workshop. Firstly, A stereo vision system calibration method by using a plane calibration plate was introduced. Then, threshold segmentation algorithm was used for recognizing and extracting the workpiece from the background and the Canny operator and polygonal approximation method were used for extracting the workpiece contour and the contour feature points. Furthermore according to the matching algorithm based on the epipolar constraint, a gray dependent based intensive fine matching algorithm was proposed and the 3D coordinate data of the workpiece contour key feature points were obtained. Finally, the data of matched feature points were processed and fitted, and the key dimensions of workpiece were obtained. On a Computer Numerical Control(CNC) machine in the workshop, an experiment to measure the 3D dimensions of a workpiece with a cant and holes shows that the measurement accuracy can reach ±1.3% or more.