A new stereo precise matching method for free-form surface measurement with computer vision at subpixel level was proposed, which combined many algorithms. Firstly, using projection grating, distortion stripes were created on the curve surface. Some points of stripe edges were detected with wavelet transform. Secondly, a non-supervisor clustering searching fitting algorithm was discussed. So all of edge points were divided into different groups according to stripe edges, the edge points were fitted by interpolating cubic B-spline, and the fitted curves were recognized and marked. Lastly, corresponding points on every curve were exactly matched based on epipolar constraint, and the coordinates of the corresponding points were calculated. Therefore, the stereo matching at subpixel level for double cameras of converging optical-axis is completed successfully.