Accurate measurement of mechanical parts can provide a basis for parts processing and manufacturing, virtual simulation, defect detection and other applications. The binocular vision method based on fringe projection is widely used in the topography measurement of mechanical parts due to its high-precision characteristics. In this method, the quality of the absolute phase curve determines the final accuracy of the 3D reconstruction. Due to the influence of environmental noise, discontinuity of the scene, and complex surface structure of the parts, absolute phase curve often has a large number of noise points and varying degrees of deformation. In order to improve the reconstruction accuracy, a corresponding improvement method is proposed for the problems existing in the absolute phase curve. Aiming at the noise points far away from the phase curve that are difficult to remove, a region-by-region phase correction method is proposed. This method divides the phase curve of each row into several regions, and uses the median of the phase value in the region as the threshold to remove such noise points. For the deformed area of ​​the phase curve, the curve fitting method is introduced on the basis of the point primitive-based stereo matching method, and the disparity map obtained after the matching is refined, which can improve the poor matching result in the deformed area of the phase curve and improve the robustness and accuracy of the matching method. The method is used to reconstruct complex part FSW inverter, and the standard deviation can be reduced to within 0.1 mm. Experiments have proved that the reconstruction accuracy of this method is higher. For mechanical parts with complex surface structure, the proposed method can achieve accurate measurement.