To address issues such as incorrect feature point matching， missing matches， and duplicate matches in the traditional stereo matching of structured light-based 3D reconstruction， this study introduced enhancements to the Gaussian filtering in the SURF algorithm through the integration of adaptive median filtering with wavelet transform. Additionally， a secondary feature matching approach based on the OKG algorithm was proposed. The proposed algorithm first employed adaptive median filtering in conjunction with the wavelet transform algorithm to achieve image smoothing and noise reduction. Subsequently， preliminary feature point extraction and matching were performed. The scale space was then divided into multiple grids. Within each grid， the FAST algorithm was employed to extract scale space feature points， the ORB operator was utilized to extract feature points from the left and right images， and these points were described using BRIEF descriptors. The K-D tree nearest neighbor search method was applied to constrain feature point selection， and the GMS algorithm was utilized to eliminate mismatches. Finally， a comparative analysis was conducted between the SURF-OKG algorithm proposed in this paper and traditional feature matching algorithms. The effectiveness of the proposed algorithm was verified through the 3D reconstruction of step blocks. Experimental results reveal that the correct matching rate of the SURF-OKG algorithm is 92.47%. In the case of step blocks with a width of 40 mm and an accuracy of 0.02 mm， the mean error in width measurement is 1.312 mm， with no maximum error exceeding 1.72 mm， meeting the experimental requirements of the structured light 3D reconstruction system.