In order to improve the speed and accuracy of point cloud matching in the structured light three-dimensional reconstruction system, a collaborative matching method of two-dimensional view and three-dimensional point cloud across feature points is proposed in this work. First, the normalization of the projected images to be spliced is realized through projection transformation and dimension mapping. After preprocessing, the endpoints and bifurcation points are extracted as key points, and the similar points are triangulated and similarly matched to obtain the initial point set. The initial point set is mapped to three-dimensional space. Second, kd-tree search is used to obtain the centroid of the double neighborhood, and the point set is further screened according to the triangle similarity relationship formed by the three points. Finally, the quaternion method is used to complete the rough splicing, and then an improved iterative closest point (ICP) algorithm is used to complete the fine splicing. Experimental results show that the matching accuracy of the proposed algorithm is 98.16%, the matching time is 3 s, and the center of gravity distance error of the coarse splicing overlap area is 0.018mm. The proposed algorithm has high robustness for two-dimensional image perspective transformation, smooth texture, and uneven light.