Based on the homography between a multi-source image and three-dimensional (3D) measurement points, this letter proposes a novel 3D registration and integration method based on scale-invariant feature matching. The matching relationships of two-dimensional (2D) texture gray images and two-and-a-halfdimensional (2.5D) range images are constructed using the scale-invariant feature transform algorithms. Then, at least three non-collinear 3D measurement points corresponding to image feature points are used to achieve a registration relationship accurately. According to the index of overlapping images and the local 3D border search method, multi-view registration data are rapidly and accurately integrated. Experimental results on real models demonstrate that the algorithm is robust and effective.OCIS codes: 100.0100, 120.0120, 120.2650, 150.0150. doi: 10.3788/COL201210.091001.Three-dimensional (3D) registration and integration are difficult but necessary steps in performing structured light 3D measurements [1] . A multiple-view measurement is needed to fully cover an object because of object occlusion, shadows or limited view depth, and so on. The overlapping layers of the same object must be integrated before the whole object is completely measured. The goal of registration is to determine the transform function of a view from one coordinate system to another [2−8] . Different methods of registration have been developed [2−8] , and can be divided into three categories. The first method involves the use of supporting hardware, such as turntable methods [2] , gyroscopic self-positioning devices [3] , etc. This method relies on precision supporting devices, and has a limited application range. The second method adds some additional features to the surface of the object and identifies the designed features to achieve registration. This method includes the paste landmark technique [4] and fixing the calibration plate on the object [5] , among others. However, one main drawback of this method is that the object may not be completely measured because of the obstruction caused by the additional feature. The third method is completely based on the 3D measurement data. This technique starts from an approximate registration, and iteratively reaches the final registration by minimizing an error functional, such as the iterative closest point algorithm [6] and various improved algorithms [7−8] . This method has the smallest measurement constraint and is most widely applied. However, its biggest drawback is the requirement of a good initial position; otherwise, convergence is not guaranteed [7−8] . To maximize the convenience of application, 3D measurement data and the related measurement process are used to simplify the complex problem. The matching scale-invariant features from two-dimensional (2D) gray images and two-and-a-half-dimensional (2.5D) range images are accurately recognized. The homographic relationship is then used to solve 3D registration and integration. The novelty of this approach is characterized by the fo...