Existing point cloud registration algorithms cannot solve problems of variable scale and registration accuracy of point cloud models simultaneously. Hence, this paper proposes a two-stage variable scale point cloud model registration algorithm. In the first stage of the algorithm, a dynamic scale factor is added to approximately estimate and adjust the scale of the target point cloud model. Spatial rotation transformation is then performed at three angles to divide the grid points, and the grid point spacing is set to 30°. This improves the convergence speed of the algorithm and prevents a local optimum, thus providing a good initial position for the second stage of registration. The second stage is optimized based on a scale iterative closest point (SICP) algorithm to match the point cloud model more precisely. A comprehensive comparison experiment is performed on different registration algorithms, and the experimental results show that in the case where there is a large rigid body transformation between two point cloud models and the scales are significantly different, the proposed algorithm has an order of magnitude of registration error of 10 ^-30--10 ^-4.