Determining the patient’s position is the premise and key of surgical navigation. Currently, most work is done by establishing sparse markers to locate the patient. In contrast, point cloud registration can obtain more accurate target positions and states by converting point clouds collected at different times and from different perspectives to the same coordinate system through rotation, translation and other operations. However, in the actual registration process, the collected point clouds are usually incomplete clouds from a certain perspective and contain noise interference, resulting in a trade-off between stability and speed in the coarse registration process. In this paper, we segment the complete point cloud from different perspectives. Then, we calculate the point distribution characteristics of each part according to the point cloud center of gravity and normal vector and use the characteristics of the actually captured point cloud to quickly infer possible perspectives. In addition, we suggested another way which combine 2D facial features with a 3D point cloud during the registration process. We extract facial features from 2D images to infer the head position and orientation to replace the traditional coarse registration process. Compared with the existing registration methods, the proposed methods reduce meaningless matching through more stable distribution features and 2D features, thereby improving the registration speed. Our methods have been tested to obtain stable registration results in less time.