A point cloud registration algorithm based on structured light and CT images is proposed to address model sensitivity to noise and low accuracy issues in medical image registration using existing deep learning algorithms. The primary aim is to improve registration accuracy and algorithm robustness whilst avoiding radiation caused by X-rays during image acquisition. First, project encoded structured light onto the body surface of an intraoperative patient, and principal component analysis is used to obtain the point cloud for the body surface after spindle correction. Second, CT scans are performed on preoperative patients and three-dimensional reconstruction and sampling are conducted to obtain their body surface point clouds. Finally, a dynamic graph convolutional network model based on feature reuse and attention mechanisms is constructed, combined with an image iterative registration algorithm for point cloud registration on the back of the human body. The entire process effectively integrates the information from the two modes, with the advantages of no radiation, high accuracy, and short time consumption.