Abstract: Registration between medical images and physical space is an important procedure in 3D surgical navigation systems. A registration method based on ICP (iterative closest point) algorithm was presented. Two point sets, one of which is acquired from physical space by the optical localizer and the other is from a 3D model reconstructed by the marching cube method, were used to calculate the transformation matrix between preoperative CT image space and intraoperative fiducial marker space. The registration procedure was divided into the following two phases. PCA(principal component analysis) was used to calculate the eigenvectors of the two point sets respectively to achieve initial registration. ICP method was used to make the initial transformation matrix to converge into the best solution, in which the k-d tree structure was used to accelerate the procedure. A spine model experiment was carried out and a virtual 3D measurement environment was set up to evaluate the accuracy of the registration method. Error analysis was conducted by adding Gaussian noise to the point sets. The result shows that under this condition the final average registration accuracy is less than 1mm.