Abstract: One algorithm to automatically register point cloud data in 3-D model reconstruction was proposed. In the algorithm, the initial registration transformations were firstly calculated according to the curvature and normal vector of the point cloud. Then the best initial registration transformation was selected by geometric hashing method and the initial registration was completed by this transformation. Next,the ICP (iterative closest point) algorithm was improved by redefining the nearest point: For one measured point of A ,the nearest three points in B point cloud were firstly found; the triangle was formed by this three points and the foot point of this triangle was taken as nearest point of measured point. Then the improved ICP algorithm was used for accurate secondary registration. Finally,this algorithm was used to register the point data in previous model. And the results show this algorithm has the better performance in registration accuracy.