The Research Of Point Cloud Registration Algorithm Based On Parallel Computing

With the development and popularization of 3D scanning technology,3D application technology has been well developed. It has received a wide range of attention and extensive research especially in the field of reverse engineering, defect detection, protection of cultural relics, medical surgery,3D printing and game entertainment etc. In order to obtain a complete point cloud data, it needs multiple scans to splice and registration, because of partial occlusion and other issues. The 3D point cloud registration is the key technology of 3D scanning. This thesis mainly studies the technology of point cloud registration and GPU parallel technology, based on the previous research. This thesis mainly studies and implements the ICP parallel algorithm using the approximate nearest point and Kinect field scanning registration system. The main research results of this thesis are as follows:The efficiency of kd-tree search algorithm is low and its combination with ICP algorithm is not closely related to the search algorithm. Aiming at this problem, this thesis proposes the kd-tree approximate nearest point search algorithm based on ICP algorithm’s character and implements its GPU parallel algorithm. Through the non-recursive kd-tree contribution, the dynamic control kd-tree level and the control of the backtracking space, it completes the acquisition of the corresponding point set. Finally, the experimental results show that the approximate nearest point search algorithm can effectively improve the efficiency of the search for the set of corresponding points. As a result it improves the efficiency of the ICP algorithm.Aiming at the problem of low efficiency of ICP algorithm on the setting of large data and equipment limitations of the GPU algorithm nowadays, this thesis proposes to transplant the algorithm, which is implemented in this thesis, from CPU to GPU by using the OpenCL platform, to achieve the parallel approximate nearest point ICP algorithm. Finally, the experimental results show that the ICP algorithm in this thesis has better generality and efficiency in GPU.This thesis designs and implements the system of the Kinect field scanning registration, which is mainly to complete the model registration process through field data collection. Combined Kinect depth camera, feature extraction algorithm and the algorithm which is implemented in this thesis, the system includes data acquisition, 3D reconstruction, model preprocessing, and model registration. System experimental effect chart shows that the system can deal with complex scenes, and it is robust to noise.