Point Cloud Data Registration Method Study Based On Somatosensory Camera

With the development of society and the advancement of the IT industry, three-dimensional reconstruction technology plays an increasingly important role in various fields. In the three-dimensional reconstruction process, the registration of point cloud data is key technology, so the registration research of three-dimensional data is research focus in reconstruction technology. However, with the improvement of three-dimensional scanning equipment, data scale and registration accuracy have increased. The classical point cloud registration method does not meet the real-time needs, so based on the traditional ICP method, in this paper, a novel ICP algorithm based on Hausdorff distance and optimized KD tree is proposed. The main content of this paper is about registration of 3D point cloud data collected from multi-view into whole 3D point cloud data. Specific contents are as follows:First, a point cloud data from different angles of the object through Kinect camera, the data point of the main curvatures and Hausdorff distance as the basis, the use of point cloud distribution plot will be divided into key and non-key points, key points to retain the full point cloud geometric characFirstly, point cloud data from different angles of the object are get through Kinect camera. According to the data point of the main curvatures and Hausdorff distance, point cloud are divided into key points and non-key points by normal distribution graphics. Key points retaining the full point cloud geometric characteristics are used for point cloud registration, but non-key points,having unobvious geometric characteristics of points, do not match. Pretreatment of point cloud registration makes the registration more targeted.Secondly, the method of optimization KD tree of mid-value threshold segmentation is proposed, which is intended to improve the query efficiency of KD tree, ensuring KD tree becoimg its equilibrium state. First, attribute values of the node are sorted from smallest to largest, the middle finger is adopted as segmentation threshold, while point cloud data are structured into a KD tree binding spatial wide distributed. Optimal K-D tree makes the minimum number of layers, improving the efficiency of point cloud search.Again, the disadvantage about lower speed of complex data calculation in the traditional ICP algorithm has been improved, the improved Hausdorff distance and optimal KD tree are used to find matching point of key point to conduct least squares iterations until the proper registration to meet the correct convergence accuracy. The advantages of high-precision improves ICP algorithm suitable for mass data reverse engineering.Finally, in order to verify the correctness and effectiveness of this algorithm, comparing to the classical algorithm, several groups of different types of point cloud registration experiments are conducted. Experiments show that registration error of our algorithm is significantly lower than the classical algorithm and has an obvious advantage on the registration accuracy. When the amount of data gets increased, the superiority gets more obvious. The improved registration algorithm is applied to the objects which needs a higher registration precision and has a more obvious concavity and convexity.