Research On Kinect Point Cloud Scene Stitching Algorithm

Virtual reality technology (VR) has been widely used in the education field. The virtual scene modeling is a key problem to be solved in virtual reality. Based on the Kinect equipment, the Kinect Fusion algorithm can capture a real 3d scene reconstruction in computer. As reconstruction range of this algorithm is limited, this paper presents a technique of matching scene. The core issue of this technique is the 3d point cloud registration.Generally speaking,3d point cloud registration consists of two steps, namely the initial point cloud registration and the accurate point cloud registration. Dior Aiger proposed 4-points congruent sets for robust pairwise surface registration(4PCS algorithm), which is an initial registration algorithm. This algorithm makes no assumption about starting poses, and noise has no effect on registration. However this algorithm has a quadratic time complexity in the number of data points. This limits its applicability to acquisition of large environments. The most widely used Accurate registration algorithm is iterative closest point algorithm (ICP) algorithm. ICP algorithm searches the nearest distance point in two point cloud by finding corresponding to each other and multiply iterates objective function to complete point cloud data of the final registration. This algorithm has a simple function and fast convergence speed, but due to the registration of the ICP algorithm is proposed for surface, when point cloud contains planar, the algorithm is robustness.The main work and innovations are as follows:(1) In order to lower the quadratic complexity, a key change is maked to the original 4PCS algorithm. The method of changing the process of extracting a certain point pairs is as following:rasterizes the point cloud recursively and creates a sphere which centers on every point in the point cloud and uses a given distance as its radius. Determining whether the grid and the surface of the sphere intersect in a certain range of threshold value, if intersect, records the present grid and divides the child gird recursively until the size of the gird and the given threshold values are equal. Then verifies the point in the grid, if the verification is correct, the point pair is the point and the center of sphere. This method lowers the time complexity of extracting the point pairs, and improves the speed of initial registration. Though the registration accuracy of the improved algorithm has a small range of error, it greatly improves the registration speed compared with the original 4PCS algorithm.(2) In order to improve the registration precision of ICP algorithm, puts forward an improved scheme:before using ICP algorithm, preprocess the point cloud model. In the pre-processing step, extracts the plane model in point cloud and removes it, obtaining the point cloud without planar. Then uses ICP algorithm for accurate registration, which promoting the registration accuracy. The registration result shows that the improved ICP algorithm on point cloud with plane is more accurate than the original ICP algorithm.