A multi-object oriented iterative closest point algorithm in augmented reality

Abstract

The three-dimensional point cloud registration is a key technology of augmented reality and three-dimensional reconstruction. Also the Interactive Closest Point (ICP) is one of the most widely used algorithms of three-dimensional point cloud registration. However, the current ICP algorithms often come with an inaccurate initial value of the rotation matrix and the translation vector and cannot be used in scenarios with multiple objects. Therefore, we propose an ICP algorithm for multiple objects, which has an accurate initial rotation matrix and a translation vector. Firstly, the point cloud segmentation is applied to get multiple objects based on the Pass-through Filter and the Normal Estimation algorithms. The main spatial features involved are geometry and texture features. Secondly, the centroid of the regular point cloud data model such as cups, desks, etc. of each segmented point cloud, is calculated. Moreover, the singular value decomposition algorithm is used to obtain the rotation matrix of each point cloud model respectively. Finally, the translation vector of each point cloud model is obtained by combining the centroid and the rotation matrix. Experimental results show that the proposed method solves the problem of the inaccurate initial position and can be used for three-dimensional point cloud registration with multiple objects being compared with the existed ICP algorithms. At the same time, the ICP registration efficiency of a three-dimensional point cloud with single object using the proposed method is also improved about 5 percent.