Energy-based automatic recognition of multiple spheres in three-dimensional point cloud

Abstract

The emerging RGB-D sensors make three-dimensional data acquisition more economy and flexible. However, an ensuing great challenge is how to efficiently and effectively understand and apply the huge amount of three-dimensional data. Automatic recognition of primary geometric shapes in three-dimensional point cloud, such as spheres, can provide some abstractions and possibly semantic information to solve the challenge. An energy-based method for automatic recognition of multiple spheres in three-dimensional point cloud is proposed from the point of view of data labeling. It first generates initial sphere models by randomly sampling. Second, it constructs the energy function, and then labels three-dimensional points by minimizing the energy. Third, it refines obtained labels and parameters further. Four, it iterates the above steps until the energy does not decrease. Finally, multiple spheres are recognized from three-dimensional point cloud. Experiments with synthetic and real data validate the proposed method. It outperforms the Hough-based and the RANSAC-based method in accuracy and robustness. More importantly, it alleviates the dependence of existing algorithms on distance thresholds, the requirement of the unknown number and parameters of spheres, and the requirement of a huge sampling number to generate initial models.