Research On Recognition Of Multi Geometric Primitives In 3D Point Cloud Based On Energy Optimization

With the rapid development of RGB-D sensor technology, 3D data acquisition becomes more economy and flexible. However, how to effectively and efficiently understand and apply the massive 3D data is becoming a challenge that needs to be solved urgently. Automatic recognition of geometric primitives in three-dimensional point cloud, such as plane, sphere and cylinder, is a basic problem of computer perception of the world. Solving this problem can alleviate that challenge in a certain extent, and more important is that it can reduce the difficulty of the computer to perceive the environment and narrow the semantic gap between the high-level semantics information and low-level visual features. It can further help the computer to analyze and perceive the world like human beings. In practical applications, it is helpful for the robot to complete intelligent operation. Therefore, automatically recognizing geometric primitives in the 3D point cloud has an important scientific significance and application value.To recognize multi geometric primitives in 3D point cloud, a new energy-based optimization framework is proposed from the view of data labeling. Firstly, it generates a set of initial geometric primitives by randomly selecting point sets; secondly, it calculates the geometric primitives energy, labels the 3D points and minimizes the sum of energy; thirdly, it refines obtained labels and parameters of geometric primitives; finally, it iterates above steps until the energy does not decrease, and outputs labels of 3D points, inliers and parameters of geometric primitives. By this framework, the recognition algorithms of plane, sphere and cylinder are proposed. Experiments with synthetic and real data validate the proposed method in stability and accuracy. On the basis of the above, three kinds of recognition algorithms of plane, sphere and cylinder are fused, and a algorithm of recognizing multiple geometric primitives simultaneously is proposed. In the end, a system is designed and implemented, which can recognize multiple geometric primitives in 3D point cloud, such as plane, sphere and cylinder.The proposed method outperforms the Hough-based and the RANSAC-based methods in accuracy and robustness. More importantly, it alleviates the dependence of existing algorithms on distance thresholds, and automatically determine the number of geometric primitives in 3D. It overcomes the limitation of the existing energy-based model fitting algorithm which needs to generate more initial models. It improves the flexibility and expands the application range of energy optimization algorithm.