Simultaneous identification of points and circles: structure from motion system in industry scenes

To address the issue of dynamic 3D circular features recognition in robot arm grasping, we propose a feature-based and incremental simultaneous points and circles structure from motion system. First, we represent the 3D target scene with sparse point cloud from multiple observations. The fundamental points construction pipeline combines the benefits of feature-based mapping and probabilistic depth estimation, which reduce the computational cost of generating practical 3D structures. Second, accurate 3D circles are extracted from the keyframes and are optimized in the backend. The circles construction pipeline attempts to find potential circles in the produced sparse point cloud and apply an adjusted level set method to do a novel 3D circle optimization process, which can work on keyframes smoothly. The integrated system is compared with other real-time construction systems and outperforms in industry scenes with more stable land-marks. Meanwhile the experimental results illustrate the ability of capturing circular features in target scenes.