Object-driven Automatic Reconstruction

With the rapid development of robot technology and the extensive use of RGB-D cameras,the autonomous reconstruction has become a reality.The traditional robot scanning algorithm focuses on the real-time location and mapping of the robot.It is sel-dom involved in the quality of the 3D reconstruction and the understanding of the scene and need human's assistance to complete the scanning and reconstruction work.The ex-isting autonomous reconstruction algorithms generally require maps as prior knowledge in advance to guide the robot's actions.We can see that there are still many problems to be solved to fully construct the autonomous robot 3D scanning system.In order to achieve autonomous reconstruction,the key lies in understanding the object level of the scene.Specifically,the point cloud after reconstruction is analyzed online,and then the target object is extracted from it.The next move by robot is planned according to the analysis results.Among them,the main algorithms are object segmen-tation,object recognition and viewpoint selection.Although there have been many re-lated research,they either can not run in real-time or are not robust enough,especially for the incomplete input.In this paper,an object-driven autonomous reconstruction al-gorithm is proposed to solve the above problems.Some new or improved methods are applied in object matching,object segmentation and next-best-view(NBV)selection,and a robot autonomous reconstruction system is built which can complete the work in one pass.In this paper,an object-driven multi-level partial matching algorithm,an object-driven multi-class global segmentation algorithm and an object-driven NBV selection algorithm are proposed respectively.For the pre-segmented point cloud,this paper first uses a multi-layer local matching algorithm to calculate the matching rate.At the coarse level,we use a series of feature reduction methods to speed up.At the fine level,we carefully screen the correspondence to improve precision.According to the matching results,we use the global multi-class segmentation algorithm to get the global optimal segmentation of all the point clouds and extract objects.Finally,for the point cloud with matching rate under the threshold,a new NBV algorithm is proposed in this pa-per.We define the conditional information gain and select the optimal view using prior knowledge.In order to verify the correctness and robustness of the algorithm,this paper has passed sufficient experimental verification in each part to ensure that each algorithm has a certain degree of improvement in accuracy or speed compared with the existing algorithms.At the same time,the complete autonomous reconstruction system is tested in virtual and real scenes.The result is very efficient and convincing.