基于点云数据的杂乱场景分割算法研究

目前处理桌面上单个对象已经被解决，然而处理复杂的场景时由于杂乱和遮挡而引起相当多的问题。虽然当前最先进的方法在基准性能上继续逐渐提高，但是它们也变得越来越复杂。针对杂乱场景中多个物体的分割问题，提出了一种基于RGB-D点云数据的分割方法。该方法先将场景点云超体聚类分解为基于体素网格的邻接图，然后对邻接图的边缘进行分类创建凸度图，再通过区域生长合并具有凸关系的分块从而得到未知物体。此外，提出用欧几里得算法对区域生长进行改进，发现对于碗和杯子这类具有内部凹面的物体有较好地分割效果。我们在对象分割数据库和手动提取场景中的实验结果，表明该方法可以在杂乱的桌面场景中分割各种形状的对象。

Research on Segmentation Algorithm of Scattered Scene Based on Point Cloud Data

Currently dealing with a single object on the desktop has been resolved, but dealing with complex scenes due to clutter and occlusion caused a considerable number of problems. While current state-of-the-art methods continue to improve in benchmark performance, they are becoming more complex. Aiming at the problem of segmentation of multiple objects in clutter scene, a segmentation method based on RGB-D point cloud data is proposed. The method firstly decomposes the scene cloud superclass cluster into the adjacency graph based on the voxel mesh, then classifies the edge of the adjacent graph to create the convexity graph, and then merges the convexity with the convexity through the regional growth to get the unknown object. In addition, it is proposed to improve the regional growth with the Euclidean algorithm and find that the objects with internal concave surfaces such as bowl and cup have a better segmentation effect. The results of our experiments in the object segmentation database and the manual extraction of the scene indicate that the method can segment objects of various shapes in cluttered desktop scenes.