基于语义连通性的室内RGB-D深度图像超体素融合与分割

Super-Voxel Fusion and Segmentation of Indoor RGB-D Depth Images Based on Semantic Connectivity

室内复杂场景三维点云数据具有粘连性和非结构性,针对传统方法对室内场景进行三维重建不能较好地分离其所包含物体的问题,本文提出了顾及语义连通性的RGB-D深度图像超体素分割与融合方法。首先,从空间位置、颜色特征、结构特征、法向量、部分与整体的关系等方面定义语义连通规则;其次,实现了RGB-D深度图到三维点云的转换,并对转换后的点云构建了Kd-tree索引;最后,利用点云过分割算法VCCS对深度图点云进行超体素分割,并采用局部凸包连接算法融合生成的超体素顾及语义连通性。实验结果表明,当Rvoxel=0.007 5,Rseed=0.8,wc=2.0,ws=0.4,wn=1.0时,场景分割效果最为理想,且场景中的桌面及其上面的铁盒、碗及柱状物体都能较好地区分。

The 3D point cloud data of indoor complex scenes are adhesive and unstructured,and for the problem that traditional methods for 3D reconstruction of indoor scenes cannot better separate the objects they contain,this paper proposes a method for super-voxel segmentation and fusion of RGB-D depth maps that takes into account semantic connectivity.Firstly,the semantic connection rules are defined from the spatial position,color feature,structure feature,normal vector,the relationship between part and whole,etc.Secondly,the conversion of RGB-D depth map to 3D point cloud is realized,and build the transformed point cloud the Kd-tree indexes.Finally,the over-segmentation algorithm of VCCS is used to segment the depth map point cloud,and the Locally Convex Connected Patches is used to integrate the generated supervoxel with semantic connectivity.The experimental results show that when Rvoxel=0.0075,Reed=0.8,w.=2.0,w,=0.4,w,=1.0,the scene segmentation effect is the best,and the tabletop and the iron box,bowl and columnar object above the scene can be well distinguished.