基于超体素的区域聚类的复杂场景分割

Complex scene segmentation based on supervoxel region clustering

针对目前算法对复杂场景分割效果不理想问题,本文提出了基于超体素的区域聚类算法。该算法首先利用最新的算法对点云过分割得到超体素;其次通过计算种子超体素与其第一个邻接块的高程差的方式,把即将合并的区域判定为与地面或与其平行平面和其他面,分别采用不同的法向量夹角阈值和自动获取的正交距离阈值进行当前聚类区域与相邻块之间的相似性度量合并,在聚类过程中为了防止欠分割现象的发生,对每次有新的超体素加入时,就更新当前聚类块的法向量等几何信息;最后利用区域生长和改进前后的算法进行了定量和定性的比较,实验分析表明本文算法对复杂场景分割具有准确性和可靠性。

In this paper,for the problem that the current algorithm is not ideal for complex scene segmentation,a region clustering algorithm based on supervoxel is proposed.Firstly,by using the latest algorithm,the hypervoxel is obtained by over-segmented the point cloud.Secondly,by calculating the elevation difference between the seed supervoxel and its first adjacent block,the area to be merged is determined as planes or parallel to the ground and other surfaces,The similarity measure between the current clustering region and the adjacent blocks is merged by using different normal vector included angle threshold and automatically obtained orthogonal distance threshold.In order to prevent under segmentation in the clustering process,each time when a new supervoxel is added,the geometric information such as the normal vector of the current cluster block is updated.Finally,both quantitative and qualitative comparisons are performed through two types of algorithms.The experimental results show that this algorithm is accurate and reliable for complex scene segmentation.