摘要
针对在复杂城市环境中难以有效自动提取树木信息的问题,该文首先基于移动激光扫描(MLS)数据,利用超体聚类对点云数据结构进行组织管理;然后从超体素局部上下文信息中提取去趋势几何特征,结合去趋势几何上下文特征,采用随机森林分类器对树木进行初始语义标记;接着,基于局部上下文信息进行迭代正则化,在全局图模型上进行整体优化,从而对初始语义分类结果进行空间平滑;最后根据语义标记结果,基于图割算法实现单木分割。该方法基于超体素的结构可以有效地保持场景中目标的几何边界,而且提升了处理效率。去趋势几何特征可以克服局部上下文中的冗余和显著性信息,使得获取的特征更具代表性。实验结果表明,该方法在3个数据集的树木语义标记结果达到90%左右,对结构简单且稀疏分布的树木都能正确提取。
Aiming at the problem that it is difficult to effectively extract tree information in complex urban environments,this paper proposes a method for automatically extracting trees in a complex urban environment from mobile laser scanning(MLS)data.The method uses supervoxel clustering to organize the point cloud data structure,and then extracts the detrended geometric features from the supervoxel local context information.Combined with the detrended geometric context feature,the tree was initially semantically labeled with a random forest classifier.Next,iterative regularization is performed based on the local context information,and overall optimization is performed on the global graph model to spatially smooth the initial semantic segmentation result.Finally,based on the semantic label results,the single-tree segmentation is implemented based on the graph cut algorithm.The method is based on the super-voxel structure to effectively maintain the geometric boundary of the objects in the scene and improve the processing efficiency.Detrended geometric features can overcome redundant and saliency information in the local context,making the acquired features more representative.The experimental results show that the proposed method achieves more than 90%of the results of tree semantic segmentation in the three datasets and correctly extract trees with simple structure and sparse distribution.
作者
李启才
赵闯姓
LI Qicai;ZHAO Chuangxing(Institute of Province,Xining 810001,China)
出处
《测绘科学》
CSCD
北大核心
2020年第9期117-122,147,共7页
Science of Surveying and Mapping
关键词
点云分割
移动激光扫描系统
超体聚类
图割算法
局部上下文特征
单木分割
point clouds segmentation
mobile laser scanning system
supervoxel
graph-based segmentation
local context
trees segementation
