局部形状特征概率混合的半自动三维点云分类

A semi-automatic 3D point cloud classification method based on the probability mixture of local shape features.

三维激光扫描获取的点云数据可用于数字城市建设、三维模型获取、场景分析与物体测量等领域.但因遮挡和噪声的影响,加之扫描场景复杂,采样精度受限,使得不能直接运用经典的曲面和三维空间理论对点云数据进行有效分析和处理.分类是点云数据预处理的重要方式之一.提取近邻四面体体积、近邻法向量差异度、主方向差异度和主曲率值4个局部形状特征,采用概率混合策略构建了一种点云数据的半自动分类方法,可实现平面点集、柱面点集和其他点集的有效区分.其中,概率混合策略是依据近邻点平均距离和单指标类别一致程度估计每个特征推断形状的概率,通过混合加权,依据概率赋权函数最大值准则进行局部形状推断.可实现用户交互,以便处理不同扫描尺度和精度的点云数据.采用本文方法对模拟生成的点云、单棵树木点云、街道场景点云、旷野自然场景扫描点云以及航空机载扫描点云等多组数据进行了实验,结果表明,基于局部形状特征的概率混合方法对各种点云数据均具有良好的分类效果.

Point clouds captured by three dimensional scanner have been used in many fields, including modeling of digital cities, acquisition of three dimensional shapes, scene analysis and object measuring. However, due to the limitation of the sampling process and the complexity of scanned scenes, most traditional methods of surface model ing and three dimensional space analysis cannot work effectively when dealing with the point cloud data. Classifica tion is therefore an important way for point cloud preprocess. Four features, namely the volume of a tetrahedron constructed by 4 neighboring points, the deviation of normal directions of neighboring points, the deviation of prin- cipal directions of neighboring points, and the values of principal curvature, are mixed with probabilities for semi- automatic classification of the three dimensional point cloud data. With the new method, a point cloud is to be divid ed into three classes： plane points, cylinder points and other points. The initial classification result is labeled accord- ing to its single shape feature value. The probability mixture is completed by estimating the probability of inferring a shape from a local point set with respect to each feature, generating a mixture with weighted sum, and maximizing the mixture probability function, while the probability is estimated with the average distance between a point and its neighbor points together with the consistency ratio of initial labels of the point to its neighbors. User interactions are invoked to make the choice of classification thresholds and the setting of weights, which is helpful when dealing with point cloud with different space scale and scanning point resolution. Experiments show that the proposed method works well for various kinds of point cloud data sets, including point clouds generated by simulation, and those corresponding to a single pine tree, a street scene, a country scene, and an airborne big scene.