点云轮廓特征点是确定物体几何形状的关键，在目标检测和定位等邻域发挥着重要作用。本研究的目的是直接利用点云邻域特征来提取点云轮廓特征点。首先，利用Cholesky分解确定主、次特征向量，并分别构建基于主、次特征向量作为法向量的折边和边界点提取的邻域特征投影平面。然后，通过构建点的邻域维度特征信息熵模型来确定最佳的邻近点数，分析投影平面上目标点与邻域点构成的向量夹角分布特性，并基于方位角的特点，提出基于邻域特征分布的边界点精细提取方法。最后，提出基于四元素的投影平面上邻域点的二维视图形成方法，建立二维视图下基于点到直线距离及直线两侧点数偏差的多参数的折边点提取模型。实验结果表明，本文方法优于有序点霍夫变换法，面分割法和二值图像法。在抗噪声性能方面，本文方法能够在不同噪声情况下提取出轮廓特征点，且稳健性优于二值图像法、区域聚类曲率法和区域增长法。另外，本文方法的准确率、召回率及F1分数都高于90%，本文方法的F1分数比区域聚类曲率法高了4.2%，比霍夫变换法高了32.4%。而且，本文方法不仅适用于规则平面形状的建筑，也适用于不规则曲线形状建筑的轮廓特征点提取。

The feature points of point cloud profiles are the key to determine the geometry of objects, and play an important role in target detection and location. The objective of this study is to extract the point cloud contour feature points directly by using the point cloud neighborhood features.First, the Cholesky decomposition was used to determine the main and secondary eigenvectors, and the neighborhood projection plane based on the main and secondary eigenvectors as normal vectors was constructed respectively. Secondly, the optimal number of neighboring points is determined by constructing the entropy model of neighborhood dimensional feature information, and the angular distribution characteristics of the vector composed of target points and neighborhood points on the projection plane are analyzed. Based on the characteristics of the azimuth Angle, a fine extraction method of boundary points based on neighborhood feature distribution is proposed. Finally, a two-dimensional view formation method of neighborhood points on the projection plane based on quaternion method is proposed, and a multi-parameters extraction model based on the distance from point to line and the deviation of points on both sides of the line is established. Experimental results show that the proposed method is superior to ordered point Hough transform, patch segmentation and binary image methods. In terms of noise immunity, the proposed method can extract contour feature points under different noises, and its robustness is better than that of binary image method, region clustering curvature method and Regional growth method. In addition, the accuracy rate, recall rate and F1 score of this method are all higher than 90%. The F1 score of the proposed method is 4.2% higher than that of the region clustering curvature method and 32.4% higher than that of the Hough transform method. The conclusion that the method in this paper is not only suitable for regular planar building shapes, but also suitable for extracting contour feature points of irregular curved building shapes.