基于机器学习的室内点云语义分割——以门为例

Semantic Segmentation of Indoor Point Cloud Based on Machine Learning——Take the Door as an Example

精确的室内三维点云数据为室内建模等领域提供数据支持。因此,在获取室内点云后,需要对室内不同的物体的点云数据进行语义分割,以方便后续的建模等处理。本文利用特征工程的思想,在原始点云7维特征的基础上,人工构建点云12维特征空间。选取室内门的30000个点云数据作为训练数据,以整个房间中的2000000个点云数据作为测试数据,对所有点云数据构建12维特征空间,使用线性向量机模型进行预测,并且研究使用7个的特征和12个特征对分割精度的影响。实验结果表明:采用线性模型,在12维的特征空间上分割的精度较高。

Accurate indoor 3D point cloud data provide data support for indoor modeling and other fields.Therefore,after obtaining the indoor point cloud,the point cloud data of different objects in the room need to be semantically segmented,so as to facilitate the subsequent modeling and other processing.In this paper,based on the idea of feature engineering,based on the original point cloud 7-dimensional features,the 12-D feature space of point cloud is constructed manually.30,000 point cloud data of the indoor door are selected as training data,and 2000000 point cloud data in the whole room are used as test data,a 12-dimensional feature space is constructed for all point cloud data,and a prediction is performed by using a linear vector machine model,7 features and 12 features on the segmentation accuracy.Experimental results show that using linear model,the segmentation accuracy in 12-dimensional feature space is high.