基于多信息深度学习的3D点云语义分割

3D point cloud semantic segmentation based on multi-information deep learning

点云语义分割是三维点云数据处理的基础步骤,是三维场景理解分析、重建和目标识别的关键环节。针对目前对三维点云进行语义分割使用的点云信息少和精度不高的问题,本文在利用点云三维坐标信息的基础上,增加了点云RGB信息和所属房间的归一化坐标信息,从而丰富了神经网络输入端的信息量,进一步提高了模型的分割精度,最后利用PointNet++对改进后的三维点云语义分割效果进行检验,实验表明:在丰富了网络输入端的数据信息后,模型的总体准确度提高了6.65%。

Point cloud semantic segmentation is the basic step of 3D point cloud data processing,and is the key link of 3D scene understanding analysis,reconstruction and target recognition.In view of the current problem of lack point cloud information and low accuracy used for semantic segmentation of 3D point clouds,this paper adds point cloud RGB information and normalized coordinate information of the room to which it belongs based on the use of point cloud 3D coordinate information to enrich the amount of information at the input end of the neural network and further improve the segmentation accuracy of the model.Finally,PointNet++is used to test the improved 3D point cloud semantic segmentation effect.The experiment shows that after enriching the data information and inputting the network,the model accuracy is improved by 6.65%.