基于CSegNet的三维点云室内语义分割研究

Research on indoor semantic segmentation of 3D point cloud based on CSegNet

人工智能的发展程度取决于机器对环境的感知能力,即机器对场景理解的能力,这就离不开其中的核心技术—三维点云语义分割,虽然基于深度学习的三维点云语义分割网络层出不穷,但都存在局部特征利用不充分的问题,导致分割效果不佳。本文针对这个问题,受PointNet++和TSegNet的启发,设计了具有双层结构的CSegNet模型。首先,编码部分使用逐点卷积来取代PointNet层捕捉点云的局部特征;其次,解码部分使用PointDeconv层恢复原始点云结构;最后加入边缘卷积来更有效的提取局部特征,并解决边界不平滑问题。在S3DIS数据集上验证表明,CSegNet能更好的利用局部特征来进行分割,最终得到oAcc为88.7%,mIoU为73.7%的分割准确性。

The development of artificial intelligence depends on the machine′s ability to perceive the environment, that is, the machine′s ability to understand the scene, which is inseparable from the core technology-3D point cloud semantic segmentation.Although there are numerous 3D point cloud semantic segmentation networks based on deep learning, they all suffer from the problem of insufficient utilization of local features, resulting in poor segmentation results.For this problem, a CSegNet model with a two-layer structure inspired by PointNet++ and TSegNet is designed in this paper.Firstly, for the encoding part, the point-by-point convolution is used to replace the PointNet layer to capture the local features of the point cloud;secondly, for the decoding part, the PointDeconv layer is used to restore the original point cloud structure;finally, the edge convolution is added to extract local features more effectively and solve the boundary non-smooth problem.Validation on the S3DIS dataset shows that CSegNet can better utilize local features for segmentation, resulting in a segmentation accuracy of 88.7 % for oAcc and 73.7 for mIoU.