针对点云语义分割过程中存在的大量点云数据的相邻关系丢失以及无法捕获部分点云特征的关键信息等问题,提出了一种基于改进PointNet++的室内点云语义分割模型.首先利用中垂线通道采样获取到更具代表性的采样点,从而提高采样结果的信息丰富度;在此基础上使用采样点邻域特征自适应分组,使组内采样点的分布特征和邻域内的点云特征更加接近,然后引入注意力机制,以实现对点云数据的多层次、多维度的建模和表达;最后通过实验进行性能对比分析.实验结果表明,模型对室内点云进行语义分割相较于PointNet++模型的整体准确率提高了 5.6%,因此语义分割网络改进模块能够帮助神经网络提取到更优的点云特征信息,从而提高语义分割网络模型的性能.
In response to the prevalent issues encountered in the process of semantic segmentation of point clouds,such as the loss of neighboring relationships among abundant point cloud data and the inability to capture critical information pertaining to partial point cloud features,this paper proposes an enhanced indoor point cloud semantic segmentation model based on the improved PointNet++.Initially,a representative set of sampling points is acquired through the employment of a perpendicular channel sampling technique,thereby enhancing the information richness of the sampling results.Subsequently,adaptive grouping of sampling points within their neighborhood is executed,which enables a closer alignment of intra-group sampling point distribution characteristics and point cloud features within the neighborhood.Following this,an attention mechanism is introduced to facilitate multi-layered and multi-dimensional modeling and representation of the point cloud data.Lastly,performance comparative analysis is conducted through experimentation.The experimental results indicate that the utilization of this model for indoor point cloud semantic segmentation yields an overall accuracy improvement of 5.6%when compared to the PointNet++model.Consequently,the proposed semantic segmentation network enhancement module in this study facilitates the neural network's extraction of superior point cloud feature information,thereby enhancing the performance of the semantic segmentation network model.