基于改进DGCNN的室内点云语义分割方法研究

Semantic Segmentation of Indoor Pointclouds Based on Improved DGCNN

室内点云场景广泛应用于虚拟现实和机器人导航等领域。然而,室内场景中物体的密集性、复杂性以及多遮挡等带来的数据不完整和多噪声问题,一定程度上限制了点云语义分割的准确度。针对上述问题,提出了一种基于改进DGCNN的室内场景语义分割方法。首先,利用K-近邻算法和点云的几何距离在动态图的基础上构造静态图,提升对动态图形中几何位置关系的利用程度;接着,对于点云数据不完整和多噪声的现象,引入了残差卷积模块,使模型可以捕捉更多层次的特征信息,提高特征提取的能力;其次,对于物体复杂性和多遮挡的现象,引入了通道注意模块和空间自注意力模块,提高模型的语义分割精度。在公开数据集S3DIS上的实验结果表明,本文模型相较于原始DGCNN模型对室内点云语义分割的整体mIoU指标提高了9.2%。

Indoor point cloud scenes are widely used in fields such as virtual reality and robotic navigation.However,issues such as data incompleteness and noise caused by the density,complexity,and occlusion of objects in indoor scenes limit the accuracy of point cloud semantic segmentation to some extent.To address these problems,an indoor scene semantic segmentation method based on an improved DGCNN is proposed.First,the K-nearest neighbors algorithm and geometric distances of the point cloud are used to construct a static graph on the basis of the dynamic graph,enhancing the utilization of geometric positional relations in the dynamic graph.Then,to address data incompleteness and noise,a residual convolution module is introduced,enabling the model to capture more hierarchical feature information and improve feature extraction capabilities.Furthermore,to tackle the complexity and occlusion of objects,channel attention and spatial self-attention modules are introduced,enhancing the accuracy of semantic segmentation.Experimental results on the public S3DIS dataset show that the proposed model improves the overall mIoU metric for indoor point cloud semantic segmentation by 9.2%compared to the original DGCNN model.