Research On Point Cloud Scene Semantic Segmentation Method Based On Deep Learning

As a commonly 3D processing data,point cloud contains more geometric structure information than 2D image,and has widely used in numerous fields.Semantic segmentation technology for point cloud has become a research hotspot in numerous such as autonomous driving scene reconstruction.The existing semantic segmentation methods of point cloud scene based on deep learning,pay more attention to the extraction of local geometric features between points,ignoring the semantic relationships and spatial distribution information between features.At the same time,the obtained features cannot be fusion effectively,resulting in inaccurate semantic segmentation results.To solve those problems,this thesis designs two semantic segmentation methods for point cloud scene based on deep learning,as follows:(1)A point cloud scene semantic segmentation method based on multi-feature fusion has proposed.It takes the 3D coordinates of points and the optimized edge feature as the network’s inputs.Firstly,based on the difference in coordinates between the 3D coordinates and adjacent points,the geometric offset is extracted through a multi-layer perceptron to enhance the local geometric feature.Then,the enhanced local geometric feature are used to calculate the offset of semantic to enhance the local semantic feature,and make the network learn the local feature better.Then the spatial attention is introduced to extract the global semantic correlation of feature.Finally,attention pooling is used to effectively aggregate the feature learned by each module.The experimental results on S3 DIS data show that,the m Io U of the network model reaches 67.5% and the OA reaches 87.2%.Compared to DGCNN,it has improved by 10.6%and 2.9%,respectively,and the segmentation performance is better.(2)A point cloud scene semantic segmentation method based on orientation coding has proposed.Firstly,orientation coding is introduced to extract local feature in different orientations to reduce the impact of point cloud sparsity.Secondly,the volume ratio insensitive to the point position is used to construct the global feature to improve the tolerance of geometric deformation of similar objects.Then,the spatial geometric feature has embedded into the network model to optimize the capturing ability of salient features,and the aggregation point features are pooled by attention.Finally,an adaptive feature fusion strategy is proposed to reduce the feature loss in the decoding stage.The experimental results on S3 DIS data show that the m Io U of the network reaches 69.2%,which is 12.3% higher than DGCNN and 1.9% higher than Rand La Net.(3)Based on the two methods in this thesis,a semantic segmentation system based on Vue framework is designed and implemented.The system includes four modules: login,data preview,semantic segmentation,and result display.The system Implemented semantic segmentation model training and testing for point cloud scene,so that the user can easily perform the semantic segmentation of the point cloud scene.