Research On Semantic Segmentation Method For Point Clouds Based On 3D Shape Saliency Features

In recent years,research on 3D vision has become a focus of current research,with the rapid development of 3D data acquisition devices.Point clouds,the basic representation of3 D data,that can well describe the spatial geometric information of 3D objects or scenes.Semantic segmentation of 3D point clouds has important applications in fields such as autonomous driving,industrial control,and remote sensing mapping.3D shape is one of the most effective methods for studying semantic segmentation of 3D point clouds,and this paper combines deep learning methods to study semantic segmentation of 3D point clouds using 3D shape.(1)Semantic segmentation of 3D objects is mainly divided into coarse-grained and fine-grained,and fine-grained semantic segmentation has a wide range of applications in mechanical arm grasping and industrial inspection.For fine-grained semantic segmentation,this paper proposes an Fine-grained semantic segmentation network for enhancing local salient of laser point clouds(ELSFNet)based on enhanced point cloud local salient features.ELSFNet improves the furthest point sampling algorithm using the geometric curvature of point clouds,enhancing the feature calculation ability of local data areas;at the same time,it creates a multi-scale high-dimensional feature extractor to extract high-dimensional features at different scales.The attention mechanism is introduced in a seq2 seq manner to fuse multi-scale high-dimensional features and obtain semantic segmentation context information.Finally,experiments are conducted on the public dataset Shape Net,and the average m Io U of ELSFNet is 85.2%,which is higher than the current mainstream networks.(2)Point cloud data contains rich geometric information,and different geometric information performs differently in semantic segmentation tasks.3D shape is an important means of studying semantic segmentation of 3D point clouds.This paper explores the impact of different geometric information on point cloud semantic segmentation accuracy and proposes an algorithm based on unsupervised geometric information encoding.This algorithm extracts geometric features of point clouds,such as curvature,density,keypoints,and normal vectors,and uses unsupervised methods to group them based on point cloud geometric features.Neural networks are used to encode each region.Finally,comparisons are made with the current popular grouping method k-nearest neighbors and encoding method octree encoding on the public dataset Model Net40,and both have certain advantages in accuracy.(3)For point cloud semantic segmentation and classification tasks,this paper proposes an Semi-Supervised Semantic Segmentation Network for Point Clouds Based on 3D Shape(SBSNet)combining shape encoding.SBSNet uses furthest point sampling improved based on geometric curvature,grouping through spherical queries,to capture local multi-scale features.The shape encoding information is used as a self-prior module,and the attention mechanism is used to fuse local features with shape encoding to achieve point cloud semantic segmentation and classification.Finally,comparisons are made with current popular networks on public datasets Model Net10,Model Net40,and Shape Net,and SBSNet has certain advantages.