| As an important branch of computer vision,3D point cloud data understanding plays an important role in intelligent medical,smart city and autonomous driving,etc.However,due to the particularity of 3D point cloud,the traditional deep learning technique for 2d image processing is no longer applicable.This paper focuses on the semantic segmentation of 3d point cloud data and studies the parametric convolution design,deep learning network structure design and lightweight network structure design based on residual network.The main contents are as follows:Firstly,aiming at the problems of insufficient feature extraction ability and poor robustness in the current methods based on parametric convolution design,the cubic convolution model is proposed to improve the local geometric feature extraction ability of point cloud semantic segmentation model based on the concept of structure regularization of local point cloud.The multi-layer convolution framework was designed and the experimental research was carried out on the open data set to verify the effectiveness of cube convolution in point cloud feature extraction proposed in this paper,so as to provide guarantee for high-precision semantic segmentation of point cloud data.Secondly,based on the mainstream network framework in the 2D image field,cube convolution is combined with residual network to construct a 3D point cloud feature residual learning structure,so as to enhance the representation ability of the feature extraction network.The k-nn interpolation algorithm is combined with the feature pyramid network to build an up-sampling network on the feature of point cloud and realize the multi-scale semantic feature fusion of the high level feature and the low level feature.A comparative experimental study is carried out on the open data set to verify the performance of the 3d point cloud residual feature pyramid semantic segmentation model based on cubic convolution proposed in this paper in point cloud semantic segmentation,which laid a foundation for the optimization of network structure and further improvement of performance.Finally,aiming at the problems of too many parameters,low computing efficiency caused by larger model size,a lightweight cyclic grouping convolution is proposed and the u-net network principle and attention mechanism are introduced to the point cloud residual network design;A comparative experimental study was carried out on the open data set to verify that the proposed 3d point cloud attention residual semantic segmentation model can significantly improve the efficiency of the algorithm on the basis of ensuring the semantic segmentation accuracy of point cloud data. |
