| With the development of sensor technologies such as lidar,the cost of acquiring 3D point cloud data has been significantly reduced.3D point cloud semantic segmentation provides technical support for areas such as autonomous driving,target tracking,and 3D scene reconstruction.However,point clouds exist in the form of discrete points,which are unordered and lack texture information,posing certain difficulties for semantic segmentation tasks.The 2D image data captured by cameras possess rich texture information.Therefore,3D point cloud data and 2D image data have certain complementarity.In this paper,we propose a semantic segmentation network that combines the characteristics of 3D and 2D data and integrates point cloud and texture information.(1)To address the challenge of coordinate system inconsistency in the fusion of point cloud and image texture information,this paper proposes the Point-Select algorithm,IMPC-Fuse algorithm,and annotation algorithm.The intrinsic calibration algorithm is used to calibrate the industrial camera(SW-TC2020C-R315YQ),and the Point-Select algorithm assists in calibrating the Li DAR(SURVEY3D A260)to the camera’s extrinsic parameters,laying the foundation for point cloud texture information fusion.An edge detection algorithm is utilized to extract texture information from color images,while the IMPC-Fuse algorithm matches each point cloud in the 3D scene to the corresponding texture information in the 2D image,achieving the fusion of 3D point clouds and texture information.The labeling algorithm is used to label the semantic segmentation dataset.In this paper,the Fusion School(FSCH)dataset is created using hardware,and data fusion experiments and street scene segmentation experiments are carried out using the KITTI and FSCH datasets.It is demonstrated that the addition of texture information can enhance data representation capabilities and improve semantic segmentation accuracy.(2)In response to the issues of low segmentation accuracy and slow speed in point cloud semantic segmentation networks,this paper proposes the Point Cloud Graph Convolutional Network(PCGCN)for point cloud semantic segmentation.Firstly,the simplified graph convolution module of PCGCN achieves fast learning and training by reducing redundant information during feature aggregation.Secondly,residual connections are used for information transmission between feature layers,reducing information loss during feature transmission and enhancing feature learning to improve accuracy.Finally,experimental analysis is conducted on S3 DIS and Shape Net datasets,demonstrating that the simplified model has certain advantages in training time,and the accuracy of PCGCN is improved.The robustness experiment results show that the model has good robustness.(3)To address the lack of texture information in point clouds,this paper presents the Fusion Graph Attention Convolutional(Fusion-GAC)network.Fusion-GAC introduces a dual-channel graph convolution module,which converts texture information into chromaticity values and extracts local features of 3D point clouds.The dual-channel design allows the network to quickly extract data features and achieve high accuracy in a short time.Notably,Fusion-GAC combines spatial attention mechanisms to extract features and fuses multi-scale features to enhance the generalization ability of the network.Finally,on the S3 DIS dataset,Fusion-GAC achieves the OAcc of 91.58% and the MIo U of 78.55%.On the self-made Semantic segmentation KITTI(SSKIT)dataset,it is proven that adding texture information improves semantic segmentation accuracy,and the performance of Fusion-GAC is better than Point Net++ and DGCNN networks.In response to these results,this paper analyzes the impact of dual-channel graph convolution on speed,the impact of RGB chromaticity values on the accuracy,and the robustness of the model. |
