Deep Learning Point Cloud Semantic Segmentation Assisted Dense Mapping Research

Point cloud is a map represented by a set of discrete points.The most basic points include XYZ three-dimensional coordinates,and can also have RGB color information.The point cloud is just some points that are independent of each other.There can be hundreds or thousands of them,sparse and disordered.Deep learning methods can be used to construct dense 3D maps with semantic information.As a powerful support for Simultaneous Localization and Mapping(SLAM),it has a wide application value for automatic driving.PointNet is a unified architecture.It is the first to directly use the point cloud as input,and output each classification label or each point label.The result has achieved advanced results.However,the PointNet method has not been proven to be able to segment large-scale point clouds.In practical applications,the number of point clouds for lidar scanning results is in the millions.Meanwhile,the PointNet uses each point as input independently and fails to take advantage of the point cloud local structure.In response to these two shortcomings,this thesis proposes two improvements:1.Based on the superpoint graph network,this thesis proposes a method for preprocessing the PointNet input point cloud.Through the operations of geometric segmentation and superpoint embedding,the number of large-scale point clouds is effectively reduced,and a function of simultaneously processing a million-scale point cloud is completed,achieving a segmentation accuracy of 77.12%,which is both robust and low complexity.2.Based on Kernel Correlation Network(KCNet),this thesis proposes an improved semantic segmentation algorithm using point cloud local structure.KCNet effectively improves the accuracy of PointNet in point cloud classification and segmentation through fine-grained structure.This thesis proves that it is as efficient in semantic segmentation task and improves the accuracy to 80.6%.In addition,the improved algorithm assists the ORB-SLAM2 with dense point cloud mapping,which makes the SLAM system have the function of semantic mapping.3.Finally,this thesis proposes an improved algorithm for the Recurrent Slice Network(RSNet).The slicing network,through the orderly processing of the disordered point cloud,enables the traditional deep learning network to have the ability to process point clouds,and has achieved excellent results in object classification,semantic segmentation and partial segmentation.This thesis proposes to improve the local dependency module by using Deep Bidirectional Recurrent Neural Network(Deep BRNN)with more powerful learning ability,which improves the segmentation accuracy.