Point Cloud Semantic Segmentation Based On Deep Learning And Research Of Dense Indoor Point Cloud Map

In recent years,the advancement of robotics technology has promoted the domestic robotics industry's rapid development.How to enable the robot to model the surrounding environment and complete the perception and understanding of the scene is an essential topic in robot intelligence research.In environmental modelling,Simultaneous Localization and Mapping(SLAM)realizes that the robot can know where it is and understand the current scene.Still,the dense point cloud map constructed by SLAM has noise points,and the lack of semantic information makes the robot unable to perceive and interact with the map.In scene perception,the neural network can learn the feature distribution in the space.Deep GCNs is a type of graph convolutional neural network.The deep graph convolutional neural network can complete the task of semantic segmentation of three-dimensional point clouds.However,it has some shortcomings in the prediction module and does not make full use of the connections between local features.Due to the limitation of the neural network itself,some semantic segmentation results are incorrectly predicted.Given the deficiencies of the above robots in environment modelling and scene perception,this thesis proposes the following improvements:(1)This thesis adopts the outlier detection algorithm and the super voxel segmentation algorithm for the problems of noise points and lack of pre-segmentation information in the point cloud map.A set of dense point cloud mapping system based on quad-rotor UAV was designed,and the indoor dense point cloud mapping was completed in the actual scene.The outlier detection algorithm eliminated the noise points in the map to a certain extent.And use the super voxel segmentation algorithm to aggregate each object's local features in the point cloud map to complete the map presegmentation.(2)Deep GCNs model does not fully use local features in the prediction part.This thesis proposes the Deep GCNs-Att model based on the dual attention network,which improves the prediction module of the original Deep GCNs network.This thesis uses S3 DIS as the data set.Compared with Deep GCNs,the proposed model has fewer Params and FLOPs.In addition,the experiment results show that when using the 6-fold cross training method on the same backbone network layer,the Mean Intersection over Union increases by about 2.16% and the Overall Accuracy is 84.7%.(3)This thesis proposes a curvature-based Markov random field optimization algorithm for misdetected points in semantic segmentation.First,the algorithm utilizes two different downsampling methods to filter the redundant information in semantic segmentation and uses Markov random field to optimize the semantic segmentation results.This thesis proves the effectiveness of Markov random field through experiments.Compared with the results of Deep GCNs-Att semantic segmentation,it improves the Mean Intersection over Union by 0.78% and the Overall Accuracy by0.36%.This thesis designs a set of dense point cloud mapping systems based on drones.It creates dense maps in actual indoor scenes,uses the outlier detection algorithm and the super voxel segmentation algorithm to preprocess the map,and proposes Deep GCNs-Att semantic segmentation model and curvature-based Markov random field optimization algorithm.Finally,the effectiveness and feasibility of the above two algorithms are verified on the public data set S3 DIS.