Research On Segmentation Of 3D Scanning Point Cloud Of Hull Block In Storage Yard

The accuracy detection technology of the hull block occupies a very important position in the ship construction process.The total station measuring tool currently used by the shipyard is time-consuming,labor-intensive and inefficient.In recent years,3D scanning technology has emerged,which has the advantages of fast measurement speed and high measurement accuracy.The introduction of 3D scanning technology in block accuracy detection can overcome the problems of low surveying efficiency and scarce measuring points.However,when 3D scanning technology is used to scan the block,it will cover the entire storage yard background,and useful block point clouds need to be segmented and extracted.Therefore this thesis focus on the segmentation problem of hull block 3D scanning point cloud in storage yard,and proposes a segmentation method based on deep learning.This method first preprocesses the scanned point cloud of the yard scene;then uses the point cloud segmentation network trained in this thesis to segment the preprocessed point cloud;finally achieves the segmentation and extraction of hull block points from the scene.The research work includes the following aspects:1.Aiming at the problem that there is no public hull block point cloud data set in the ship field,a method for making ship block point cloud data sets combined with preprocessing operations is proposed.Firstly,the collected original scene point cloud is preprocessed,and the number of original point clouds is downsampled to a level that can be processed by deep learning,and the original geometric structure characteristics are greatly preserved;secondly,the data set is enhanced by means of translation,rotation,scene splicing,etc.,which solves the problem of sparse hull block point cloud data samples;finally,all samples are standardized and made into HDF5 format data sets,which can be used for subsequent deep learning;2.Aiming at the problem that there is no segmentation network that can be directly used to extract the hull block point cloud from the storage yard scene in current deep learning field,the network structure of the existing Point Net and DGCNN networks is improved.A weighted cross-entropy loss function is proposed to increase the loss weight of the network predicting block points as background points.The improved model has achieved 88%segmentation accuracy on the test set,and through the visualization of the segmentation results,the network trained in this thesis is also robust against the incomplete block point cloud problem.3.Based on the trained segmentation network model,this thesis develop point cloud segmentation software with graphical user interface.The software can easily and quickly realize the segmentation of the block point cloud.After inputting a scene point cloud,it only takes 3-5 seconds to segment the block point cloud,and the segmentation result can be displayed visually.