Recognition And Segmentation Of 3D Scanning Point Cloud Of Ship Segmental Components Based On PointSIFT

Ship measurement is an important technology in the process of ship construction,and it is the key technology to ensure the efficiency of ship construction.With the development of large-scale ships and Intelligent manufacturing,on one hand,shipyards require higher technology in measurement methods;on the other hand,with the using of three-dimensional laser scanning,a measurement technology which can measure large areas quickly with high precision,the application of this technology in ship measurement will certainly become the trend of future development.At present,the measurement methods used by shipyards have serval disadvantages of low efficiency,while the application of three-dimensional scanning technology in ship construction is still in its infancy.Compared with the traditional measurement method,the result of three-dimensional laser scanning measurement is composed with a large number of three-dimensional coordinates of the measured object surface points.In the past,the processing method of measurement results for key points in hull segmentation is not suitable for three-dimensional laser scanning data,but the point cloud segmentation algorithm should be used to segment the measurement data first before using subsequent error detection methods to detect.At present,there is no point cloud segmentation algorithm for ship hull point clouds segmentation.So in this paper a point cloud segmentation algorithm for hull segmented point clouds is used based on PointSIFT deep learning model,which can solves the point cloud segmentation problem for hull segmented point clouds.The technical route of this paper is as follows:Firstly,building a dataset of point cloud segmentation of ship hull segmentation.The dataset is derived from the CAD model files collected from production design using by shipyards in construction.It covers a variety of representative structural styles such as bottom section structure,side section structure,deck section structure,bulkhead structure and so on.The collected model files are read,sampled and annotated by software.In the process of labeling,the category of each point in the point cloud will be labeled,and the label results will be automatically saved.Finally,700 different hull segment point clouds are generated as training sets and 200 as training sets by means of data enhancement.Then,research on the structure of PointSIFT deep learning model and the model structure is adjusted according to the characteristics of hull segment point cloud to achieve higher segmentation accuracy.In order to make the model to have a certain generalization ability,even if the model trained by the point cloud generated by CAD model can be used in the segmentation of the actual hull segment point cloud,the network structure of PointSIFT is improved.In order to make the model extract the feature of complex hull components,the number of input points in the input layer is adjusted to 16384 in this model,and the number of parameters in each layer is adjusted to 16384.The number of parameters in the output layer is less than the original model,which prevents over-fitting during training and increases the generalization ability of the model.Finally,the accuracy of the training model is 79% in the testset,and the prediction accuracy of the strong component is 89% and the recall rate is 75%.The accuracy of predicting common components is 67%,and the recall rate is 83%.The results prove the availability and feasibility of the algorithm.The algorithm used in this paper can effectively segment point clouds of ship hull segments.