Deep Learning-based 3D Pipelines Reconstruction

Pipelines are the basic building blocks of many industrial sites,such as electricity,oil and chemical sites.They are used for the short and long distance transmission of liquid or gas.With the development of economy and optimization of industrial structure,a large number of industrial sites are facing industrial upgrading.Pipeline reconstruction became an important means of understanding the current situation of the site.In addition,with the rapid development of the internet and the popularization of 3d digital world,factory digitization plays an important role in enterprise management High quality 3D models of power-plants,petrochemical plants and other industrial sites are crucial in many applications,including disaster simulations,monitoring and executive training.Today,modern laser scanners can capture three-dimensional surfaces and geometric shapes with high precision generating dense samples of point clouds so that high quality point cloud becomes the popular representation of 3D scene.Thus it becomes an important problem to reconstruct high quality 3D pipeline model from point cloud.Although pipes are merely cylindrical primitives which can be easily defined by their axis and radius,they often consist additional components such as flanges,valves,inlets,elbows,tees and etc.3D pipes are typically dense,consisting a wide range of topologies and geometries,with large self-occlusions.Thus,reconstruction of a coherent 3D pipe models from large-scale point clouds is a challenging problem.As the basic shape of the pipeline,obtaining the cylinder from the point cloud is the key to the pipeline reconstruction.The existing works adopt the geometric method to obtain the radius and axis position of the cylinder through fitting three-dimensional cylinder or two-dimensional circle,and then detecting the possible connection relationship to complete the reconstruction of the whole pipeline.This kind of work often has the local problem,the result heavily depends on the threshold,and the connection relationship need to be predicted through established ruleIn this work we take a prior-based reconstruction approach in which we learn recognition of parts in the scene.This reduces the complexity of the general pipe reconstruction problem into a combination of part detection and model fitting problems We utilize a convolutional network to learn 3D point cloud features and the classification into various classes.The pipe classification is noisy and we apply robust clustering and graph-based aggregation techniques to compute a coherent pipe model Our method shows promising results on pipe models with varying complexity and density both in synthetic and real cases.The main contributions of this article are as follows:1.An automatic prior-based approach that reduces the complexity of the general pipe reconstruction problem into a combination of part detection and model fitting problems is proposed.2.The proposed approach combines neural network and geometry,and makes use of their advantages to solve the reconstruction problem well,which reduces the time complexity and has promising results.3.The proposed approach can recover more details such like flanges,tees and etc.