Edge Extraction For 3D Point Clouds Based On Topology Awareness

In recent years,with the development of sensor technology and the rapid growth of the amount of point cloud data,3D point clouds have been widely used in many fields,such as autonomous driving,indoor navigation,AR/VR and urban modeling.In practical applications,we often extract the low-level features of point clouds to provide clues for high-level perception tasks.Particularly,the edge line structure,which is a kind of concise and expressive representation for 3D models,plays an important role in 3D modeling,registration,localization and other tasks.However,in the real world scenario,edge extraction for 3D point clouds is a challenging task.Firstly,the unstructured and irregular point cloud data are difficult to process directly;secondly,the scale of real scene point cloud data is often huge.Occlusion,noise and other factors can affect the quality of the point clouds;thirdly,it is difficult to define the edges of point clouds by explicit and formalized rules.At present,there is still much room for improvement in the research on edge extraction for 3D point clouds.In this thesis,we present an edge extraction deep learning network for point clouds,and we do further research from the perspective of topological constraints.Our ultimate goal is to design a topology-aware edge extraction algorithm for 3D point clouds based on deep learning.The main work and contributions of this thesis include:1.We present a parametric edge extraction network,3D-GMcGAN(3D-Guided Multi-conditional GAN),for point clouds based on GAN.By converting the operations in point space to parameter space and introducing initial guidance into the network,the performance of the network has been improved.This method avoids the complicated definition of point cloud edges and introduces high-level semantic information,which make the extracted point cloud edges more consistent with human perception.In addition,we build a large-scale dataset with edge annotations,which can be used as a benchmark for point cloud edge detection task.2.We design a topology loss function,which introduces the persistent homology theory in topology into the deep learning frameworks,to optimize the distribution of predictions from the perspective of topology.Further more,we design two general ways to simply embed the topology loss function into the existing point cloud segmentation networks,to improve the performance of the original networks.3.We present a topology-aware edge extraction algorithm for point clouds.The edge points are extracted through a topology-aware point cloud segmentation network,and then are fitted into lines through a line fitting algorithm with false alarm filtering.This robust method can detect edges well in the presence of noise.