Local Feature Description And Matching For Point Clouds

Establishing reasonable point-to-pint correspondences between 3D point clouds is a fundamental yet critical issue in 3D computer vision and pattern recognition,it is a research hotspot and has been applied to many areas such as 3D reconstruction,object recognition and tracking,localization and mapping,and grasping.Such correspondences usually rely on local feature description and matching.The objective of local geometric feature description is to represent the geometric information contained in the local surface with a 1D feature vector;feature matching is supposed to find those consistent correspondences between point clouds using extracted local features.However,measuring noise,varying data resolutions,clutter,and occlusion possess great challenges to this task.With these regards,this dissertation provides the following contributions that effectively address these concerns.First,a TOLDI descriptor is proposed that is highly distinctive and robust to noise,varying data resolutions,clutter,and occlusion.We first propose a new local reference frame for the local surface that leverages normals and weighted projection vectors for axes calculation to achieve rotation invariance and full spatial information encoding.Our local reference frame is also highly repeatable and general for other existing 3D local feature descriptors.Then,we introduce a new feature representation that performs 3D-to-2D projection orthogonally in the local reference frame,resulting in three depth images that is further concatenated into a 1D vector.Benefited from the strong repeatability of the proposed local reference frame and high discriminative power of our suggested feature representation,our TOLDI descriptor simultaneously achieves superior feature matching performance on three datasets that address shape retrieval,point cloud registration,and 3D object recognition scenarios,respectively.Second,an RCS descriptor and its binary variants are proposed to achieve light-weight representations.RCS solves the high dimensionality problem of TOLDI while still being distinctive and robust.Key to the RCS descriptor are the multi-view mechanism and contour signature representation.We demonstrate that contour is a very effective yet compact cue for feature representation.In addition,we propose three approaches to binarize RCS,i.e.,thresholding,quantization,and geometrical binary encoding.The resultant RCS binary versions lose very few distinctiveness while being ultra-fast for matching and storage.Third,a deep learning-based approach is proposed for both low-level and high-level feature fusion.We observed that current fusion methods,either for descriptor generation or enhanced matching,are typically based on linear operations such as concatenation and min pooling.However,they will lead to redundant features and less comprehensive utilization of complementary information provided by different features.We present the first deep learning-based approach to fusing local geometric features non-linearly.Specifically,we show that more compact and distinctive representations can be achieved by optimizing a neural network model under the triplet framework.A new loss is proposed as well that fully leverages all pairwise relationships within the triplet.The experiments demonstrate the advantages of our approach in terms of both feature matching and geometric registration.Forth,a consistency voting-based feature matching method is proposed as a robust matcher regarding high outlier ratios.Since false matches are always distributed unorderly,while correct matches have strong consistency.Thus,we leverage such property and design a votingbased mechanism to judge the correctness of a correspondence,being able to find correct correspondences within a set contaminated by severe outliers.Our method is efficient and superior to conventional feature similarity-based matchers.Fifth,we apply our proposed methods to a real-world application,i.e.,the pose estimation and reconstruction of non-cooperative targets.For synthetic and scanned model data,our method manages to compute the pose information including angular velocity and principal axis when the object undergoes rotation,translation,noise,and scanning distance variation.With such pose information and scanned point cloud sequences,we can reconstruct the object accurately.Above techniques are finally embedded in an applicable software.We have proposed several general local feature description and matching methods and solved a real-world application problem.Our methods are heuristic for the research community and significant in terms of application prospects for the industry community.