Efficient Robust Global Registration of 3D Data

点击翻译按钮获取中文摘要

Registration is the process of transforming two or more partially overlapping data sets to align***in the same coordinate reference frame. When the data are 3D point clouds, as acquired with range***sensors such as LiDAR or stereovision cameras, then the problem is one of surface registration,***which is an early processing step enabling a large variety of important emerging applications, such as autonomous driving and Simultaneous Localization and Mapping, automated scene reconstruction and object modelling, and 3D***object retrieval and recognition.******When an initial estimate of the transformation between data sets exists, then the problem is one***of refinement, known as local registration. Effective solutions to local registration have been known***for some time, prominently the Iterative Closest Point Algorithm and its many variants. More recently, the community has turned its attention to the more general and difficult problem of global registration, wherein no initial estimate of the transformation exists. Most approaches to global registration follow a probabilistic search and are heuristic. While an approach has recently been proposed, solutions that are both guaranteed and efficient remain elusive, especially when the region of overlap between the data sets is cluttered or occluded, or the degree of overlap is small.******The proposed research will build upon and significantly extend my previous work into global registration. My students and I will in the short-term extend our investigation of Virtual Interest Points to address global registration under non-rigid transformations. A second short-term goal will consider the impact of applying calculated boundaries to the transformations used to seed the local minima search process in Potential Well Space Embedding. A long term goal of the research is to develop more effective metrics and methods to evaluate the quality of a registration result. A second long term goal is to explore bridges between heuristic suboptimal and branch-and-bound optimal registration.******The advancement of these global registration techniques promise to enable an important set of applications, such as the use of realtime embedded range sensors for flexible indoor localization, the advent of which will be as transformative to indoor navigation as GPS has been in outdoor environments. Another related application is that of self-driving automobiles, which continually monitor their environments with range sensors, and for which effective global registration will enable enhanced navigation, recognition, and collision avoidance capabilities. The students in this research program will gain expertise in the general field of Computer Vision, specifically global registration. Their work will advance knowledge in this area, and the unique skills that they acquire will generate opportunities and job creation to the benefit of Canadian industry in this vital and exciting field.