An efficient point-set registration algorithm with dual terms based on total least squares

Abstract

Point set registration (PSR) is competitive with related techniques because it purposefully captures the overall structure between two point-set patterns. Typically, the point set registration problem can be divided into two sub-problems: (1) search point set correspondence (PSC); (2) estimate spatial transformation matrix (STM). Searching for the best PSC is a classical combinatorial explosion problem, and estimating the STM is a continuous space optimization problem. Also, two feature point sets detected by two low-quality images include point-position errors and often involve bilateral outliers composed of such feature points that cannot form a correspondence relationship. To address the above problems, we propose an efficient PSR algorithm with dual (symmetrical) terms based on the total least squares (DT-TLS), which can correct errors-in-variables and suppress multiple outliers. Meanwhile, the framework of soft decision-making is presented, and a TLS-based criterion is constructed to efficiently exploit the probability and global structures of two-point sets. Such TLS-based criterion with single row-orthonormal STM includes two interesting dual (symmetrical) terms that can be conveniently exploited to suppress bilateral outliers. The experimental results show that DT-TLS achieves better performance than the state-of-the-art algorithms in some multi-view computer vision tasks, indicating that our proposed algorithm is suitable for solving PSR problems.