A Novel Quaternion-Based 2D-3D Registration Algorithm with Line Correspondence

Image’s registration includes 2D-2D, 3D-3D and 3D-2D registration. This paper only concentrates on the 2D-3D registration, the image’s attitude is represented by a rotation matrix R, while the position is a translation vector T. Traditional approaches mainly focus on points correspondences, and state-of-the-art approaches concentrate on high-order structures, i.e. lines, rectangle, parallelepiped etc. Mathematically, most existing solutions adapt either linear optimization or iteration methods. However, they need the position, attitude initialization, which is not always available in real scene, and they do not guarantee to find global solutions. In this paper, instead of solving these polynomials directly, we introduce a novel approach (say qLR), which treat these multivariate polynomial equations as “monomials” and express R in a quaternion vision, resulting in dramatic decrement of the number of equations. We continue to utilize line correspondence since it commonly appears in real-scene and is easy to extract. Experiments on both simulation and real dataset attest qLR’s robustness and efficiency. Overall, our approach shares the following advantages: do not need any initialization and guarantee to find the global optimality if it exists; computational-cost is only linear to the number of measurements; robust to noise and separate the calculation of rotation from translation.