A probabilistic spectral graph matching algorithm for robust correspondence between lunar surface images

The correspondence between key points is an important problem in lunar surface image processing, and further lays the foundation for the navigation of a rover and the terrain reconstruction of the lunar surface. However, the problem is still challenging due to the existence of large scale and rotation transformations, reflected view of the same scenery, and different illumination conditions between acquired images as the lunar rover moves forward. Traditional appearance matching algorithms, like SIFT, often fail in handling the above situations. By utilizing the structural cues between points, in this paper we propose a probabilistic spectral graph matching method to tackle the point correspondence problem in lunar surface images acquired by Yutu lunar rover which has been recently transmitted to the moon by China's Chang'e-3 lunar probe. Compared with traditional methods, the proposed method has three advantages. First, the incorporation of the structural information makes the matching more robust with respect to geometric transformations and illumination changes. Second, the assignment between points is interpreted in a probabilistic manner, and thus the best assignments can be easily figured out by ranking the probabilities. Third, the optimization problem can be efficiently approximately solved by spectral decomposition. Simulations on real lunar surface images witness the effectiveness of the proposed method.