北京理工大学 增强现实智能医疗实验室

Abstract—Feature matching of endoscopic images is an important and challengeable task for many clinical applications, such as tissue surface reconstruction and object tracking. In this study, we proposed a locality preserving based motion consensus method for endoscopic image feature matching. Firstly, a local distance constraint is applied to maintain the local structure of initial matches derived from the ASIFT algorithm. Secondly, bilateral affine motion boundaries are estimated from the local structure preserving based matches to obtain precise motion constraint. Initial matches that meet the criterion of adaptive threshold of the bilateral affine motion boundaries are considered as final matches. Through considering both locality and global motion coherence of feature points, the proposed method can effectively find reliable matches from initial matches of large outlier ratios. We test our method and four state-of-the-art methods on simulated-nonrigid deformation and simulated-tool occlusion endoscopic images. The proposed method outperforms the other state-of-the-art methods in Precision, Recall, F1-Score, and Accuracy.