SAFD: A Statistical Analysis-Based Feature Point Descriptor

Describing and matching feature points is one of the most basic tasks in many computer vision tasks. For example, in tasks such as structure from motion, simultaneous localization and mapping, and image stitching, the quality of feature point matching directly determines the accuracy of the vision task, which in turn affects the availability of specific applications such as VR, MR, autonomous driving, and robot navigation. The current mainstream feature point method has the problem that the number of mismatches is relatively large, which will cause a lot of interference to computer vision tasks based on feature point matching. Traditional feature point descriptors only use the feature point and very little information in its neighborhood, so mismatches are prone to occur. The main contribution of this paper is to propose a feature point descriptor and a corresponding matching method. The core idea is to use multiple dimensions of information to describe a feature point, and then integrate the differences in each dimension to evaluate the similarity between the two feature points, and then improve the matching rate of feature points. The experimental results show that the feature description and matching algorithm proposed in this paper is superior to the popular ORB algorithm in terms of matching rate.