A Robust Feature Matching Method based on Semantic Geometric Constraints

Feature matching is crucial in computer vision applications, which aims to determine the correspondence between images. Different from traditional methods, this paper proposes a robust feature matching method based on semantic geometric constraints, which adopts the idea of two-stage matching. First, the image semantic information is used as a priori, and the region matching association is performed through the 3D-2D dual constraint method. On this basis, a matching trust factor is introduced, and a semantic geometric consistency feature matcher is constructed in combination with the VGG feature descriptor network to achieve robust feature matching. Experimental results show that this method has better matching performance for scenes with viewpoint changes and motion blur, and the AUC accuracy of relative pose estimation for indoor and outdoor scenes has been improved.