Summary: The most important part in image feature matching is to retrieve feature vectors via a distance function. This paper focuses on better extracting feature points and establishing points’ neighborhoods more quickly and accurately. First, the convex hulls of speeded up robust feature (SURF) feature points are divided into Delaunay triangles. Then, the indexes of the Delaunay connections are built by sampling, clustering and quantization. Finally, we construct a matching grid of the pairwise points by a voting algorithm, which improves matching efficiency without using any relevant structural information. The paper proposes a novel matching method based on SURF feature and Delaunay triangular meshes. Experiment results verify that, the method is able to extract more feature points and achieve feature matching with a higher accuracy while maintaining time cost.