In order to improve accuracy of large-scale scene model in three-dimensional (3D) reconstruction, we extract two kinds of partial stable invariant features, under the premise of ensuring the efficiency of the algorithm, and use a multi-feature fusion method to match image features. Considering both problems of the joint modeling based on aerial and urban street images, we propose a matching method based on the two kinds of partial stable features. The method comprises the following steps. Firstly, we extract ASIFT (Affine Scale Invariant Feature Transform) feature points and MSER feature areas, and improve the MSER (Maximally Stable Extremal Regions) algorithm to get the two stable features described by SIFT (Scale Invariant Feature Transform) feature descriptor; secondly, we use the homography matrix to match features by the feature matching algorithm; finally, we parallelly optimize feature matching by using graphics processing unit(GPU). A large number of experiments and comparison results show that more than twice correct matching pairs can be obtained by the proposed algorithm than other two algorithms.