In the process of visual SLAM(Simultaneous Localization and Mapping), the image feature matching algorithm of RANSAC(Random Sampling Consensus) is usually used to estimate the image model randomly, which is easy to cause the uncertainty of algorithm time complexity, and then lead to excessive image matching time consumption. It is difficult to meet the real-time requirements of visual SLAM. In order to improve this problem, this paper uses the algorithm of PROSAC (Progressive Sampling Consensus) to screen image features and reject mismatched feature points, which effectively improves the efficiency and robustness of image feature matching, and further enhances the stability and real-time performance of visual SLAM. Experimental verification shows that compared with the current visual SLAM feature matching algorithm, the computational efficiency of this paper is significantly improved.