In order to solve the problems of RANSAC, such as high number of iterations, poor real-time performance and unstable robustness in the front end of simultaneous localization and mapping (SLAM), an improved image matching algorithm based on the fusion of quadtree method and PROSAC algorithm is proposed. The mismatching elimination algorithm of quadtree method + PROSAC algorithm is implemented, and the improved ORB-SLAM2 algorithm is tested on EuRoC data set. The results show that compared with ORB-SLAM2 system, the proposed algorithm reduces the average absolute trajectory error by 39.28% and the relative pose error by 35.45% on Vicon Room 1 03 dataset, and has higher mapping accuracy.