ORB-SLAM2 is prone to scale quantization error due to discrete Gaussian pyramid and quantization scale. Aiming at this defect, a twin filter algorithm is proposed. The quantization error is reduced by constructing descriptors in the logarithmic polar coordinate system, and twin descriptors in the Cartesian coordinate system are constructed in the same pyramid, and the distance of the descriptors is used as the filter to improve the scale invariance of feature points and enhance their matching accuracy. At the same time, aiming at the situation that the qurdtree algorithm of ORB-SLAM2 excessively pursues the discreteness but neglects the feature point quality problem, this paper proposes the quadtree algorithm with limited depth. The feature point extraction threshold and the pyramid layer where the feature points are located are used to set the adaptive depth threshold to reduce the division times of weak feature point regions and thus reduce the number of weak feature points extraction. Experiments show that the proposed algorithm can effectively improve the dispersion of feature points, the number of correctly matched feature points and the matching accuracy, and has higher trajectory accuracy.