Abstract: The feature points of moving objects in the actual environment are added to the calculation of camera pose, and the feature points in the static environment are over-sparse, which will lead to the low accuracy and poor robustness of the traditional visual simultaneous localization and mapping (SLAM) algorithm of mobile robot in pose estimation. A bilateral semantic segmentation algorithm based on branching hole convolution was designed to divide the environment into potential motion region and static region. Combined with geometric constraints to perform secondary judgments of static feature points and judgments of mobile feature points without prior dynamic markings, the dynamic feature points were removed from all the feature points uniformly extracted in advance, and only the static feature points were used to solve the camera pose and construct the static environment map. The experiment on the TUM common dataset shows that the positioning accuracy of SLAM in dynamic environment of the proposed algorithm is obviously better than that of other existing methods. The mapping experiment was carried out in the real environment with moving objects, and compared with the ORB-SLAM2 algorithm, the map constructed in the dynamic environment was clearer.