The invention relates to a semantic SLAM robustness improvement method based on instance segmentation, and the method comprises the steps: firstly carrying out the instance segmentation of a key framethrough an instance segmentation network, and building prior semantic information; calculating a feature point optical flow field to further distinguish the object, identifying a real moving object in the scene, and removing the feature points belonging to the dynamic object; and finally, performing semantic association, and establishing a semantic map without dynamic object interference. Compared with the prior art, the semantic map is established by adopting a method of combining deep learning and optical flow, and the depth map is added on the basis of the color map, so that the system isendowed with the capability of establishing the dense three-dimensional point cloud semantic map. In addition, a Mask-RCNN framework is adopted for real-time semantic segmentation, and object dynamicinformation can be calculated through mutual combination of dynamic feature points estimated by optical flow information and pixel-level semantic information. According to the method, deep learning and optical flow are mutually combined, so that the robustness of the whole system is remarkably improved, and the method can be applied to real-time semantic map construction in a dynamic scene.