At present, in the simultaneous localization and map construction (SLAM) algorithm, the more mature technology is the map construction in the static environment. However, there are currently moving objects in most environments, which will affect the accuracy and integrity of slam mapping. In this paper, we propose an improved vision SLAM based on ORB-SLAM2 that adds an independent parallel target detection thread to effectively filter low dynamic objects in the scene. Finally, we conduct an experimental evaluation using the TUM dataset, and the results show that our algorithm has a great improvement over ORB-SLAM2 in positioning accuracy and running time.