At present, many factors restrict the application of the simultaneous localization and mapping ( SLAM) in real environment, and the interference of dynamic objects in indoor environment is one of the urgent problems to be solved. This article proposes a visual SLAM system based on ORB-SLAM3 and assisted by the instance segmentation network. The system places the segmentation task at the back end, and combines RGB-D camera input and KCF algorithm at the front end to detect semantic information from the back end. After tracking and transmitting, the system uses this information to track the motion state of key points in the Bayesian probability framework. Compared with current methods based on detection or segmentation, this system uses a lighter scheme to segment and track moving objects in the scene. With the further assistance of Bayesian filtering model, it not only realizes accurate dynamic interference filtering, but also optimizes the real-time problem of system operation caused by CNN network preprocessing. Experiments on the TUM RGB-D dataset show that the system can achieve high positioning accuracy at a speed of about 16 fps, with an average lead of 78. 56% compared with ORB-SLAM3 and 11. 85% compared with DynaSLAM.