Aiming at the problems that traditional simultaneous localization and mapping (SLAM) based on static scene assumptions has poor robustness and low accuracy of position estimation in dynamic scenes, and that the ORB-SLAM2 system can only construct sparse point cloud maps but not dense point cloud maps. Adding and comparing two YOLO target detection networks to effectively detect a priori dynamic information in a tracking thread. First, YOLO dynamic target detection network and dynamic feature rejection module are added to the tracking thread to improve the localization accuracy of the system; second, the dense map building thread is added to complete the construction of dense point cloud maps based on the key frames with the corresponding bit poses. Experiments on the publicly available TUM RGB-D datasets fr3/s_static, fr3/w_static and fr3/w_xyz show that compared with the original ORB-SLAM2 algorithm, the system localization accuracy is improved by more than 98% at most in a high dynamic environment, and more than 13% at most in a low dynamic environment after combining with YOLOv5m. After combining with YOLOv7, the positioning accuracy of the system is improved by more than 98% in high dynamic environments and more than 12% in low dynamic environments, which effectively improves the positioning accuracy and robustness of the system in dynamic environments. The 3D dense point cloud map is also constructed, which lays the foundation for subsequent applications in robot autonomous navigation, obstacle avoidance and path planning.