Dynamic SLAM Method Based on Cavity Complementation

Visual simultaneous localization and mapping(SLAM) is a core technology in the fields of intelligent robots and unmanned driving. Usually, most visual SLAM focused on the static scene cannot be easily applied in dynamic scenes, and there is no semantic information in the map. Some visual SLAM methods have been applied in dynamic scenes. They use neural networks to eliminate dynamic objects so as to reduce the dynamic interference, but the void left in the deleted image will have a great impact on pose estimation and map construction. Therefore, we propose a dynamic environment based on the cavity completion semantic SLAM method. First, the semantic segmentation network and motion consistency detection are combined to eliminate dynamic objects.Second, the eliminated image sequences with holes are inputted into the complete network for deep completion. Finally, the completed image sequence is used for loop detection and three-dimensional scene reconstruction. Experiments are conducted on the KITTI and TUM datasets. The results show that the accuracy of the proposed pose estimation method achieved better results compared with ORB-SLAM2 and the current advanced SLAM methods.