DFS-SLAM: A Visual SLAM Algorithm for Deep Fusion of Semantic Information

Abstract

In order to solve issues such as low accuracy, poor robustness and tracking loss in dynamic environments in traditional Simultaneous Localization and Mapping (SLAM) techniques, a semantic SLAM system named DFS-SLAM (Deep Fusion Semantics SLAM) is proposed. The algorithmic framework of DFS-SLAM is heavily inspired by ORB-SLAM2 and is robust in dynamic environments. In this letter, The DFS-SLAM proposed in this letter deeply integrates the semantic segmentation network and geometric method to effectively reduce the impact of dynamic objects on map construction. Meanwhile, a static point optimization method is proposed to improve the camera posture by using the semantic information of static objects, which effectively compensates for the loss of features after eliminating dynamic objects. The validation results on the TUM dataset show that DFS-SLAM has high accuracy and robustness in dynamic environments. In highly dynamic environments, DFS-SLAM shows a significant improvement in mapping accuracy compared to the ORB-SLAM2 algorithm. In most scenarios, it outperforms similar semantic SLAM algorithms such as DS-SLAM and DynaSLAM in accuracy as well.