Lifelong SLAM Based on Environmental Semantic and Depth Information in Dynamic Scenes

The Simultaneous Localization and Mapping (SLAM) algorithm is robust enough when used on static scenes. However, it is less robust when used in dynamic scenes where objects move for an extended period in indoor scenes. In order to address the mentioned issues, this paper proposes an indoor dynamic scene localization and map construction algorithm that combines environmental semantic with depth information. Based on the ORB-SLAM2 framework: First, we perform semantic segmentation on keyframes to obtain the mask, then we combine the mask and optical flow algorithm to estimate the motion properties of environmental objects and remove dynamic features. Second, keyframe pose, semantic label, and depth information are combined to identify scene loop. Finally, we filter the keyframes twice and construct a static semantic map. We conducted experiments in a dynamic simulation environment and TUM dataset. The result shows that the proposed algorithm can maintain high localization and loop detection accuracy in dynamic scenes, and the static semantic map can reliably describe the scene, reflecting better lifelong localization robustness.