SIG-SLAM: Semantic Information-Guided Real-Time SLAM for Dynamic Scenes

Visual Simultaneous Localization and Mapping (Visual-SLAM) providing the camera's location and the map of surroundings, becomes more and more popular in the field of AR/VR, robotics, unmanned vehicle driving and so on. Traditional SLAM assumes that the scene is static, which results in the bad performance in dynamic scenes. In this paper we present SIG-SLAM, a real-time dynamic SLAM system that is built on ORB-SLAM2, adding the capability of dynamic object pixel-level semantic segmentation. SIG-SLAM can still perform well in dynamic scenes via taking the semantic information into account. Besides, by adding another independent thread for semantic segmentation, tracking thread does not need to stop to wait for semantic information any more. Finally, a novel approach to utilize the semantic information efficiently is proposed, making sure the real-time performance of our system. We evaluate our system in the public dataset TUM. The result shows that compared to other state-of-the-art algorithms, our proposed method can provide competitive accuracy while ensuring good real-time performance.