Real-Time RGB-D SLAM System with Semantic and Depth Information Fusion in Dynamic Environment

Abstract

Simultaneous localization and mapping (SLAM) is crucial for intelligent mobile robots to move autonomously in unknown environments. However, many current visual SLAM systems heavily rely on static scene assumptions, which severely limit their applicability in dynamic environments. Additionally, many SLAM systems based on semantic segmentation are unable to operate in real time, rendering them impractical for robotic applications. In this paper, we propose a real-time semantic RGB-D visualization SLAM system within the ORB-SLAM3 framework. Firstly, we add two modules, a real-time semantic information acquisition module and a fast dynamic feature point removal module that integrates semantic information and depth information. Subsequently, the dynamic feature point removal module is incorporated into the tracking thread of ORB-SLAM3. We conduct experimental evaluations on the TUM dataset and real environments. The results demonstrate that our algorithm is one of the most accurate robust and real-time systems in dynamic scenes.