DEP-SLAM: A Dynamic Environment Perception SLAM System with Large Language Models

Simultaneous localization and mapping (SLAM) is crucial for robot navigation and building environment maps in real time. Most visual SLAM (VSLAM) systems assume that objects in the environment are static. However, in highly dynamic environments where this assumption fails, the system’s efficiency can be seriously affected. Furthermore, current VSLAM systems are unable to interact with the environment or adapt their operating strategies to environmental changes. In this paper, we introduce DEP-SLAM, a dynamic environment perception SLAM system combined with large language models. The lightweight object detection network YOLOV7-Tiny is used to obtain semantic information. Large language models are used to dynamically perceive changes in the environment. To minimize manual hyper-parameter tuning, we propose an enhanced geometric constraint method to better filter out and eliminate dynamic feature points. Experimental results demonstrate the superiority of DEP-SLAM over existing schemes, especially in terms of accuracy and robustness in highly dynamic indoor environments.