Visual SLAM Based on YOLOX-S in Dynamic Scenes

Abstract

Simultaneous Localization and Mapping (SLAM) is the foundation of various field such as moving robots, which has achieved good results in static environments. However, when dynamic factors enter the environment, wrong correlations will be generated, resulting in degraded accuracy and even losing tracking. In this paper, a SLAM system based on ORB-SLAM2 for dynamic environment is proposed. Based on RGB-D camera, the system uses YOLOX-S to detect dynamic objects and combines depth information to filter dynamic points. While reducing the influence of dynamic factors in the environment, the static information in the environment is retained as much as possible and the number of reliable feature points is increased. Finally, experiments are carried out on public datasets. The proposed system effectively improves the accuracy of SLAM in dynamic environment, and achieves good real-time performance due to the use of lightweight object detection network.