Proceedings of the
The Nineteenth International Conference on Computational Intelligence and Security (CIS 2023)
December 1 – 4, 2023, Haikou, China
Research on an Improved Real Time Semantic RGB-D SLAM System for Dynamic Environments
School of Air Transport, Shanghai University of Engineering Science, China.
ABSTRACT
Currently, most visual SLAMs rely on static assumptions, making them prone to failure in dynamic environments. In recent years, some have used semantic information to address the impact of dynamic objects in visual SLAM. However, such methods cannot handle dynamic objects without prior information and rely on high computational costs, making the system unable to run efficiently. To address such issues, we propose a real-time semantic RGBDSLAM system suitable for dynamic environments. The system is based on ORB-SLAM3 and adds semantic segmentation threads. In order to reduce computational costs, semantic segmentation is only performed on key frames and semantic information is updated, and the tracking thread is designed as a non blocking structure. To handle unknown dynamic objects, we propose a geometric module using clustering algorithms, which combines the system with ORB using the Technical University of Munich (TUM) dataset. The comparison with SLAM3 shows that the system significantly reduces trajectory error.
Keywords: SLAM, SegNet, ORB-SLAM3, DBSCAN, Semantic.
Download PDF
School of Air Transport, Shanghai University of Engineering Science, China.
ABSTRACT
Currently, most visual SLAMs rely on static assumptions, making them prone to failure in dynamic environments. In recent years, some have used semantic information to address the impact of dynamic objects in visual SLAM. However, such methods cannot handle dynamic objects without prior information and rely on high computational costs, making the system unable to run efficiently. To address such issues, we propose a real-time semantic RGBDSLAM system suitable for dynamic environments. The system is based on ORB-SLAM3 and adds semantic segmentation threads. In order to reduce computational costs, semantic segmentation is only performed on key frames and semantic information is updated, and the tracking thread is designed as a non blocking structure. To handle unknown dynamic objects, we propose a geometric module using clustering algorithms, which combines the system with ORB using the Technical University of Munich (TUM) dataset. The comparison with SLAM3 shows that the system significantly reduces trajectory error.
Keywords: SLAM, SegNet, ORB-SLAM3, DBSCAN, Semantic.
Download PDF