Scale aware dense dynamic SLAM for monocular, stereo and RGBD cameras

Accurate real-time dense mapping with precise scale and localization is crucial for autonomous robot navigation, particularly in dynamic environments. However, existing methods often rely on a single sensor type and lack robustness to dynamic scenes. To address these challenges, a generalized framework for real-time, scale-aware dense mapping with dynamic robustness is proposed in this paper, referred to as SDMFusion. SDMFusion supports monocular, stereo, and RGB-D cameras and is built upon the ORB-SLAM3 system. Three core modules are integrated into SDMFusion. The scale-depth optimization module recovers the absolute scale for monocular and refines the depth maps. The dynamic feature rejection module segments dynamic objects, combining geometric constraints and moving consistency checks to facilitate dynamic feature rejection. The real-time anti-dynamic reconstruction module generates high-quality dense maps of static regions using optimized depth, dynamic masks, and camera poses. Extensive experiments on KITTI, TUM RGB-D, BONN RGB-D, and real-world datasets validate the effectiveness of our approach. The results demonstrate that SDMFusion achieves superior overall performance in accuracy and robustness compared to ORB-SLAM3 and other advanced dynamic SLAM methods. Furthermore, our method effectively eliminates dynamic regions from the dense maps.