DMSG-SLAM: Cascaded Semantic and Geometric Filtering for RGB-D Tracking and Mapping in Dynamic Environments.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Traditional visual SLAM struggles with localization drift in dynamic environments due to moving objects.
• Existing semantic segmentation and depth-based masking methods have limitations, including under-segmentation and missed detections of truncated objects.

Purpose of the Study:

• To propose DMSG-SLAM, a cascaded visual SLAM framework designed to improve localization accuracy in dynamic environments.
• To fuse depth-mask, semantic information, and geometric constraints for robust dynamic object handling.

Main Methods:

• A lightweight object detection network with depth consistency generates initial masks for dynamic feature removal.
• A rotation-aware local epipolar geometric filtering mechanism suppresses residual features near object boundaries.
• Adaptive epipolar thresholding and TSDF-based dense volumetric mapping enhance surface reconstruction and filtering under challenging motion.

Main Results:

• DMSG-SLAM effectively removes dynamic features and mitigates perceptual blind spots caused by occlusion or truncation.
• The system demonstrates competitive accuracy in highly dynamic environments, outperforming ORB-SLAM2.
• Localization performance showed improvements of up to 90% on the TUM RGB-D dataset.

Conclusions:

• DMSG-SLAM offers a robust solution for visual SLAM in dynamic environments by effectively integrating multiple information sources.
• The proposed methods significantly enhance localization accuracy and map consistency compared to existing approaches.