Resource-aware structural point-line RGB-D SLAM for dynamic indoor environments
TL;DR
A temporally maintained object-level mask and optical-flow consistency are used to filter unreliable observations, while 3D line landmarks provide complementary structural constraints from the static environment for robust localization in indoor scenes with moving objects, without relying on GPU acceleration.
Abstract
This paper presents a CPU-oriented dynamic point-line RGB-D SLAM method for robust localization in indoor scenes with moving objects, without relying on GPU acceleration. The method addresses two coupled challenges in dynamic RGB-D SLAM: corrupted visual constraints from moving foreground objects and weakened geometric observability after dynamic feature suppression. To this end, a temporally maintained object-level mask and optical-flow consistency are used to filter unreliable observations, while 3D line landmarks provide complementary structural constraints from the static environment. The backend jointly optimizes camera poses, point landmarks, and line landmarks in local and global bundle adjustment. For line optimization, we adopt a point-to-line reprojection residual that constrains observed 2D segment endpoints to lie on the projected image line of a 3D line landmark. This avoids direct endpoint-to-endpoint matching and improves robustness to changes in visible line extent caused by occlusion, viewpoint variation, and line fragmentation. After loop closing or map merging, duplicate lines are conservatively fused using geometric and descriptor consistency. Experiments on public dynamic RGB-D datasets demonstrate competitive trajectory accuracy, particularly in scenes with observable static structure, while revealing limitations under strong rotational motion.