RSD-SLAM: A Robust Saliency-Driven Visual SLAM System in Indoor Environments

The region of interest (ROI) with abundant and structured textures provides robust features in an indoor environment, which can effectively facilitate accurate simultaneous localization and mapping (SLAM). However, most existing visual SLAM systems generally treat ROI and non-ROI uniformly, resulting in ineffective employment of ROI. To meet this gap, we propose a robust saliency-driven visual SLAM system for indoor environments, coined RSD-SLAM. It can increase the focus on valuable ROI with the saliency maps obtained from a novel saliency prediction (SP) model. Specifically, we first design a saliency map construction method for visual SLAM, enabling the SP model to accurately describe ROI, which generates the first indoor SP dataset integrating geometric, semantic, depth, and low-level visual information. Second, we develop a global stability constraint module for the SP model to enable the capability of keeping temporal consistency and illumination invariance. Third, we design a saliency map-based hybrid saliency-driven mechanism to increase the focus of the system on ROI. At the front end of the system, an adaptive feature-point extraction algorithm extracts more robust feature-points from the ROI, and a saliency entropy-based keyframe selection algorithm selects keyframes with the saliency value distribution of feature points. At the back end, a dynamic weighted bundle adjustment (BA) optimization algorithm heavily weights the map points of the ROI. Last, the particular focus on ROI results in a robust and accurate location. Extensive experiments, conducted on the EuRoC and TUM RGB-D datasets as well as in simulation environments, demonstrate that the proposed RSD-SLAM significantly outperforms the state-of-the-art in robustness and accuracy.