SGC-VSLAM: A Semantic and Geometric Constraints VSLAM for Dynamic Indoor Environments

As one of the core technologies for autonomous mobile robots, Visual Simultaneous Localization and Mapping (VSLAM) has been widely researched in recent years. However, most state-of-the-art VSLAM adopts a strong scene rigidity assumption for analytical convenience, which limits the utility of these algorithms for real-world environments with independent dynamic objects. Hence, this paper presents a semantic and geometric constraints VSLAM (SGC-VSLAM), which is built on the RGB-D mode of ORB-SLAM2 with the addition of dynamic detection and static point cloud map construction modules. In detail, a novel improved quadtree-based method was adopted for SGC-VSLAM to enhance the performance of the feature extractor in ORB-SLAM (Oriented FAST and Rotated BRIEF-SLAM). Moreover, a new dynamic feature detection method called semantic and geometric constraints was proposed, which provided a robust and fast way to filter dynamic features. The semantic bounding box generated by YOLO v3 (You Only Look Once, v3) was used to calculate a more accurate fundamental matrix between adjacent frames, which was then used to filter all of the truly dynamic features. Finally, a static point cloud was estimated by using a new drawing key frame selection strategy. Experiments on the public TUM RGB-D (Red-Green-Blue Depth) dataset were conducted to evaluate the proposed approach. This evaluation revealed that the proposed SGC-VSLAM can effectively improve the positioning accuracy of the ORB-SLAM2 system in high-dynamic scenarios and was also able to build a map with the static parts of the real environment, which has long-term application value for autonomous mobile robots.