Real-time Direct Monocular SLAM with Learning-based Confidence Estimation

Direct monocular simultaneous localization and mapping (SLAM) methods, for which the image intensity is used for tracking and mapping instead of sparse feature points, have gained in popularity in recent years. However, feature-based methods usually have more accurate camera localization results than most direct methods, though direct methods can work better in a textureless environment. To tackle the localization issue, we develop a novel real-time large-scale direct SLAM model, namely, GCP-SLAM, by integrating the learning-based confidence estimation into the depth fusion and motion tracking optimization. In GCP-SLAM, a random regression forest is trained off-line with pre-defined confidence measures for learning confidence and detecting the ground control points (GCPs). Then, the confidence value along with the selected GCPs is utilized for depth refinement and camera localization. Our proposed method is shown experimentally more reliable in tracking and relocalization than the previous state-of-the-art direct method when compared with feature-based and RGBD SLAMs.