Real-time visual-inertial SLAM based on RGB-D image and point-line feature with loop closure constraint

Conventional visual-inertial SLAM(VINS) primarily relying on point features often face challenges in indoor environments characterized by low textures, variable lighting, and fast motions. This is attributed to the difficulty in identifying a substantial number of dependable point features in such scenarios. Considering the abundance of structured features, such as line features in artificial environments, the RPL-VINS is introduced. This system not only harnesses line features to impose additional constraints but also seamlessly integrates them within the loop closure model. Using the Dempster-Shafer evidence theory, the system achieves a tight fusion of point-line loop closure outcomes, dramatically enhancing localization precision. Furthermore, the RGB-D camera's depth information facilitate direct recovery of feature-based 3D information, refining trajectory accuracy. By assimilating point-line features, the RPL-VINS's visual-inertial odometry (VIO) effectively trajectory and localizes the pose of each image frame. Experimental results on public OpenLORIS-Scene dataset reveal that RPL-VINS outperforms VINS-RGBD, PL-VIO, and PL-VINS in terms of performance. While ensuring real-time, it notably enhances trajectory precision and robustness, addressing the performance shortcomings of traditional algorithms in low-texture environments.