A Novel Depth Map-Assisted Monocular Visual–Inertial Odometry

Due to the complementarity between light detection and ranging (LiDAR) and images, fusing the LiDAR data can significantly enhance the accuracy and robustness of visual simultaneous localization and mapping (SLAM) systems. Cameras are smaller and less expensive than LiDAR, making them easier to deploy. Building on these properties, we propose a novel depth map-assisted monocular visual–inertial odometry (DM-VIO) that fuses image and LiDAR data through the reuse of prior LiDAR depth maps. Existing monocular visual SLAM methods often suffer from inaccurate depth estimation, while many LiDAR-based depth estimation methods fail to fully leverage depth variation. To address this, we propose a depth estimation method based on the radial basis function (RBF) interpolation. Local point clouds, extracted from the prior depth map near visual features, are projected onto two principal directions using the principal component analysis (PCA). The compressed point clouds are then used to construct interpolation functions to estimate the depth of visual features. In addition, we introduce a visual–LiDAR fusion constraint method. This method uses semantic information from images to track associations among LiDAR point clouds projected onto images. Experimental results demonstrate that the proposed method significantly improves the accuracy and robustness of monocular visual–inertial odometry (VIO), especially in challenging environments, such as low-light and large-scale outdoor scenarios.