Research on Multi-Sensor Fusion SLAM Algorithm Based on Outlier Rejection and Boundary Rendering

R3LIVE is an innovative laser-visual-inertial SLAM fusion framework. This framework leverages the error-state iterative Kalman filtering method to tightly couple Laser Inertial Odometry (LIO) and Visual Inertial Odometry (VIO) for mutual state optimization. It allows the construction of geometric structures under LIO and texture rendering under VIO, facilitating the real-time reconstruction of dense RGB threedimensional maps of the environment. However, in the VIO subsystem, when laser map points are projected onto image frames, issues such as point duplication, occlusion, and depth discontinuities at edges often arise between adjacent image frames, significantly affecting the precision of the VIO system. To address these challenges, we propose an improved R3LIVE algorithm based on outlier rejection and boundary rendering. Specifically, within the VIO system, we project map points from the previous scan into the camera coordinate system and employ a selection process to retain only those map points with optimal states for visual localization and mapping. In the global mapping thread, we use a pixel-wise comparison with neighboring pixels to identify edge-rendering points based on a specified pixel color difference threshold. For non-edge points, we apply bilinear interpolation for smooth rendering, while for edge points, we use nearest-neighbor interpolation for direct rendering. This approach not only preserves the smooth rendering effect but also mitigates the impact of noise colors on the map, effectively resolving the issue of the appearance of blue points in tree and building edges during the construction of dense point cloud maps.