Abstract: Aiming at the problems of low mapping and relocalization accuracy, as well as poor scene adaptability, for robots in indoor and outdoor scenes, a tightly-coupled light detection and ranging(LiDAR)-visual-inertial odometry via smoothing, mapping, and relocalization by stereo(LVI-SAM-Stereo)method was proposed. The LiDAR-inertial pose estimation model was constructed by utilizing point-line and point-plane distances. Multi-sensor information interaction enabled rapid initialization of stereo-inertial odometry, with the odometry pose being optimized through reprojection error minimization. A cross-modal loop closure detection mechanism combining Scan-Context with visual features effectively reduced incorrect loop closures. A bidirectional relocalization architecture was developed, where factor graph-optimized odometry provided initial pose estimation for visual tracking, while perspective-n-point(PnP)-derived visual poses assisted LiDAR point cloud registration. A thorough evaluation with both datasets and real-world experiments verified that LVI-SAM-Stereo achieved 3.10% and 5.97% higher outdoor mapping accuracy compared to tightly-coupled LiDAR inertial odometry via smoothing and mapping(LIO-SAM)and tightly-coupled LiDAR-visual-inertial odometry via smoothing and mapping(LVI-SAM), respectively. Indoor average drift decreased by 72.7% and 43.05% versus these benchmarks. The system significantly improved mapping precision and scene adaptability. The relocalization satisfied the engineering requirements for autonomous navigation of robot products.

Key words: robot, multi-sensor fusion, visual-inertial odometry, loop closure detection, relocalization