Optimization of Visual SLAM for Mobile Robots in Complex Environments: Tight Coupling of Visual-Inertial Fusion and Efficient Loop Closure Detection Strategies

This study tackles mobile robots' environmental perception challenges in complexity, presenting an advanced Visual-Inertial SLAM (VSLAM) technique that enhances accuracy, robustness, and real-time functionality. It bolsters visual odometry with ORB features for meticulous matching and selects keyframes, integrating IMU data for robust motion handling in low-textured scenes. Rear-end operations employ graph optimization and a bag-of-words-based loop closure, utilizing sliding window optimization and marginalization to curb accumulative errors, ensuring trajectory and map consistency. Experiments on KITTI, EuRoC, and TUM datasets surpass conventional methods like ORB-SLAM2 and VINS-Mono, trimming trajectory and mapping inaccuracies by 23.8% and 41.7%, respectively, with robust adaptability across motion modes and environments. System module timing analysis paves the way for real-time deployment on less powerful hardware. Future research directions include direct visual odometry, dynamic environment adaptation, multi-sensor fusion, and large-scale scene comprehension, pushing SLAM's frontier in intricate dynamics and empowering autonomous robot navigation.