ADVI-SLAM: An Adaptive Multi-Modal Data Fusion Localization Method for Mobile Robots

Current purely vision-based SLAM algorithms often suffer from visual data failure in complex environments, such as pose tracking loss and poor robustness, leading to catastrophic tracking failure or significant drift. To ensure high accuracy and generalization stability, we build upon binocular visual SLAM by adopting a tightly-coupled approach to fuse IMU inertial data. Through nonlinear optimization, we achieve deep complementarity between visual and inertial information, effectively enhancing system stability in challenging scenarios such as weak textures, sudden illumination changes, and rapid motion. To address these challenges, we introduce ADVI-SLAM, which is an adaptive multi-modal information fusion localization method for mobile robots. Furthermore, to overcome the drawbacks of traditional marginalization strategies, specifically their strong priors and tendency to introduce erroneous constraints, we innovatively design an adaptive dynamic marginalization algorithm. Based on sliding window optimization theory, this algorithm incorporates a feedback mechanism to dynamically adjust the marginalization strategy, preventing abrupt jumps in state estimation and thereby further ensuring consistency in the global manifold. Extensive experiments on the EuRoc and KITTI datasets demonstrate that the proposed ADVI-SLAM maintains high positioning accuracy while exhibiting strong environmental adaptability and system stability.