Robust Visual SLAM Framework Based on Human Saccadic Eye Movement Modeling

Abstract

Robust visual simultaneous localization and mapping (SLAM) is fundamental for robotic navigation in unfamiliar or dynamic environments; however, traditional visual SLAM systems often suffer from accuracy and stability losses due to dynamic objects, occlusions, and abrupt scene changes. To address these challenges, we propose a novel SLAM framework inspired by human saccadic eye movements, which incorporates a neurobiologically-based inhibition-of-return mechanism to suppress redundant operations, adaptively allocate computational resources, and emphasize salient information. The proposed system embeds saccade-driven attention directly within the SLAM process, integrating adaptive map maintenance and attention-based real-time feature selection. Experiments conducted across diverse indoor, outdoor, and semi-structured environments—including complex scenes with dynamic traffic—demonstrate that our approach significantly improves performance. Specifically, the saccade-inspired model achieves an absolute trajectory error (ATE) of 0.139 m in dynamic outdoor sequences, outperforming baseline methods such as ORB-SLAM3 (0.176 m) and VINS (0.168 m), and maintains a median feature retention ratio of 69.2% over 3000-frame indoor sequences (vs. 54.6% and 48.1% for baselines). Robustness tests under occlusion and sensor disturbance confirm stable localization and faster recovery. These results validate the effectiveness of human-inspired attention mechanisms for enhancing SLAM robustness, precision, and resource efficiency in real-world, dynamic conditions.