Real‐Time Monocular 2D Occupancy Grid Mapping for Autonomous Navigation of Ground Robots

ABSTRACT This study proposes a lightweight and real‐time 2D occupancy grid mapping method for autonomous ground robots using a monocular visual‐inertial system. While traditional simultaneous localization and mapping (SLAM) systems typically produce sparse point clouds optimized for localization, these point clouds lack the completeness required for direct use in navigation. To bridge this gap, the proposed method transforms sparse keyframe‐based point clouds into a navigable 2D occupancy grid map through a reflection probability model, without relying on depth sensors or dense 3D reconstruction. To address the common problem of missed obstacle detection, especially when the camera is positioned too close to obstacles, a novel multi‐conditional filtering algorithm has been introduced. This algorithm combined multi‐frame feature tracking with geometric constraints to robustly identify both static and dynamic obstacles, even in low‐texture environments. By avoiding computationally intensive 3D mapping and dense data processing, the method achieves low memory consumption and excellent real‐time performance. Experiments conducted in real‐world indoor environments demonstrate that the proposed system outperforms Octomap, improving map accuracy by an average of 6.76 percentage points (7.6% relative increase) and map recall by 5.13 percentage points (5.5% relative increase). In addition, it requires less memory and processing time, highlighting its superior mapping quality and robustness. The results confirm that this method effectively connects localization‐level sparse visual SLAM data with navigation‐level environment understanding, offering a practical and efficient solution for autonomous navigation on resource‐constrained platforms.