Qualitative Analysis of RGB-D SLAM Using RTAB-Map for Indoor and Outdoor GNSS-Denied Environments

Abstract

This study presents a qualitative and semi-quantitative analysis of the Real-Time Appearance-Based Mapping (RTAB-Map) RGB-D SLAM framework utilizing an Intel RealSense D435i sensor running on ROS 2 Humble across both structured indoor and unstructured outdoor GNSS-denied environments. Operating within a controlled 4.5 m x 3.5 m closed-loop path at a fixed sensor height of 0.75 m, the system’s performance was evaluated based on visual odometry stability, 3D point cloud reconstruction quality, loop closure efficacy, and computational resource consumption. The experimental results demonstrate that the indoor environment provides highly favorable conditions characterized by rich geometric textures and stable artificial illumination, enabling the Visual-Inertial Odometry (VIO) pipeline to achieve robust tracking, two confirmed returns to the start, and an exceptionally low final drift of 0.025 m with a peak CPU usage of 25.8% and 4.8 GB of RAM. In contrast, the outdoor environment is very challenging, with ambient sunlight causing significant active infrared depth sensor saturation, dynamic vegetation and moving obstacles. This results in unrecoverable tracking losses, 0 successful loop closures, a final drift above 5.00 m, and increased computational overhead reaching 41.8% peak CPU and 6.5 GB of RAM. The results demonstrate the reliability and efficiency of pure RGB-D SLAM for indoor navigation, but also expose the intrinsic limitations of this method in uncontrolled outside daylight environments, thereby hinting the necessity for multi-sensor fusion techniques in the future.