Comparative analysis of real-time appearance-based mapping and cartographer algorithms using deep learning object detection

Abstract Simultaneous localization and mapping (SLAM) shares computation, memory, sensing, and middleware resources with other robotic perception tasks. This study evaluates Cartographer and RTAB-Map in ROS 2 Humble/Gazebo using a TurtleBot3 Waffle Pi, both alone and while a separate YOLOv8n object-detection node runs concurrently on the CPU. Ten runs were analyzed per configuration (40 runs total). Independent Gazebo ground truth and TF-derived SLAM trajectories were converted to planar x-y-yaw TUM trajectories, associated within 20 ms, and evaluated using translation and yaw absolute pose error (APE) and 1 m relative pose error (RPE). Baseline translation APE RMSE was 0.030965 ± 0.007947 m for Cartographer and 0.014013 ± 0.003691 m for RTAB-Map; corresponding translation RPE RMSE was 0.017614 ± 0.002281 m and 0.006209 ± 0.001787 m. With concurrent YOLOv8n, translation APE RMSE was 0.031711 ± 0.007119 m and 0.017177 ± 0.007964 m, respectively. Combined monitored process CPU reached 670.491 ± 15.741% and 624.513 ± 12.703%, while Gazebo real-time factor remained near 1.0. Because the groups were independently teleoperated, baseline-to-YOLO differences are interpreted descriptively rather than causally. The workflow integrates repeated-run localization accuracy, orientation error, resource profiling, and motion characterization under documented simulation conditions.