LiDAR and Visual Perception-Based Indoor Semantic Mapping: Comparative Study of GMapping and SLAM Toolbox

Abstrak

This study investigates semantic embedding strategies for indoor mapping by comparing Trajectory-Based Payload Embedding (TPE) in GMapping and Pose-Based Payload Embedding (PPE) in SLAM Toolbox. A custom Turtlebot3 platform equipped with a 2D LiDAR and six RGB cameras was used in the Gazebo simulation to acquire geometric and visual data. Object segmentation results from YOLOv11 were integrated into occupancy grids using two distinct embedding workflows: scan-level batch attachment in TPE and point-level graph persistence in PPE. Performance evaluation employed two metrics: pixel-level accuracy and time cost under three varied velocity conditions, followed by a comparative analysis. Results show that PPE achieved higher accuracy (mean 86.83%) and lower variability, while maintaining negligible time cost (<0.5 ms). TPE, although simpler to implement, exhibited greater sensitivity to motion dynamics and higher computational variability (average 350.47 ms). These findings highlight a trade-off between accuracy and efficiency, suggesting PPE as the more suitable approach for real-time semantic SLAM, while TPE remains useful for lightweight integration scenarios. Beyond quantitative results, the study contributes methodological insights into how embedding granularity and persistence affect semantic consistency, offering guidance for future implementations in both simulated and real-world robotic navigation.