vEMB-SLAM: An Efficient Embedded Monocular SLAM Framework for 3D Mapping and Semantic Segmentation

This study introduces a novel approach to cost-efficient and real-time 3D mapping and object detection, leveraging monocular camera systems to overcome the high hardware demands of traditional SLAM systems. The proposed vEMB-SLAM algorithm integrates advanced depth estimation models (MiDaS, Metric3D, and Zoe Depth) with semantic segmentation techniques powered by PointNet++. The system is designed for embedded platforms and utilizes modular components for data acquisition, pose estimation, and point cloud segmentation. Experimental evaluations on the newly developed AirSim Indoor Scenes Dataset (AIS dataset) demonstrate the algorithm’s capability to generate accurate 3D maps and perform robust semantic segmentation in diverse environments. The results highlight the potential of low-cost monocular systems to facilitate autonomous navigation and mapping across resource-constrained applications. This work provides a foundation for further advancements in monocular SLAM methodologies and their deployment in real-world scenarios.