Real-time Graph-Based 3D Reconstruction of Sparse Feature Environments for Mobile Robot Applications

In the last years, the demand for mobile robots in rescue and surveillance has increased considerably as these systems allow to reduce human harm and save the workforce. A major problem still is the mapping of unknown environments for localization and navigation purposes. Common 2D mapping is not sufficient for 3D environments with multiple levels and 3D structures. Unfortunately, recent approaches for 3D reconstruction suffer from high hardware requirements to meet real-time constraints and the accumulation of errors in the reconstruction result over time. Moreover, the loop closing problem could not be solved satisfactorily yet. In this paper, we propose a new approach for real-time 3D reconstruction that meets the hardware requirements of mobile robots and is capable of detecting and closing loops to reduce errors. We therefore combine and modify three state-of-the-art approaches into a 3D reconstruction system that is also working in sparse feature environments. Each of the subsystems runs in a parallel thread accelerated by the GPU and can easily be replaced by another algorithm if necessary.