Learning-Based Matching of 3D Submaps from Dense Stereo for Planetary-Like Environments

An autonomous robot typically requires a minimum capability of perceiving the surroundings and locating itself when it is deployed to an unknown environment. Such a task is generally known as Simultaneous Localization and Mapping (SLAM), for which pairwise submap matching is a common foundation for subsequent processes to construct a global map around the robot. While the task has been extensively studied and successfully accomplished with different advanced solutions, their applied domains are rather constrained within indoor or structured regions. In this paper, we enhance a seminal learning-based approach, 3DFeat-Net, with more sophisticated architectures, and evaluate them in extremely unorganized planetary-like environments. Our work demonstrates that the proposed enhancement performs better than classical feature-based algorithms, and therefore outlines a promising direction for future work.