Toward localization and mapping with heterogeneous depth sensors

Abstract: Heterogeneous collaborative Simultaneous Localization and Mapping (SLAM) can be defined as the solution to the SLAM problem that can handle different devices with different sensors, such as a monocular camera and a 3D LiDAR sensor, building a map and performing localization all at the same time. Research regarding this field is still at a very early stage, and it is hard to find solutions to the data association problem for these different types of sensor outputs. The increasing presence of devices and autonomous agents in our surroundings and everyday life is raising the question of how to make these very different devices collaborate. The goal of this thesis is to expand on previouswork regarding the use of SegMap as a map representation solution based on segmentation of point clouds that can handle heterogeneous data coming from sensor-specific and device-specific SLAM modules, and in this particular case, stereo and LiDAR data interchangeably. Stereo depth reconstructed point clouds do not share the same properties as LiDAR point clouds as is. Therefore, this thesis benchmarks the previous framework, introduces a new post-filtering module for stereo data that includes outlier filtering and point cloud registration, and finally shows results pointing out how the framework could be used robustly for heterogeneous place recognition.