Toward realtime manifold mesh reconstruction from sparse data

In the last decade the growing interest about autonomous driving and unmanned vehicles in robotics stimulated the research in the understanding of the surrounding through the reconstruction of the environment. The reconstructed map is needed to plan a path to reach a destination or to let the robot localize itself by comparing its perception with the map. In Robotics this requires the reconstruction algorithms to be scalable, incremental and realtime. The use of a mesh in the map reconstruction instead of sparse or dense points allows a continuous representation of the environment. Scalability is needed especially in large-scale environments. Realtime and incremental algorithms allow to update the map as new data are acquired and to make it available to the robot as soon as possible. Although in Computer Vision the research focused on detailed and accurate reconstruction algorithms while disregarding incremental approaches, the methods used in these two fields are useful to each other and are converging. For this reason the manifold property becomes interesting for both fields, as it allows the algorithms from Computer Vision to refine the map adding details and accuracy and consequently providing a consistent and coherent navigability. As the main contribution of this thesis we improve the performances of the state of the art in the reconstruction of manifold mesh from sparse point clouds with the objective of filling the gap between realtime mapping and detailed map reconstruction. Moreover, we extended the algorithm to manage the updates of 3D points and camera positions We tested our implementation against the state of the art results and demonstrate that our system can be used to reconstruct the environment incrementally and nearly realtime on a CPU.