This project makes it possible to densify Visual SLAM-obtained sparse maps offlie. The use case kept in mind during its development was an autonomous vehicle with limitations in weight, space, computational power and/or cost inspecting a place and navigating it using a light Visual SLAM method. With the map and the images taken by the robot, a dense, color, precise and human-friendly 3D map can be built offline using this method.
More visual results can be found in the provided slides. A demo video of the incremental creation of the dense map is available on YouTube:
This project consists in a method to create a dense three-dimensional representations of the environment in which a single 2D camera is moving with an unconstrained movenet. This is done by fusing the output of a keyframe-based visual SLAM algorithm (ORB-SLAM2) with predicted depthmaps obtained with a Deep Learning method (DeMoN).
The current results provide maps that densify the ORB-SLAM2 output extracting almost 200 times more data from each keyframe (49152 vs 250.54 3D points). The resulting map's average distance to the nearest neighbours of the groundtruths is 0.0399 meters, with a variance of 0.0006. This measures have been computed using various sequences from the TUM RGB-D SLAM dataset.
The files in this repository are divided in the following parts (files/directories):
- docs: Directory containing the documentation files and media. It contains slides summing up the technical aspects and providing qualitative and quantitative results. My Degree Thesis is available here aswell, which can be also found in the institutional Repository of the University of Alicante (RUA). The paper itself can be requested from ResearchGate or found in this Springer book.
- programs: Directory containing the source files and the corresponding CMakeLists.
- full_runCOLOR.bash: Bashscript integration of the whole project to ease and automate the exercution. It incorporates command-line based interaction with the user that allows some personalization of the execution.
This repository contains the source files and some other related documents regarding my end-of-degree thesis. This work has been supported by the Spanish Government TIN2016-76515R Grant, supported with European Regional Development Fund (ERDF) funds. I have been supported by a Spanish Government grant for cooperating with the RoVit Research Group in research tasks ID 998142. This work has been submitted to the ROBOT’19 conference, an international congress that will be held in O Porto, Portugal, in November 2019.