Monocular 3D Scene Reconstruction for an Autonomous Unmanned Aerial Vehicle

Matouš Melecký


The real-time 3D reconstruction of the surrounding scene is a key part in the
pipeline of the autonomous flight of unmanned aerial vehicle (UAV). The com-
bination of an inertial measurement unit (IMU) and a monocular camera is a
common and inexpensive sensor setup that can be used to recover the scale of the
environment. This thesis aims to develop an algorithm solving this problem for
this particular setup by leveraging the existing visual-inertial navigation system
(VINS) odometry algorithms for localisation.
Two algorithms are developed, wide-baseline matching-based and small-baseline
tracking-based. Also, an offline post-processing structure-refinement step is im-
plemented to further improve the resulting structure. The algorithms and the
refinement step are then evaluated on publicly available datasets. Furthermore,
they are tested in a simulator and a real-world experiment is conducted.
The results show that the tracking-based algorithm is significantly more perfor-
mant. Importantly, tests on the datasets and the real-world experiments suggest
that this algorithm can be practically employed in application scenarios.
Keywords: unmanned areal vehicle, 3D scene reconstruction, structure from
motion, monocular simultaneous localisation and mapping, non-linear least
Thesis Attachments

Source code, thesis in pdf, etc.


Datasets as recorded for evaluation of the 3D reconstruction of outdoor scene at Karlovo namesti.

Resulting point clouds

Resulting point clouds created by the algorithms, including the visualisation of their distance to reference, are available at:
Note that it may take some time to load the point clouds into the browser.

Video material