Graphics Reference
In-Depth Information
4.4 Autonomous Landing
In this section, we introduce autonomous landing as a high-level navigation
application. Even if autonomous landing is a specific task, its single parts can easily
be adapted to a series of different applications. Our approach comprises two research
topics, dense monocular 3D reconstruction and visual surface analysis. First, we will
describe the related work, then introduce our landing algorithm, followed by the
embedded implementation, and concluded with experimental results.
4.4.1 Related Work
Most prior work on autonomous landing of unmanned aerial vehicles addresses
landing on terrain, instead of finding elevated perches like rooftops. Due to the
severe constraints on size, weight, and power (SWaP) for especially micro aerial
vehicles, applicable methods must use much lighter, lower performance sensing
and computing resources than available on larger scale systems [
53
]. Approaches
amenable to these SWaP constraints frequently employ monocular [
10
,
26
,
40
,
60
]
and binocular stereo [
37
,
61
] camera systems to map and analyze terrain. Most
approaches perform some form of 3D terrain reconstruction, then assess planarity
and slope of appropriately-sized terrain patches. Binocular stereo vision approaches
are, due to the fixed intercamera geometry, algorithmically simpler, but are limited
by the fixed interocular baseline and also heavier due to the additional camera.
Three monocular approaches are particularly relevant here. The first tracks point
features to estimate homographies from image pairs for predominantly planar terrain,
then analyzes correlation coefficients for dense matches to segment in-plane and
out-of-plane pixels [
10
]. The second uses a recursive filter at each pixel, image
matching via gradient descent with intensity derivatives, and a plane plus parallax
formulation of structure from motion to estimate dense elevation maps from image
sequences [
60
]. Both of these address finding landing sites on the ground. The third
uses multiplanar homography alignment with tracked features to segment a planar
ground-level surface from an elevated, planar landing site [
11
].
In the last couple of years, significant progress has been made in dense monocular
3D reconstruction as well. Recently published approaches use Bayesian and varia-
tional estimation models with known camera motion [
43
,
46
,
59
]. All of them use
powerful processing hardware, such as GPUs, to achieve real-time capability. An
overview about earlier work can be found in [
55
].
4.4.2 Algorithm Description
Our approach can find flat landing platforms everywhere in the 3D model and is
not limited to dominant planes. The presented algorithm consists of three parts,
whereas the whole approach is designed to achieve reasonable short and constant
processing time even on limited computing hardware. First, we use a dense motion
