Graphics Reference
In-Depth Information
Table 4.2
Evaluation of landing site detection
Altitude (m)
Baseline (m)
Reprojection
Number of frames
FP (%)
FN (%)
error (pixel)
Visible
Successful
1.47
0.39
0.21
116
106
91.4 %
0.028
39.34
2.19
0.86
0.20
63
62
98.4 %
0.046
43.57
3.34
2.00
0.18
67
63
94.0 %
0.0026
53.14
6.51 1.16 0.28 337 325 96.3 %
10.15 2.28 0.22 480 433 90.2 %
Row 1-3
Indoor experiment at three different altitudes (
r
=
13 cm, depth accuracy at ground
level
=
3 cm),
Row 4-5
outdoor experiment (depth accuracy at ground level
=
20 cm)
Fig. 4.19
Second indoor experiment:
a
box height calculation for valid landing point with highest
confidence.
b
Feature reprojection error of feature matches and estimated baseline (depth accuracy
3cm)
Our approach is able to robustly detect the landing zone with a success rate of more
than 90 % in all experiments and a false positive rate below 0.05 %. The false positive
(FP) and false negative (FN) rates are largely defined by the quality of the disparity
input. Border-fattening effects (caused by our correlation-based stereo matching)
usually increase the FP rate, whereas missing disparity pixels on top of the target
increase the FN rates, since we treat missing data as unsafe. To mitigate these two
effects, we introduced two thresholds to maximize safety: (1) at disparity edges, we
disable all pixels that are located within half a correlation window size to the edge,
and (2) we use a percentage threshold which defines the minimum number of pixels
with valid disparity around a landing site (98 % in our experiments).
In the second experiment, in order to verify the accuracy of our 3D reconstruction,
we plotted the height of the landing point with the highest confidence for one of
the sequences (Fig.
4.19
). The error follows the expected depth accuracy of 3 cm
(min. depth acc.) within the true box height of 27.5 cm. The low average feature
reprojection error confirms valid motion estimation results.
The third indoor landing experiment consisted of landing site detection and target
approach. In this indoor experiment, the landing target also consisted of a cardboard
box to simulate an elevated landing surface. We commanded the quadrotor to fly
