IEEE - Aerospace and Electronic Systems - July 2022 - 39

Fahimi et al.
The terms Dy and Dz are the last y and z deviations from
portal center position, just before the vehicle loses sight of
portal boundaries.
EXPERIMENTAL EVALUATION
EXPERIMENTS ON FACADE DATASETS
Figure 10.
Result ofDAE on CMP and eTRIMS datasets.
LATERAL GUIDANCE
In order to verticalize the MAV path relative to the window
plane, a Vb
y;cmd command is calculated according to
Vb
y;cmd ¼ KP;xKP;ycp
where cp
cv
(14)
cv is defined as the deviation of the vehicle position
vector expressed in the portal coordinate system from
the window plane normal direction, which is better illustrated
in Figure 1. Also, KP;x
the heading is aligned before verticalization.
FINAL ENTRANCE PHASE
As the MAV gets closer to the window portal (in our
case, when the distance D is less than approximately
1m), the boundaries go out of the camera field of
view. So the described strategy is not sufficient anymore.
In this case, we calculate two approximate
velocity set points to be followed. The time to reach
the window is calculated as
T ¼
and then Vb
y;cmd and Vb
D
Vforward
z;cmd are found accordingly
Vb
y;cmd ¼
Vb
z;cmd ¼
JULY 2022
Dy
T
Dz
T
:
(16Þ
(17)
(15)
is multiplied to make sure
The ability of the proposed method to detect the correct
entrance section is tested on CMP [33] and eTRIMS [34]
datasets. Those images of the mentioned datasets that contain
an open window are chosen, and the window section
of the image is assumed to be cropped by a proper window
detection method. The results are summarized in
Figure 10. As can be seen, the DAE method shows superior
efficacy in identifying the correct entrance section
and is comparatively well in detecting the " exact " boundaries
ofwindows.
In Figure 11, examples of successful implementation of
DAE are presented. In some cases, such as Figure 11(a)-(e),
and (i), it is clear that the entrance section is identified correctly
and also the exact boundaries are detected, enabling
the MAV to use the perspective-n-point algorithm in order
to estimate 6-DoF position data. Due to some adverse conditions,
the exact boundaries could not be obtained for cases,
such as Figure 11(f)-(h), but the correct entrance section is,
however, detected to guide theMAV inside without verticalization.
In Figure 11(e), the correct section is detected
despite the presence ofan occlusion. Figure 11(f) shows that
the assumption of the entrance section as being the darkest
section in the window is valid even ifthere are some lighting
sources inside. Figure 11(c) indicates that the implemented
cost function is able to find the correct threshold value in
nonorthogonal perspectives as well.
A comparison with the ADR-Net [20] method, which is
an extension to the single shot detector[35], is also provided
in Figure 11. As it is clear, the object detection method is
able to find window frames in multiple occasions so that the
total window detection accuracy of ADR-Net is 81%
of cases while it is 75% for DAE. However, when it
comes to comparing the overlap of detected window
and actual portal section, ADR-Net has very poor performance,
hence, using its output bounding box for
guiding a robot inside a real building will inevitably
result in collisions. A valid reason for that could be
the fact that ADR-Net is trained on synthetic gates and
windows and not real-world portals. The rectangle output
of object detection methods proves not to be appropriate
for the task at hand, whereas the quadrilateral output of
DAE is a better approximation for navigation w.r.t. the
window. Moreover, the DAE method is more suitable for
a real-time implementation, since it runs much faster
than ADR-Net on a same processing unit.
IEEE A&E SYSTEMS MAGAZINE
39
IEEE - Aerospace and Electronic Systems - July 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - July 2022
