IEEE - Aerospace and Electronic Systems - July 2022 - 36

A Vision-Based Guidance Algorithm for Entering Buildings Through Windows for Delivery Drones
Figure 4.
Influence of the DAE parameter t in thresholding a sample image segment. The corresponding values for t from left to right are
f0:15; 0:20; 0:27; 0:41g.
Therefore, spanning the DAE parameter t from 0 to 1, it is
expected that the corresponding value ofs suddenly increases
in a short interval in the beginning and experiences softer
increases afterward. To better see this, we investigate the
derivative ofs with respect to t,defined as
s0 ¼
dsðtÞ
dt
:
(3)
Figure 5(a) and (b) shows how the values of s and s0
change with respect to the threshold value. For a given
cropped image, increasing t gradually will lead to sequential
sudden rises in s, each of which corresponds to the
emersion of objects with uniform brightness. Equivalently,
these emersions could be better detected at every
dominant peak in s0 chart. So the first local minima after
the main peaks in s0 plot, designated by the dashed lines
in Figure 5(b), are the proper t candidates set. A fast peak
detection method [30] is then utilized to find mentioned
local minima. Experiments show that the most desirable
DAE parameter is among this candidates set.
FINDING THE GLOBAL OPTIMUM
Another approach to determine a desirable t, is to treat the
problem as an optimization task. This is achieved by taking
two steps. First, dividing the image into a number of segments,
that ideally, each segment is a distinct section of the
building facade (a window glass, portal section, window
frame, etc.). Second, defining a cost function that is minimized
when the correct entrance section is completely white
and other sections are black.
SEGMENTATIONSTEP
One fast and fairly accurate way ofsegmentation on facades is
introduced in [12]. Inspired by this work, given an image segment
as in Figure 6(a), we first apply an edge detector, namely
a Laplacian filter, on the facade image [see Figure 6(b)]. Next,
the horizontal and vertical projections on the edge image are
summed up generating, the corresponding horizontal and vertical
histograms, as shown in Figure 6(c). Peaks ofthese histograms
are most likely to be the window boundaries and will
be considered so. By detecting the horizontal and vertical
peaks, a grid on the image frame could be created in which
every partition of the grid is likely to be a specific facade element.
In an ideal scenario, the outcome would be similar to
Figure 6(f). In practice, the image usually gets oversegmented,
as showcased in Figure 6(d). Moreover, for the success ofthis
segmentation, it is needed to have a rectified view over the
facade image.1 In the following optimization step, an effective
cost function for finding the desirable t is formulated.
OPTIMIZATIONSTEP
Once the input image is divided into a set of segments S,as
described in the previous step, the darkest segment s 2 S
will be found by taking the average ofthe value channel over
each segment, and the segment with the lowest average value
will be considered as s. Then, the binary image B will be
Figure 5.
(a) Graph ofs, the total values ofbinary image pixels with respect
to the threshold value. (b) Graph of s0, the image derivative of s.
The dashed lines designate the candidates set.
36
1It should be noted that the algorithm also worked well on some
images containing oblique views over the facade, as can be seen in
Figure 11.
IEEE A&E SYSTEMS MAGAZINE
JULY 2022

IEEE - Aerospace and Electronic Systems - July 2022

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