Aerospace and Electronic Systems - March 2019 - 22

Robust Cooperative Target Detection for a Vision-Based UAVs Autonomous Aerial Refueling Platform via the Contrast . . .
is conducted to all the morphological dilated binary maps
at all scales and the output of the MFL is described by:
Cout ðx; yÞ ¼ bs1 ðx; yÞ

ðbsi È Dsi Þðx; yÞ

(13)

where Dsi is the structuring element for morphological
dilation. In this paper, it is a disk of radius s i where s i is
the ith resolution of all scales with ascending order.

TARGET SELECTION RECOGNITION LAYER

When the frames of the cooperative target are successively
acquired, the visual measurement algorithms are utilized to
determine the relative pose information. The process of
visual measurement can be divided into four steps, cooperative target detection, feature extraction, point matching and
pose estimation. Finally, the visual measurement system is
implemented in the AAR platform designed in our work.

FEATURE EXTRACTION AND FEATURE MATCHING

After the contours of four scales are fused, the morphological closing operation is conducted to the fused contour image, so that the breaks of the salient contour of
the target can be eliminated as much as possible.
Thereby, the cooperative target area can be filled.
According to the minimum pixels of the cooperative
target in the actual situation, an appropriate threshold is
set and those connected components less than the set
threshold number of pixels are excluded. The corresponding binary images of each remaining connected
component are separated and their Hu's moment invariant vectors are calculated. The Euclidean distance
between them and the Hu's moment invariant vector of
the cooperative target is calculated subsequently. The
corresponding binary image with the smallest Euclidean
distance is selected, and in this binary image, the pixels
with value 1 form the of the cooperative target area.
In order to illustrate the advantages of the cooperative
target detection method, we perform the proposed method
on images from two challenging lighting conditions and
present the results in accordance with the above described
five layers. As a comparison, the Canny edge detection operator is used to the same images. The experiment results are
shown in Figure 4.
From Figure 4(b), we can see that OSNL prefers the
salient contour while the texture edges are restrained by
the layer, and the DL can smooth the energy maps and
retain its principal details as shown in Figure 4(c). Our
proposed detection method is able to extract the salient
contours effectively and to remove a lot of texture edges
and the cooperative targets for two challenging lighting
conditions are successfully detected, shown in Figure 4(f).
By contrast, the Canny detection results in Figure 4(g)
shows that it failed to extract the cooperative target
since there are complex contours in the binary images.
In our test, the value of four scales are set
fs 1 ¼ 1; s 2 ¼ 2; s 3 ¼ 3; s 4 ¼ 4g for strong light condition and fs 1 ¼ 11; s 2 ¼ 12; s 3 ¼ 13; s 4 ¼ 14g for low
light condition. The inhibition factor k has an important
effect on the performance of our proposed detection
method, and it is necessary to design appropriate inhibition factor k according to the specific situation. In our test,
k ¼ 1:5 is acceptable.
22

AAR PLATFORM

After target detection, the cooperative target area is
obtained. Feature extraction only concentrates on this
area, which greatly reduces the computation load and
improves the accuracy. In our platform, the main task of
feature extraction is to detect seven red markers on the
surface of the cooperative target and calculate the corresponding coordinates, respectively. Since HSV space has
better ability to represent color features of those red
markers relative to RGB space, the image is first transformed from the RGB space to the HSV space. Then HSV
segmentation [42] is utilized to extract those red markers
in this paper. Finally, the feature point coordinates of
those red markers are determined by the centroids of these
corresponding blobs and the quantity of blobs is counted
as the number of detected feature points.
After feature extraction, it is required to correctly
associate each detected corner with its physical marker on
the receiver UAV before pose estimation. Assume that
_
_ _
_
q ¼ fq1 ; q2 ; . . . ; q7 g represents actual world coordinates
_
_ _
of seven red markers, where qi ¼ ðui ; vi Þ is the ith marker,
and p ¼ fp1 ; p2 ; . . . ; pm g denotes the detected feature
point coordinates, where pi ¼ ðui ; vi Þ indicates the ith feature point coordinate obtained from feature extraction.
_
_ _
_
The projected point coordinates p ¼ fp1 ; p2 ; . . . ; p7 g can
be calculated by the standard pin-hole projection model.
The Euclidean distance matrix between the reprojected
_
_ _
_
point coordinates p ¼ fp1 ; p2 ; . . . ; p7 g and the feature
point coordinates p ¼ fp1 ; p2 ; . . . ; pm g is denoted as:
2

dðp1 ; p1 Þ dðp1 ; p2 Þ Á Á Á
6 dðp_ ; p1 Þ dðp_ ; p2 Þ Á Á Á
2
2
6
Err ¼ 6
..
..
..
4
.
.
.
_
_
dðp7 ; p1 Þ dðp7 ; p2 Þ Á Á Á

3
_
dðp1 ; pm Þ
_
dðp2 ; pm Þ 7
7
7
..
5
.
_

dðp7 ; pm Þ

(14)

The problem of point correspondence can be described
by the classical assignment model, and the Munkres algorithm [43] is adopted to resolve the optimal assignment
problem by labeling each peak. We have implemented the
algorithm in our previous work [15].

IEEE A&E SYSTEMS MAGAZINE

MARCH 2019

Aerospace and Electronic Systems - March 2019

Table of Contents for the Digital Edition of Aerospace and Electronic Systems - March 2019