IEEE Consumer Electronics Magazine - May 2018 - 14

color circle on the inside for the recognition step, which can
be detected in a wider range. The system transforms the redgreen-blue (RGB) to the hue-saturation-value (HSV) color
space and performs binarization on the saturation (S) and
value (V) maps. The circumcircles are extracted using the
connected components and then recognized using machine
learning. A few drawbacks are apparent, including that 1) the
detection result can be easily interfered with by background
noise; 2) the computational cost will be higher as the areas
for processing increase; and 3) there are a limited number of
marker patterns possible.

THE ProPoSEd VISuaL MarkEr
THe mARkeR desigN
Figure 1 shows the structure of the proposed marker. An
example of it is illustrated in Figure 1(a), and the outer colorcircle pattern that is used for detection is shown in Figure 1(b).

This color-circle pattern, which consists of a combination of
the colors red, yellow, blue, and black, is used to localize the
marker from the other objects. The border of the three colors
(excluding black) is used to detect the outside color ring. The
details will be introduced in the next section.

Qr codE dESIgn
The QR code is used for the recognition part [as shown in
Figure 1(c)] because of its potential in representing a vast
number of IDs and its ability to save text information without
a database. It also embeds a specific ID for each location.
The markers are classified by the center color point called
TAG, which yields 5% of the radius of the outer ring. The TAG
helps the proposed system determine the type of the marker
in long-range recognition. To reduce decoding time on high
resolution, the region of interest (ROI)-based process is implemented after the detection step.

THe mARkeR deTecTioN
The detection step is shown in Figure 2. First, the input image
is transformed to HSV and thereby segmented into black,
white, yellow, red, and blue areas in the segmentation step.
If S ^ x h ) V ^ x h $ T1,
(a)

(b)

(c)

FIgurE 1. The structure of the proposed marker: (a) the marker
and the (b) detection and (c) recognition parts.

Start

HSV Color Space

RGB Color Space

Segmentation

Intersection

Clustering

Centering

End

Matching

FIgurE 2. The detection flow.

FIgurE 3. The feature points on the color ring.
14 IEEE Consumer Electronics Magazine

^

MAy 2018

Z
] red, if H ^ x h 2 320 or H ^ x h 1 20
] yellow, if 30 # H ^ x h # 90
P^ xh = [
] blue, if 170 # H ^ x h # 270
] white, if V ^ x h 2 T2
\

(1)

otherwise,
P ^ x h = black,

(2)

where H ^ x h, S ^ x h, and V ^ x h are the hue, saturation, and intensity of the pixel x, respectively. T1 and T2 are threshold values
defined by different cameras.
Second, the intersection step finds two types of feature
points by the border of multiple colors as shown in Figure 3.
One is the border of blue, yellow, and black (blue circle in
Figure 3); another is the border of red, yellow, and black (red
circle in Figure 3). These feature points can be calculated as
Z a, if B ^ x h 7 blue 7 yellow 7 black b red
]]
J ^ x h = [ b, if B ^ x h 7 red 7 yellow 7 black b blue,
] 0, otherwise
\

(3)

where B ^ x h is the block centroid. The size of the block is
5 # 5 to achieve a balance between performance and robustness. The feature point values are a and b.
After the intersection step is performed, all feature points in
the input image have been captured. Next, in the clustering
step, the spatially arranged feature points in specific zones can
be clustered by using the k-dimensional-tree algorithm [20].
Two types of feature points are clustered separately. All feature
points will be clustered into their own group in this step. In
Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May 2018
