IEEE Consumer Electronics Magazine - July/August 2022 - 51

Figure 6. Proposed Network Training Curve.
tracker to detect the moment of abandoning by
calculating the distance between the hand and
luggage. Specifically, we need to trace the coordinate
of the hand luggage using the high-speed
KCF tracker. However, there are various factors
of false detection of hand luggage, such as occlusion,
noise, and image shaking while tracking
pedestrians. In the proposed system, hand luggage
regions are sorted to remove falsely
detected regions. Falsely detected objects from
the pedestrian are usually caused by occlusion,
and unlike hand luggage, they are not detected
in continuous frames. On the other hand, an
action of pedestrians dumping hand luggage is
shown in a set of consecutive frames, which will
be shown in detail in section " Experimental
Results. " In the proposed system, the detected
hand luggage is not immediately tracked, but
stored for comparison with the next frame. The
hand luggage detected in the next frame uses
the histogram comparison with the previous
frame, through which the similarity between the
two objects is measured. The histogram comparison
used as a similarity comparison method
compares color values in the hue channel. This
method is relatively simple and efficient with
less computational complexity than template
matching or feature matching methods. If the
similarity is higher than the specified threshold
using this the histogram comparison, the object
is determined as a candidate of the abandoned
luggage counted. In this system, if a similar
object is detected more than five times in a row,
it is determined as hand luggage to judge abandonment,
and this system is also used for
retracking when hand luggage tracking fails. We
measure the distance between the abandoned
luggage and the owner using the sorted hand luggage
using the KCF tracker12 to measure the distance
from the owner. The KCF tracker used as a
tracking algorithm is a method that does not use
deep learning and is a suitable for algorithm as a
tracking method in this article, because it shows
fast speed and outstanding acceptable accuracy.
Finally, to determine the abandoned object,
the distance between the owner and the hand
luggage being tracked is calculated as shown in
Figure 8. We cited Luna's survey paper22 for the
conditions, and verified them through experiments,
where W is the owner's width and D is
the distance from the hand luggage. In the proposed
system, if the distance (D) is more than
two times larger than the width (W) and this
state is maintained for 5 s, it is determined as an
abandoned object.
Experimental Results
In this article, the abandoned object is
detected based on the detected hand luggage
using the network proposed in section
" Proposed Method. " This section describes the
experimental results using the ABODA datasets,23
and compares the performance of various
methods in the sense of intersection over union
(IOU), which is a good indicator to demonstrate
the performance of the HLDNet.
The ABODAdatasets are often used to evaluate
the performance of abandoned object detection.
These datasets were created by Lin, and scenarios
in the datasets include outdoor, night, illumination
change, indoor, and crowded scenes.24 Table 2
Figure 7. Abandoned Object Decision System Framework.
July/August 2022
51
IEEE Consumer Electronics Magazine - July/August 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - July/August 2022
