IEEE Consumer Electronics Magazine - January/February 2024 - 75

system's adaptability. Common benchmark datasets
such as ImageNet and COCO are often insufficient
due to the limited availability of annotated
data with diverse object labels. There are few
labeled datasets that identify objects as threats,
requiring custom images and videos with threat
objects and proper label annotations to design the
proposed online physical threat detection system.
Pretrained models on large datasets are frequently
used and then retrained on customdatasets.
Although faster R-CNN and RetinaNet are more
accurate than YOLO, YOLO's real-time detection
capability outshines other detection algorithms.
In Jiang et al.6 YOLOv3-SPP was identified as having
high mAP and precision, but not ideal for edge
deployment due to its low speed and high storage
requirements. YOLOv5-s is recommended for
edge deployment with a compact size of 14 MB,
fast speed of 41 FPS, and high mAP and precision
of 0.803 and 0.638, respectively, outperforming
YOLOv3-SPP. GhostNet12
generated reduced boundary box
learning and implementation phases, which are
explained in detail in this section.
Learning Phase
The learning phase consists of the process of
preprocessing the dataset, using the data to
train the YOLO model, and using the validation
dataset to test the model.
To improve the algorithm's ability to predict
images, data augmentation is necessary for both
the training and validation datasets. This
includes random crop, rotation, flips, saturation,
and exposure shifts. Manual labeling is required
for the augmented images and newly captured
images containing threat labels. A minimum of
1,500 images per class and 10,000 labeled objects
per class are recommended, and adding background
images with no objects can help reduce
false positives.
To make the OD model operational, the proand
coordinate attention
(CA)13 methods can make YOLOv5-s even
more lightweight but equally efficient. Using both
techniques14
loss, classification loss, and object confidence
loss in self-constructed data.11 It demonstrated
better performance of the YOLOv4 version with a
trimmed-down CSPdarknet53 and a modified
backbone neck with spatial pyramid pooling
(SPP) and feature pyramid network (FPN) for costeffectivesmall
OD.
The modified GhostNet and CA backbone of
YOLOv5-s can be improved with the same lightweight
neck structure used in YOLOv4 to
enhance our physical threat detection model
performance. YOLOv5 uses batch normalization,
leaky ReLU activation, stochastic gradient
descent optimization, binary cross-entropy for
classification loss, and mean squared error for
coordinate regression. Mean average precision
(mAP), intersection over union (IoU), precision,
recall, and frames per second (FPS) are used to
evaluate the proposed system.
DESIGN OF THE PROPOSED
FRAMEWORK
The proposed real-time physical threat
detection framework with online learning is presented
in Figure 1. It is composed of two parts:
January/February 2024
posed YOLO model is fed with annotated thermal
and visual images. Pretrained weights are
suggested for small or medium datasets. The
YOLOv5 model's default settings obtain good
results on large and well-labeled datasets with
300 epochs recommended. Overfitting can
occur if the model is too complex. Regularization
techniques such as L1 (Lasso) and L2
(Ridge) regularization can be used in such
cases. Cross validation can also be used to
avoid overfitting. The batch size should always
be as large as the hardware allows as a lower
batch size generates poor batch normalization
statistics. YOLOv5 has approximately 30
hyperparameters for training. A genetic algorithm
(GA) is provided to optimize these
hyperparameters, producing an optimal value
by repeatedly mutating parent hyperparameters.
For best results, 300 generation cycles
are recommended, and the hyperparameters
from the best-performing cycle are chosen for
the model.
The performance of the YOLO model on the
test image dataset can be evaluated using the
validation dataset. This evaluation involves
estimating the model's localization and classification
errors, using the evaluation metrics
IEEE Consumer Electronics Magazine - January/February 2024

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2024
