IEEE Consumer Electronics Magazine - January/February 2024 - 73

are generating data exponentially, which can be
used to predict trends. Edge computing is the
architecture in which the processing and storing
of data happens close to the data source. Realtimeprocessing
at the edge is more efficient, accurate,
and necessary as the majority of the data is
created there. The three components of the edge
computing architecture are the cloud layer, the
boundary layer, and the terminal layer.2
EdgeAI deploys AI applications on edge devices
near the data source, enabling deployment of AI in
resource-constrained environments with real-time
insights, reduced latency, cost, and increased privacy.10
You only look once, a state-of-the-art object
detection (OD) system, provides high accuracy
and real-time OD at up to 45 fps, making it useful for
edge AI applications in autonomous vehicles and
surveillance systems.4 Edge AI can be used to process
real-time data from IoT sensors to detect
potential security systems threats. YOLO OD algorithms
can identify physical threats on the edge
and take immediate action. Few physical security
systems in the market can perform real-time detection
using visual, night vision, and thermal video
frames.9 Current real-time AI detection systems
require batch-trained models, which are costly and
time-consuming to update, and deploying the
detection model on the cloud causes data transfer
latency. Deploying AI at theedgewithreal-time
learning can improve accuracy, efficiency, security,
and privacy in security systems, while addressing
dynamic and evolving threats.
In this article, we propose a real-time physical
threat detection framework that combines
edge-based OD AI with online learning to
improve accuracy, efficiency, reliability, security,
and privacy. Edge deployment leads to a
fast and efficient system with minimal resources
required. Online learning allows the model to
learn from new data in real time, with fast and
affordable learning steps.3 The proposed edgebased
model ensures privacy protection by
immediately utilizing and discarding data, while
also being capable of offline operation, making it
dependable and trustworthy. Additionally, the
model's ability to operate through visible and
thermal images further enhances its reliability.
Our solution enhances physical security systems
through the combination of edge computing and
incremental online learning.
January/February 2024
BACKGROUND AND RELATED WORK
Overview of OD
OD is a computer vision technique for detecting,
locating, and classifying objects in images.
Two main types of OD algorithms are two-shot
detection and single-shot detection. YOLO4
takes an image and splits it into S S grid, where
each grid will have parameters defining an object
as [Pc;bx;by;bw;bh;c], where Pc denotes the confidence
score for the object in the box, (bx;by) represents
the center of the box relative to the grid
cell, (bw;bh) represents the width and height relative
to the whole image, and c denotes the presence
of each class in the cell. YOLO is fast,
accurate, and the best option for real-time OD,
utilizing anchor boxes and nonmaximal suppression
to improve the performance.4
Overview of Online Learning (OL)
ML has flourished with the availability of
data, storage, and processors, but data generation
surpasses its usage rate. Online learning
trains models in real time by incrementally learning
from continuous data input, eliminating the
need for data storage.3 However, faulty data
could negatively impact performance, so proper
data governance is required.
Related Work and Objectives
In airport security, deep learning approaches
have been utilized to detect potential threats in Xray
and thermal infrared (TIR) images.5 The effectiveness
of different versions of YOLO models is
evaluated using Teledyne Forward-looking Infrared
(FLIR) Thermal Dataset.6 Another study utilized
YOLOv3 and demonstrated its performance using
a combination of pretraining on Imagenet and a
custom gun dataset.7 Edge YOLO is a lightweight
version of the YOLOv4 OD algorithm that has been
optimized for edge computing.11 The algorithm features
a trimmed-down backbone using CSPNet,
enhanced feature fusion, and a connection to a
cloud-based GPU workstation for model training.
Edge YOLO outperforms other popular algorithms,
such as YOLOv3 and MobileNetv3 SSD, in terms of
both speed and accuracy on the edge.
Although research was experimented to
make real-time inferences with image and video
frames with OD and edge computing as a single
73
IEEE Consumer Electronics Magazine - January/February 2024

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