Sensors,
Journal Year:
2025,
Volume and Issue:
25(10), P. 3108 - 3108
Published: May 14, 2025
The
deployment
of
autonomous
AI
agents
in
smart
environments
has
accelerated
the
need
for
accurate
and
privacy-preserving
human
identification.
Traditional
vision-based
solutions,
while
effective
capturing
spatial
contextual
information,
often
face
challenges
related
to
high
costs,
privacy
concerns,
susceptibility
environmental
variations.
To
address
these
limitations,
we
propose
IdentiFi,
a
novel
AI-driven
identification
system
that
leverages
WiFi-based
wireless
sensing
contrastive
learning
techniques.
IdentiFi
utilizes
self-supervised
semi-supervised
extract
robust,
identity-specific
representations
from
Channel
State
Information
(CSI)
data,
effectively
distinguishing
between
individuals
even
dynamic,
multi-occupant
settings.
system’s
temporal
contrasting
modules
enhance
its
ability
model
motion
reduce
multi-user
interference,
class-aware
minimizes
extensive
labeled
datasets.
Extensive
evaluations
demonstrate
outperforms
existing
methods
terms
scalability,
adaptability,
preservation,
making
it
highly
suitable
homes,
healthcare
facilities,
security
systems,
personalized
services.
Computational Intelligence,
Journal Year:
2025,
Volume and Issue:
41(2)
Published: April 1, 2025
ABSTRACT
Human
activity
recognition
(HAR)
technology
plays
a
major
role
in
today's
world
and
is
used
detecting
human
actions
poses
real‐time.
In
the
past,
researchers
employed
statistical
machine
learning
methods
to
build
extract
attributes
of
various
movements
manually.
However,
typical
techniques
are
becoming
increasingly
ineffective
face
exponentially
increasing
waveform
data
that
lacks
unambiguous
principles.
With
advancement
deep
technology,
manual
feature
extraction
no
longer
required,
performance
on
challenging
problems
can
be
improved.
models
have
such
as
time
consumption,
inaccuracy,
vanishing
gradient
problem.
Therefore,
solve
these
problems,
proposed
study
convolutional
attention‐based
bidirectional
recurrent
neural
network
detect
activities
provided
samples.
The
input
images
first
pre‐processed
using
an
adaptive
bilateral
filtering
approach
improve
their
quality
remove
image
noise.
Then,
crucial
features
recovered
(CNN)
based
encoder‐decoder
model.
Finally,
identify
activities.
model
recognizes
with
higher
effectiveness
lower
latency.
behaviors
identified
HMDB51
dataset.
acquired
highest
accuracy
95.46%,
which
10.51%
superior
multi‐layer
perceptron
(MLP),
6.99%
CNN,
12.76%
long
short‐term
memory
(LSTM),
5.59%
Bidirectional
LSTM
(BiLSTM),
4.82%
CNN‐LSTM,
respectively.
