Computer
vision-based
3D
pose
estimation
for
automated
excavator
operation
monitoring
requires
numerous
training
images
annotated
with
labels.
Owing
to
challenges
in
collecting
such
datasets
a
field
setting,
using
synthetic
from
virtual
environments
has
emerged
recently.
However,
lack
the
realism
inherent
onsite
images,
potentially
impacting
performance
on
real
images.
This
paper
thus
proposes
generative
model
generating
realistic
multiple
backgrounds.
The
evaluation
was
conducted
by
comparing
models
trained
(Model
#1),
generated
single
background
#2),
and
backgrounds
#3).
Model
#3
exhibited
lowest
mean
angular
error
of
5.96°
data,
implying
its
superiority
generalizing
patterns.
proposed
facilitates
data
acquisition
improving
without
manual
annotation,
providing
rich
information
movements
proactive
safety
productivity
management.
Smart Construction and Sustainable Cities,
Journal Year:
2024,
Volume and Issue:
2(1)
Published: July 2, 2024
Abstract
Digital
visual
data,
such
as
images
and
videos,
are
valuable
sources
of
information
for
various
construction
engineering
management
purposes.
Advances
in
low-cost
image-capturing
storing
technologies,
along
with
the
emergence
artificial
intelligence
methods
have
resulted
a
considerable
increase
using
digital
imaging
sites.
Despite
these
advances,
rich
data
not
typically
used
to
their
full
potential
because
they
processed
documented
subjectively,
several
contents
could
be
overlooked.
Semantic
content
analysis
annotation
enhance
retrieval
application
relevant
instances
large
databases.
This
research
proposes
an
ensemble
approach
use
deep
learning-based
object
recognition,
pixel-level
segmentation,
text
classification
medium-level
(ongoing
activities)
high-level
(project
type)
still
from
outdoor
scenes.
The
proposed
method
can
annotate
without
actors,
i.e.
equipment
workers.
experimental
results
shown
this
annotating
activities
82%
overall
recall
rate.
Computer-Aided Civil and Infrastructure Engineering,
Journal Year:
2025,
Volume and Issue:
unknown
Published: May 3, 2025
Abstract
Pose
estimation
of
excavators
is
a
fundamental
yet
challenging
task
with
significant
implications
for
intelligent
construction.
Traditional
methods
based
on
cameras
or
sensors
are
often
limited
by
their
ability
to
perceive
spatial
structures.
To
address
this,
3D
light
detection
and
ranging
has
emerged
as
promising
paradigm
excavator
pose
estimation.
However,
these
face
challenges:
(1)
accurate
annotations
labor‐intensive
costly,
(2)
exhibit
complex
kinematics
geometric
structures,
further
complicating
In
this
study,
novel
framework
proposed
full‐body
directly
from
point
clouds,
without
relying
manual
annotations.
The
parameterized
using
parameters
primitives
under
kinematic
constraints.
A
unified
deep
network
designed
predict
clouds.
initially
pre‐trained
synthetic
data
provide
parameter
initialization
then
fine‐tuned
real‐world
data.
facilitate
label‐free
training,
the
self‐supervised
loss
functions
exploiting
consistency
between
clouds
excavators.
Experimental
results
construction
sites
demonstrate
effectiveness
robustness
method,
achieving
an
average
accuracy
0.26
m.
method
also
exhibits
performance
across
various
operational
scenarios,
highlighting
its
potential
applications.
