Abstract
With the development of deep learning, how to apply it to improve people's lifestyle is
becoming a research focus. In real carriages, it is usually crowded, which makes occasions
like passenger’s luggage being taken mistakenly or stolen happen frequently,
besides,
catching each
behavior is
difficult relying on
human
resources
only, therefore,
taking the
behavior of picking up
or
taking away luggages
as the reaserch perspective,
it is
of great
significance and practical value to study the action recognition algorithm for travel scenarios.
This paper
collects and tags datas which simulate the real carriages, and designs and
trains a
series of high accuracy and real-time
models on
these
datas. Specifically, for
the
human-object interaction detection algorithm PPDM, its pre-defined interaction point is not
suitable for detecting luggage’s being picked up and taken away behavior, thus in this paper,
the interaction point’s location is modified to suit the task, besides, a pose branch is added
to
enhance the new network’s detection ability of the interaction point
inspired by
other
methods. Furthermore, to
improve
the
performance
of human-object
interaction
detection
on
luggage’s being picked up and taken away,
this paper extracts the core ways that
dominate
the
performance in
human-object
interaction,
that is, object detection and pose
estimation, and explores the factors that influence the performance such as tagging choices,
data augmentation in object detection and manual
feature design, machine
feature learning
in pose estimation
separately. Finally,
how
model fusion
influences
the final performance
will be explored in this paper.
The experimental results show that specialized modification for interaction point’s pre-
defined location and the extra pose branch improve the precision and recall of classification
compared to the original method;
object detection
enhanced by data
augment and pose
information processed by multi-layer perceptron
are all
able
to
archive the
goal
that both
precision and recall exceed 90%. On this basis, model fusion combining the output ofobject
detection and
pose estimation reaches best recognition accuracy compared with other
methods in this paper, while it just costs a little bit more inference time.
Key words: deep learning,
picking
up
and taking away luggage, action
recognition,
model fusion
II
VIP享7折，此内容立减5.97元 
