12-10-2016, 04:49 PM
1458878087-068217012.pdf (Size: 420.65 KB / Downloads: 9)
Abstract—This paper proposes a hand posture recognition as
an assistive tool for elderly care. The system uses Haar-like
feature, Principle Component Analysis (PCA), and simulation
web server for detection, recognition, and alert by order. Our
system interprets the hand sign captured by web camera and
alerts a caregiver via smart phone. Besides, a caregiver can also
periodically check needs of an elder from a smart phone. So a
caregiver can do other tasks more efficiently because he/she need
not to be with an elder at all times. The empirical results show
satisfactory performance of PCA and contour detection with
hand posture recognition.
INTRODUCTION
From the World Health Organization (WHO) report[1], the
number of aging people is rising every year. Most of them need
caregivers especially the ones who cannot take care by
themselves. Many researchers have developed various assistive
technologies for lone senior people such as the Physical
Parameter Measurement (PPM) to measure blood pressure[2]
and activity recognition[3]. Systems in [4-5] can measure and
detect abnormal situation. We found that most of them were
developed for supporting on emergency cases. Only a few of
them facilitate needs of elders.
This paper presents a novel approach to recognize the needs
of elderly especially those elders who are speaking impaired
and mobility impaired. Normally, a caregiver must stay nearby
taking care of the elder at all times. So a caregiver cannot do
other works when he/she looks after more than one elder. We
contribute an assistive technology on elderly care notification
system that uses hand sign recognition for solving this
problem.
Hand detection and recognition have two general
approaches: one is a 3D hand model based approach and
another one is an appearance based approach. Hand is an
uncertain model because hand has 27 degrees of freedom. The
first model, the 3D-hand model based approach must use many
image databases because hands have so many shapes. So this
technique requires complicated computation of extracting
features and has unclear view management [6].
The second model, an appearance based approach
compares features extracted from video sequence images. The
advantage of this approach is a real time performance. The
appearance based can separate into data-glove based and vision
based approach. We adopt the vision based approach because
the data-glove based approach has bulky equipment and high
cost. The vision based system has a simple feature that finds
skin color region. The skin color technique is feasible and fast;
however, it has some drawbacks in light condition and requires
no other objects having a color-like skin at scene. Examples of
vision based technique on skin color approaches appeared in
[7-8].
Recently, popular methods were proposed with a local
invariant feature [9]. They were successful on hand posture
recognition and got high accuracy by using AdaBoost
algorithm and Scale-Invariant Feature Transform (SIFT). Viola
and Jones[10] created Haar-like feature since 2001. Haar-like
feature calculates the difference of sum intensity in black and
white area. Then, the machine learning such as AdaBoost was
used to select and combine weak classifiers for generating
strong classifiers. They achieved high performance on accuracy
and computation time. Many researchers have applied Haarlike
with other methods for better results [11-12]. Many
approaches applied hand detection and recognition to control
robot[13], translate sign language [14-15], and do some other
purposes[16-17]. However, our method expands their
techniques to recognize a demand from an elder, so that a
caregiver can check the needs from the elder via smart phone.
We use Haar-like feature, Principle Component Analysis
(PCA), and contour detection for detection and recognition
hand posture. Contour detection is used to improve the system
performance. We upload an outcome of the recognition event
on internet and implement an android application to read that
interpretation.
This paper is organized as follows: section two is a system
overview; section three explains our method; section four and
five give information about experimental setup and
experimental results; the last section provides conclusion of the
paper.
SYSTEM OVERVIEW
Our notification system can separate into two steps and four
block diagrams as shown in Fig.1.
A. Detection and Recognition Hand Posture
The system detects hand region from images of webcam
video sequence with Haar-like feature. After it gets hand
region, the system then applies contour detection before using
PCA to classify hand as the trained command identification.
B. Connection between System and Caregiver
We use XAMPP[18], a free program for web server
simulation, and eclipse it with add-ons for coding and testing
our android application[19]. Our android application is used to
read data from web and notice a caregiver. Because today’s
internet is very common and anyone can access by smart
phone. So we decide to implement our system on internet.
However, our experiment is tested on simulation web server; it
has not been tested in the registered domain name.
CONCLUSION
In this paper, we proposed hand posture recognition to
perceive the needs of elderly. We focused on an elder who has
difficulty movement and/or speaking problems. Thus we have
developed an assistive tool using hand posture for
communication between an elder and a caregiver. The elder
can pose hand on any camera because the camera used in this
research is a normal web camera type. The caregiver
effortlessly perceives the needs of elderly because he/she can
access the demand messages anytime from his/her phone. Thus
a caregiver can do other works while looks after elderly.
Experimental results show that contour detection technique
improves recognition rate than using Haar-like feature with
AdaBoost algorithm and PCA alone. Our system achieves realtime
performance of 87.45 percent and 70.29 percent accuracy
for the tested method and simulation situation, respectively.
However, an accuracy of our android application depends on
the system’s recognition rate. Future work will improve on the
robustness of our system in real situation such as illumination
changes and various distances between camera and an elder.