Page 125 - ITU KALEIDOSCOPE, ATLANTA 2019

ELDERLY HEALTH MONITORING SYSTEM WITH FALL DETECTION USING MULTI-
FEATURE BASED PERSON TRACKING

1
Dhananjay Kumar , Aswin Kumar Ravikumar , Vivekanandan Dharmalingam , Ved P. Kafle
2
1
1
1 Department of Information Technology, Anna University, MIT Campus, Chennai, India
2 National Institute of Information and Communications Technology, Nukui-Kitamachi, Koganei, Tokyo, Japan

ABSTRACT of both. The sensor-oriented surveillance systems generally
utilize accelerometer and GPS sensors to locate the person
The need for personalized surveillance systems for elderly [3]. Although, these sensors provide highly accurate real-
health care has risen drastically. However, recent methods world coordinates, there exists a possibility of sensors being
involving the usage of wearable devices for activity misplaced, not worn by the user, or worn by the wrong user,
monitoring offer limited solutions. To address this issue, we thus restricting the tracking ability of the system. Although
have proposed a system that incorporates a vision-based other alternative methods employing devices like thermal
deep learning solution for elderly surveillance. This system sensors have been proposed to work, they work only within
primarily consists of a novel multi-feature-based person a short range [4-5]. On the whole, sensor-based tracking
tracker (MFPT), supported by an efficient vision-based techniques heavily rely on the assumption that users
person fall detector (VPFD). The MFPT encompasses a continuously wear the devices. In general, sensor-based fall
combination of appearance and motion similarity in order to identification involves the use of triaxial accelerometer
perform effective target association for object tracking. The sensor [6] which records real-world 3D coordinates. The
similarity computations are carried out through Siamese continuous analysis of coordinate information also poses
convolutional neural networks (CNNs) and long-short term difficulty in differentiating between daily activities such as
memory (LSTM). The VPFD employs histogram-of-oriented- sleeping, sitting and standing, implying the need of more
gradients (HoGs) for feature extraction, followed by the sophisticated and accurate systems based on artificial
LSTM network for fall classification. The cloud-based intelligence techniques such as deep learning. Vision-based
storage and retrieval of objects is employed allowing the two fall detection using optical flow and convolutional neural
models to work in a distributed manner. The proposed networks (CNNs) [7] can be used to extract temporal features
system meets the objectives of ITU Focus Group on AI for needed for improving system performance. However,
Health (FG-AI4H) under the category, “falls among the existing customized techniques such as curvelets [8] do not
elderly”. The system also complies with ITU-T F.743.1 extract deep features for human representation to detect falls.
standard, and it has been evaluated over benchmarked
object tracking and fall detection datasets. The evaluation Visual object tracking can be categorized into two broad
results show that our system achieves the tracking precision categories namely detection-based tracking and detection-
of 94.67% and the accuracy of 98.01% in fall detection, free tracking. The detection-based tracking consists of three
making it practical for health care system use. The HoG main components: moving object detection, object
feature-based LSTM model is a promising item to be classification and localization, and object tracking [9]. The
standardized in ITU for fall detection in elderly healthcare moving object detection component identifies the salient
management under the requirements and service description objects that are present in the current frame using bounding
provided by ITU-T F.743.1. boxes. Object classification is carried out in identifying the
detected objects and segregating into specific classes, while
Keywords – CNN, fall detection, HoG, LSTM, object the tracking is performed for target association in subsequent
tracking frames. On the other hand, detection-free tracking does not
involve recognition of different objects, rather it utilizes
1. INTRODUCTION motion features in order to locate moving objects. Typical
detection-free tracking involves the usage of optical flow,
According to the United Nations (UN) report on ageing background subtraction in order to eliminate static objects in
world population [1], the population of elderly people will each frame. Tracking based on background subtraction in
rise to 2 billion by the year 2050. The guidelines for video usually requires manual intervention to identify scene-
Integrated Care for Older People (ICOPE) [2] released by the specific objects [10]. An optical flow-based tracking
World Health Organization (WHO) clearly indicates that algorithm also requires additional support from appearance
accidental fall is one of the common reasons for the decline modelling in order to produce accurate results. Optical flow
in the health of the elderly. The surveillance system for along with blob analysis yields better traffic surveillance
effectively monitoring the elderly’s health can be achieved systems [11]. However, these algorithms are only capable of
by either a sensor or vision-based system, or a combination

978-92-61-28401-5/CFP1968P-ART @ ITU 2019 – 105 – Kaleidoscope