ABSTRACT-Sensor systems can be used to assist elderly people
living alone at home. Sensor systems are able to perform different functions
ranging from monitoring of older adults to predicting their functional health
status.Movement tracking of elderly people can be performed by several
different types of sensor equipments. Sensor systems can
also be deployed in the homes of older adults living alone for functional
There are many people of accidental
deaths at home about 13 thousand every year, and about 80%, 10 thousand is
elderly people. In comparison, look up traffic death that is about 5 thousand
people,which means elderly people should pay attention to take bath and when an
accident occurs they would need some help.In such cases sensor systems can be
used to assist elderly people living alone at home. They can be used for
alarming or visualisation of information. Direct automated health assessments can
also be made based on sensor data.
1. Sensor systems for monitoring elderly
people living alone at home
The Doppler effects means phenomenon of
wave frequency is observed differently
by the presence of relative speed between source of wave and observer3.
Therefore we use the Doppler radar by observing the shift of the frequency due
to the Doppler effects to measure the moving speed of the observation target
only rather than position.
Install at a wall of high position of
dressing room toward a bathroom to detect fall down or drown in a bathtub by
momentum calculation of the Doppler sensor.
2) Living room
Install at a corner of high position of
a room covered whole spaces to detect fall down by momentum calculation of the
3) X-band Doppler Sensor Module
This specification covers the general
requirements for Xband microwave Doppler module. This module is designed for
motion sensing applications. It consists of DRO (Dielectric Resonator
Oscillator), balanced Schottky Barrier Diode mixer and Micro- strip Patch
2. Sensor systems for predicting
functional health status of elderly people living alone at home
this approach first a large feature set is created from the sensor data.
Subsequently, machine learning methods are used for the selection of the best
features and the best model. For feature selection and model selection, two
off-the-shell algorithms for regression are compared: linear regression
combined with correlation-based ranking and regression forests.
plan that indicate sensor placement in a house
1) SENSOR DATA
Each house is equipped
with a sensor system of approximately 16 sensors mostly passive infrared motion
sensors. Requirements are that the main areas in the house should be covered by
passive infrared motion sensors, making it possible to track a person through
the house. For detecting specific activities additional sensors can be
installed: passive infrared motion sensors for presence, a float sensor for
toilet’s flush and contact switches on doors and cabinets. When the behaviour
of the resident triggers the sensors, events are generated and stored remotely
as triples(label, timestamp, value),where label is the sensor id and value=0,1.
extraction method defines location related features that are calculated for
many time intervals throughout the day. For each time interval following
features are calculated:
Duration of stay in an area
Number of transitions to an area
Total number of transition between areas
Total agitation in the form of total number of sensor
Modelling the relation between the sensor data
and the health status is a regression problem as the features extracted from
the sensor data as the health metrics are on a continuous scale. Feature
extraction method will probably result in a redundancy in the features. Several
features will refer to the same concept. Instead of manually choosing relevant
concepts, a feature selection mechanism should be part of the solution. Because
the complexity of problem is unknown two off-shell regression methods are chosen
that differ in expressive power. a)Linear regression b)Regression forest
ordinary least squares) is used in combination with correlation based ranking
as a feature selection procedure1.
The goal of OLS
is to minimise the differences between the observed responses in some arbitrary
dataset and the responses predicted by the linear approximation of the data.
The equation for
For the linear regression model the train
data are used to rank the features based on their correlation with the health
Regression forest is chosen as a non-linear method. The advantage is
that they generalize well and have internal feature selection, which makes them
suitable for this problem. Forest has the advantage that it generalizes better
than a single tree, no explicit feature selection has to be done and the model
can capture non-linear relations in the data. The regression forest is an
ensemble method of many trees, where each regression tree is generated from a
sample drawn with replacement (bootstrap sample), and for a subset of features2.
A regression tree is a variant of a decision tree that is suited for
(nonlinear) regression problems instead of classi?cation problems. At each node
the data are split such that a simpler model (or weak learner) can handle the
data. Atypical objective function that should be minimized to ?nd the optimal
split is the mean squared error (MSE).For the regression forest before using
the training data to train a model,2-fold cross validation is used to optimize
the parameters, namely forest size and number of features for each tree. With
these parameters, a forest is trained and test data are used to calculate the
Mean Absolute Error(MAE).
Integrated sensor networks realize
giving reassurance for a family and elderly people as a whole service. The
features of this service are inexpensive even initial and running cost because
thinking about a way of data communication, and wide range cover in home that
several different types of sensor using depends on circumstances, and easy to
mounting and maintenance that sensors are solar cell so no need of changing
battery and no need of wiring by
wireless, and new construction work is not needed for dedicated communication
line of this service.
It is possible to predict health from
domestic sensor data, even with little assumptions on the features and with a
simple feature selection and modelling scheme. Although prediction of health
metrics is possible, changes in functional health — which are at least as
valuable to a caregiver — can be predicted with significantly better precision.
This opens up opportunities to better and faster detection of problems or
health degradation, and is likely to have a great impact on clinical practice.