Improving Activity Monitoring Through a Hierarchical Approach

Xavier Rafael-Palou, Eloisa Vargiu, Guillem Serra and Felip Miralles

Barcelona Digital Technology Center, Barcelona, Spain

Keywords:

Activity Monitoring, Telemoniroting, Sensor-based approach, Ambient Assisted Living.

Abstract:

Performance of sensor-based telemonitoring and home support systems depends, among other characteristics,

on the reliability of the adopted sensors. Although binary sensors are quite used in the literature and also in

commercial solutions to identify user’s activities, they are prone to noise and errors. In this paper, we present

a hierarchical approach, based on machine learning techniques, aimed at reducing error from the sensors.

The proposed approach is aimed at improving the classiﬁcation accuracy in detecting if a user is at home,

away, alone or with some visits. It has been integrated in a sensor-based telemonitoring and home support

system. Results show an overall improvement of 15% in accuracy with respect to a rule-based approach. The

system is part of the BackHome project and is currently running in 2-healthy-users’ home in Barcelona and in

3-end-users’ home in Belfast.

1 INTRODUCTION

Activity monitoring is an increasingly important re-

search area due to the fact that it can be applied to

many real-life, human-centric problems, such as el-

dercare and healthcare. Recent studies have shown

that physical activities in daily life are an important

predictor of risk of hospital readmission and mortal-

ity in patients with chronic diseases (Yohannes et al.,

2002) (Pitta et al., 2005).

Monitoring users’ activities allows therapists,

caregivers, and relatives to become aware of user con-

text by acquiring heterogeneous data coming from

sensors and other sources. Moreover, activity mon-

itoring provides elaborated and smart knowledge to

clinicians, therapists, carers, families, and the pa-

tients themselves by inferring user habits and be-

haviour. Various methods of subjective and objec-

tive physical activity assessment tools have been de-

veloped. Subjective methods, such as diaries, ques-

tionnaires and surveys, are inexpensive tools. How-

ever, these methods often depend on individual ob-

servation and subjective interpretation, which make

the assessment results inconsistent (Meijer et al.,

1991). On the other hand, objective techniques use re-

mote monitoring techniques relying on sensors, such

as home-automation, wearable and/or environmental

ones (Warren, 2000).

A lot of telemonitoring systems have been pro-

posed in the literature (Carneiro et al., 2008) (Cor-

chado et al., 2010) (Mitchell et al., 2011) and are cur-

rently adopted in real environments (Scanaill et al.,

2006), enabling the healthcare provider to get feed-

back on monitored people and their health status pa-

rameters.

Sensor-based telemonitoring systems rely on a

conjunction of sensors, each one devoted to monitor a

speciﬁc status, a speciﬁc activity or activities related

to a speciﬁc location. Sensor technology can range

from vital signal devices –such as blood pressure

monitors, heart rate monitors and devices which can

measure body temperature– to sensors which can de-

tect presence in a room or detect a door being opened

(Nugent et al., 2008). Once all of the data have been

recorded it is then necessary for data processing to

take place to identify if the person requires a form

of assistance since an unusual activity has been rec-

ognized. Of course, this requires a certain degree of

intelligence which should take into consideration the

current state of the environment, the performed ac-

tivity and/or some physiological data (Cook and Das,

2007). Due to issues regarding personal privacy, tech-

nical installations, and costs of technology the most

adopted sensors are anonymous binary sensors (Nu-

gent et al., 2008). Binary sensors do not have the

ability to directly identify people and can only present

two possible values as outputs (“0” and “1”). Typical

examples of binary sensors deployed within smart en-

vironments include pressure mats, door sensors, and

movement detectors. A number of studies reporting

159

Rafael-Palou X., Vargiu E., Serra G. and Miralles F..

Improving Activity Monitoring Through a Hierarchical Approach.

DOI: 10.5220/0005437701590168

In Proceedings of the 1st International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AgeingWell-

2015), pages 159-168

ISBN: 978-989-758-102-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

the use of binary and related sensors have been under-

taken for the purposes of activity recognition (Tapia

et al., 2004). Nevertheless, sensor data can be con-

sidered to be highly dynamic and prone to noise and

errors (Ranganathan et al., 2004).

The increasing social demand for intelligent tele-

monitoring systems makes necessary to put more em-

phasis in activity recognition methods that deal with

environments prone to errors (Jafari et al., 2005). In

this paper, we make a step forward in this direction

by introducing a novelty and effective discriminative

method based on machine learning. The goal is to dis-

cover the very initial but crucial information regard-

ing the user location and recognition whether s/he is

alone or with visits under complex, noisy and unstable

environments. The method has been evaluated at two

healthy user homes in Barcelona where it is currently

running. Moreover, the system has been installed and

it is working in 3-end-users’ home in Belfast. Evalua-

tion is performed by using a novelty technique to gain

further user objectivity based on the information col-

lected from an activity tracking mobile application.

The rest of the paper, is organized as follows. Sec-

tion 2 illustrates the implemented sensor-based tele-

monitoring and home support system. In Section 3,

we present the adopted approach aimed at improving

habit recognition whereas in Section 4, we summarize

our preliminary results that show the improvement in

adopting the proposed solution. In Section 5, we re-

call relevant work related to the partial detection of

some of the activities that concern us and with set-

tings similar to those presented in this work. Section

6 ends the paper with conclusions and future work.

2 THE SENSOR-BASED

TELEMONITORING AND

HOME SUPPORT SYSTEM

To monitor users’ activities, we develop a sensor-

based telemonitoring and home support system (SB-

TMHSS) able to monitor the evolution of the user’s

daily life activity. The implemented system is able to

monitor indoor activities by relying on a set of home

automation sensors and outdoor activities by using

Moves

. Information gathered by the SB-TMHSS is

also used to provide context-awareness by relying on

ambient intelligence (Casals et al., 2014). Monitoring

users’ activities through the SB-TMHSS gives us also

the possibility to automatically assess quality of life

of people (Vargiu et al., 2014). In this Section, we

http://www.moves-app.com/

brieﬂy describe the SB-TMHSS, the interested reader

may refer to (Miralles et al., 2014) for further details.

The high-level architecture of the SB-TMHSS is

depicted in Figure 1. As shown, its main components

are: home; healthcare center; middleware; and intel-

ligent monitoring system.

Figure 1: Main components of the SB-TMHSS.

At home, a set of sensors are installed. In particu-

lar, we use presence sensors (i.e., Everspring SP103),

to identify the room where the user is located (one

sensor for each monitored room); a door sensor (i.e.,

Vision ZD 2012), to detect when the user enters or ex-

its the premises; electrical power meters and switches,

to control leisure activities (e.g., television and pc);

and pressure mats (i.e., bed and seat sensors) to mea-

sure the time spent in bed (wheelchair). The system

is also composed of a network of environmental sen-

sors that measures and monitors environmental vari-

ables like temperature, but also potentially danger-

ous events like gas leak, ﬁre, CO escape and pres-

ence of intruders. All the adopted sensors are wire-

less z-wave

. They send the retrieved data to a col-

lector (based on Raspberry pi

). The Raspberry pi

collects all the retrieved data and securely redirects

them to the cloud where they will be stored, pro-

cessed, mined, and analyzed. The proposed solution

relies on z-wave technology for its efﬁciency, porta-

bility, interoperability, and commercial availability.

In fact, on the contrary of other wireless solutions

(e.g., ZigBee), z-wave sensors are able to communi-

cate with any z-wave device. Moreover, we adopt a

solution based on Raspberry pi because it is easy-to-

use, cheap, and scalable. We are also using the user’s

smartphone as a sensor by relying on Moves, an app

for smartphones able to recognize physical activities

http://www.z-wave.com/

http://www.raspberrypi.org/

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

160

and movements by transportation. Among the activ-

ity trackers currently on the market, we select Moves

because it does not need user intervention being al-

ways active in background. The user interacts with

the overall system through a suitable interface aware

of end-user needs and preferences.

The middleware, which acts as a SaaS, is com-

posed by a secure communication and authentication

module; API module to enable the collector transmit-

ting all the data from sensors to make them available

to the intelligent monitoring system; and further utili-

ties such as load balancing and concurrency.

In order to cope with the data necessities of the

actors of the system (i.e., therapists, caregivers, rela-

tives, and end-users themselves), an Intelligent Mon-

itoring system has been designed. It is aimed to con-

tinuously analyzing and mining the data through 4-

dimensions: detection of emergencies, activity recog-

nition, event notiﬁcations, and summary extraction.

In order to cope with these objectives, the Intelli-

gent Monitoring system is composed of the follow-

ing modules: PP, the pre-processing module to encode

the data for the analysis; ED, the emergency detection

module to notify, for instance, in case of smoke and

gas leakage; AR, the activity recognition module to

identify the location, position, activity- and sleeping-

status of the user; EN, the event notiﬁcation module to

inform when a new event has been detected; and SC,

the summary computation module to perform sum-

maries from the data.

The healthcare center receives notiﬁcations, sum-

maries, statistics, and general information belonging

to the users through a web application.

3 THE HIERARCHICAL

APPROACH

The SB-TMHSS described in the previous section is

aimed at recognizing activities and habits of a user

who lives alone. One of the requirements of the

implemented SB-TMHSS was to be cheap and no-

intrusive. In other words, we use the minimum num-

ber of sensors depending on the user’s home conﬁgu-

ration, avoiding camera or wearable sensors. In par-

ticular, we decided to not use a camera for privacy

reason and in accordance with the requirements com-

ing from the end-user of the proposed system. More-

over, the sensors are wireless and rely on wi-ﬁ con-

nection to send data to the collector. Let us note that

we decided to do not adopt a wired solution because

is more expansive and intrusive. These requirements

imply that we have to take into account with errors

and noise coming from this conﬁguration and to ﬁnd

a solution to avoid them. In fact, sensors are not 100%

reliable: sometimes they loose events or detect them

several times. When sensors remain with a low bat-

tery charge they get worse. Moreover, also the Rasp-

berry pi may loose some data or the connection with

Internet and/or with the sensors. Also the Internet

connection may stop working or loose data. Finally,

without using a camera or wearable sensors we are

not able to directly recognize if the user is alone or

if s/he has some visits. Although, as said, a wireless

solution is not 100% reliable, .

Figure 2: An example of the sliding window approach,

where M means “motion event” and D means “door event”.

In order to solve this kind of limitations with the

ﬁnal goal of improving the overall performance of

the SB-TMHSS, we propose an approach based on

machine learning techniques. In this initial solution,

we only consider motion and door sensors. The in-

telligent monitoring system continuously and concur-

rently listens for new data in a given window, accord-

ing to a sliding window approach (Datar et al., 2002).

For each window, data are pre-processed by the PP

and analyzed. As an example, let us consider the Fig-

ure 2 where once the current window recognizes a

door event at time tb, it looks for the previous one

in the window or before (in the example ta). Then,

the period from that door events (i.e, tb −ta) is clas-

siﬁed by the hierarchical classiﬁer. Seemly, when the

event tc has been recognized, the period from tb and

tc is classiﬁed. Finally, the period from tc to the end

of the window is classiﬁed. In case of no door events

have been recognized, the period from ta to the end

of the window is classiﬁed.

The hierarchical classiﬁer, depicted in Figure 3, is

composed of two levels. The upper is aimed at recog-

nizing if the user is at home or not, whereas the lower

is aimed at recognizing if the user is really alone or if

s/he received some visits.

The goal of the classiﬁer at the upper level is to

improve performance of the door sensor. In fact, it

may happen that the sensor registers a status change

(from closed to open) even if the door has not been

opened. This implies that the SB-TMHSS may reg-

ister that the user is away and, in the meanwhile, ac-

tivities are detected at user’s home. On the contrary,

the SB-TMHSS may register that the user is at home

and, in the meanwhile, activities are not detected at

user’s home. To solve, or at least reduce, this prob-

lem, we built a supervised classiﬁer able to recognize

ImprovingActivityMonitoringThroughaHierarchicalApproach

161

Figure 3: The hierarchical approach in the activity recogni-

tion module.

if the door sensor is working well or erroneous events

have been detected. First, we revise the data gathered

by the SB-TMHSS searching for anomalies, i.e.: (1)

the user is away and at home some events are detected

and (2) the user is at home and no events are detected.

Then, we validated those data by relying on Moves,

installed and running on the user smartphone. In fact,

Moves, among other functionality, is able to localize

the user. Hence, using Moves as an “oracle” we build

a dataset in which each entry is labeled depending on

the fact that the door sensor was right (label “1”) or

wrong (label “0”).

The goal of the classiﬁer at the lower level is to

identify whether the user is alone or not. The input

data of this classiﬁer are those that has been ﬁltered

by the upper level, being recognized as positives. To

build this classiﬁer, we rely on the novelty detection

approach (Markou and Singh, 2003) used when data

has few positive cases (i.e., anomalies) compared with

the negatives (i.e., regular cases); in case of skewed

data. Let us recall here that novelty detection is the

identiﬁcation of new or unknown data that a machine

learning system has not been trained with and was not

previously aware of. In particular, we rely on the ap-

proach presented in (Sch

olkopf et al., 2001) that tries

to estimate a function f that is positive on the dataset

and negative on the complement. The functional form

of f is given by a kernel expansion in terms of a poten-

tially small subset of the training data; it is regularized

by controlling the length of the weight vector in an as-

sociated feature space. The expansion coefﬁcients are

found by solving a quadratic programming problem,

which we do by carrying out sequential optimization

over pairs of input patterns.

4 PRELIMINARY

EXPERIMENTAL RESULTS

The SB-TMHSS presented in this paper is part of

BackHome

, an European R&D project that aims to

provide telemonitoring and home support using Brain

Computer Interfaces (BCI) and other assistive tech-

nologies to improve autonomy and quality of life of

disabled people (Vargiu et al., 2012). The system is

currently running in ﬁve homes, two in Barcelona and

three in Belfast. In Barcelona installations have been

made at healthy-users’ home, whereas in Belfast at

BackHome end-user’s home (Edlinger et al., 2015).

To train and test the proposed approach, we con-

sider a window of 4 months for training and evalua-

tion (training dataset) and a window of 1 month for

the test (testing dataset). Experiments have been per-

formed at each level of the hierarchy. First, we per-

formed experiments to identify the best supervised

classiﬁer to be used at the upper level of the hierar-

chy. Subsequently, we applied the novelty detection

algorithm on the data ﬁltered by the classiﬁer at the

upper level, to validate the classiﬁer at the lower one.

Finally, we measure the performance of the overall

approach.

4.1 Is the User at Home?

First of all we build the training dataset with door

events (gathered by the door sensor) in a window of

4 months. Those data have been then validated by re-

lying on the information coming from Moves. The

entries are manually labeled in two classes Correct

data and User not at home according to the following

criteria:

• Correct data: 0 if the data gathered from the door

sensor differs from the data gathered from Moves;

1 otherwise.

• User not at home: 0 if the user is at home; 1 oth-

erwise.

http://www.backhome-fp7.eu/backhome/index.php

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

162

Table 1: Results for the high-level classiﬁer during the training (T) and evaluation (E) phase.

Classiﬁer Parameter(s) Accuracy (T) Accuracy (E)

LR C = 0.005 0.945 ± 0.09 0.885

SVM γ = 0.01, C = 1.0 0.945 ± 0.09 0.885

SVM γ = 1.0, C = 0.452 0.853 ± 0.12 0.943

SVM γ = 0.05, C = 0.452 0.930 ±0.11 0.885

SVM γ = 0.01, C = 0.257 0.945 ±0.09 0.885

RF n estimators = 5 0.943 ± 0.09 0.942

RF max f eatures = 12 0.930 ± 0.09 0.942

AB n estimators = 15 0.918 ± 0.10 0.823

First, we deﬁne and implement a rule-based sys-

tem to verify if an approach based on rules may

help in improving the overall performance. The re-

sults coming from the rule-based system have been

then compared with those manually validated using

Moves. Unfortunately, results show a very few im-

provement with an accuracy of 77%. Thus, we decide

to implement a supervised classiﬁer.

The data labeled as 1 for the class Correct data

have been used to extract the following features:

• number of motion events divided by their dura-

tion, calculate after a door event (from t = i door

event to t = i + 1 door event) [Feature 1];

• number of motion events divided by their dura-

tion, calculated before a door event (from t = i−1

door event to t = i door event) [Feature 2];

• number of motion events happened during minute

before a door event (t = i) [Feature 3];

• number of motion events happened during the 2

minutes before a door event (t = i) [Feature 4];

• number of motion events happened during the 5

minutes before a door event (t = i) [Feature 5];

• number of motion events happened during minute

after a door event (t = i + 1) [Feature 6];

• number of motion events happened during the 2

minutes after a door event (t = i + 1) [Feature 7];

• number of motion events happened during the 5

minutes after a door event (t = i + 1) [Feature 8].

The dataset is then used to train four well-

representative and successful families of supervised

classiﬁers (Fern

andez-Delgado et al., 2014): a Lo-

gistic Regression (LR) classiﬁer, a Support Vector

Machine (SVM), a Random-Forest (RF) and an Ad-

aBoost (AB).

The dataset has been divided in a training and

in an evaluation set and a 10-fold cross-validation

method has been used. The classiﬁers have been then

tested with an independent dataset. Table 1 shows re-

sults during the training phase and evaluation phase.

As shown, the best performance has been obtained by

relying on the SVM (with γ = 1.0 and C = 0.452), see

Figure 4, which shows different combination of the

parameters kernel coefﬁcient (γ) and penalty parame-

ter of the error term (C).

The best classiﬁer has been then used with the

testing dataset and, on average, it obtained a F

0.97 and an accuracy of 0.968. Finally, it showed an

improvement of 20% with respect to the rule-based

approach.

4.2 Is the User Alone?

The data ﬁltered by the classiﬁer at the upper level,

belonging to the monitored window of 4 months, are

the training dataset of the classiﬁer at the lower level.

The corresponding dataset is composed by 57 normal

instances (i.e., the user was alone) and 8 anomalies

(i.e., the user received visits).

First of all, also in this case, we deﬁned and imple-

mented a rule-based classiﬁer to verify if a rule-based

approach could solve this problem. The adopted rules

together with the number of anomalies detected by

each one are the following:

• Number of movement events every 60 seconds

greater than 6: 23 anomalies detected;

• Number of movement events every 60 seconds

greater than 10: 1 anomaly detected;

• Simultaneous movement events: 20 anomalies de-

tected;

• Simultaneous movement events in less 2 seconds:

23 anomalies detected;

• Maximum number of movement events in 60 sec-

onds greater than 6 or simultaneous moves in less

2 seconds: 28 anomalies detected;

• Maximum number of movement events in 60 sec-

onds greater than 6 or simultaneous moves: 21

anomalies detected.

Since the rule-based approach is not able to cor-

rectly recognized anomalies, we use, also in this case,

an SVM classiﬁer (one-class SVM with RBF, non lin-

ear). The following features have been considered:

ImprovingActivityMonitoringThroughaHierarchicalApproach

163

Figure 4: Cross-validation results of the different setting of parameters in the SVM classiﬁer.

• maximum number of motion events in intervals of

60 seconds [Feature 1];

• maximum number of motion events in intervals of

120 seconds [Feature 2];

• maximum number of motion events in intervals of

180 seconds [Feature 3];

• number of motion events happened in a range of

5 seconds [Feature 4];

• number of motion events happened in a range of

2 seconds [Feature 5];

• number of motion events happened simultane-

ously [Feature 6];

• minimum number of seconds between two con-

secutive motion events [Feature 7];

• average of seconds between two consecutive mo-

tion events [Feature 8].

The classiﬁer has been trained by considering the

normal instances and then evaluated introducing the

anomalies. Figure 5 shows the results obtained con-

sidering two features at time. In particular, for each

pair of features the frontier, the training observations

and the observations in case of normal instances (reg-

ular) or anomalies (abnormal) have been shown. Ta-

ble 2 shows the overall results .

Results during the evaluation phase show that the

system is able to correctly recognize all the anoma-

lies. According to the obtained results, we select the

classiﬁer with the regularization parameter (ν) = 0.01

and γ = 0.1.

Similarly to the classiﬁer at the upper level, the

system has been tested with the data coming from

the 1-month window of monitored events. Results

showed an average accuracy of 0.94.

4.3 Overall Results

Finally, we tested the performance of the overall hi-

erarchical approach. Once both classiﬁers have been

Table 2: Results for the classiﬁer at the lower level. The

table reports the classiﬁcation error calculated as the ratio

between the number of detected anomalies and the number

of instances in the dataset.

Parameter(s) Error (T)

ν = 0.01, γ = 0.1 0.0701

ν = 0.01, γ = 0.5 0.0877

ν = 0.01, γ = 1 0.1578

ν = 0.05, γ = 0.1 0.0877

ν = 0.05, γ = 0.5 0.0877

ν = 0.05, γ = 1 0.1403

ν = 0.1, γ = 0.1 0.1052

ν = 0.1, γ = 0.5 0.1052

ν = 0.1, γ = 1 0.1403

ν = 0.5, γ = 0.1 0.4912

ν = 0.5, γ = 0.5 0.5087

ν = 0.5, γ = 1 0.4912

trained, we tested the performance of the overall ap-

proach with the testing dataset corresponding to a

window of 1 month. We compared the overall results

with those obtained by using the rule-based approach

in both levels of the hierarchy. Results are shown

in Table 3 and point out that the proposed approach

outperforms the rule-based one with a signiﬁcant im-

provement.

Table 3: Results of the overall hierarchical approach with

respect to the rule-based one.

Metric Rule-based Hierarchical Improv.

Accuracy 0.80 0.95 15%

Precision 0.68 0.94 26%

Recall 0.71 0.91 20%

0.69 0.92 23%

To highlight the performance of the proposed ap-

proach, let us consider the Figure 6 that shows a com-

parison between the real data, labeled during the val-

idation phase (on the left), and the data classiﬁed by

relying to the approach proposed in this paper (on the

right).

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

164

Figure 5: Feature analysis of select novelty outlier method.

Figure 6: Comparison between real labeled data and data classiﬁed by the hierarchical approach.

Thanks to the proposed approach, the system is

able to recognize if the user is at home, away and/or

if s/he had some visits. An example of results in rec-

ognizing if the user is at home or away and if s/he

received some visits is given in Figure 7.

Figure 7: Example of results in recognizing if the user is at

home or away or if s/he received some visits.

5 RELATED WORK

There is a large literature on recognition of activities

at home (Van Kasteren et al., 2008) (Ye et al., 2012).

At the same time, we ﬁnd a great variability in the

settings of the experiments either in the number of

sensors and their type, individuals involved or the du-

ration thereof. Also noteworthy is the large amount of

recognition techniques (supervised, either generative

or discriminative; and unsupervised). This diversity

makes it extremely difﬁcult to compare performances

and draw conclusive ﬁndings from those studies. De-

spite this high variability, it is noteworthy to mention

that we have not found extensive studies that analyze

altogether the detection of visits, presence or absence

ImprovingActivityMonitoringThroughaHierarchicalApproach

165

of users at home using wireless binary sensors. Even

so; we report a number of different papers with differ-

ent approaches related to the partial detection of some

of the activities that concern us and with settings sim-

ilar to those in our work.

A former study (Tapia et al., 2004) already points

out some of the difﬁculties in discriminating daily

life activities based only on binary sensors activi-

ties. The automatic recognition system was based

on rules deﬁned from the context and the duration

of the activities to identify. The data of the study

were obtained from 14 days of monitoring activities at

home. Although promising accuracies were achieved

for some activities, detection tasks such as ”leaving

home” were nothing less than satisfactory with 0.2 of

accuracy. This was because the activities were rep-

resented by rules directly deﬁned on the ﬁrings out-

putted by single sensors (i.e. door switches); so they

did not contemplate that could be activated for other

reasons and in varying times, which made reduce their

discriminating power.

A more exhaustive work regarding the use of

switch and motion sensors for tracking people inside

home is found in (Wilson and Atkeson, 2005). Tests

were done with up to three simultaneous users. High

performances were reported by the trained tracking

models. However it is interesting to note that this

type of sensors experimented occasional lag between

“entering” a room and triggering a sensor; making to

decrease the performance of the tracking models.

In (Krishnan and Cook, 2014) a more complex

template learning model (SVM) was used to auto-

matically recognize among 11 different home activ-

ities. The proposed technique was integrated in dif-

ferent window sliding strategies (e.g. weighting sen-

sor events, dynamic window lengths, or two levels of

window lengths). They used 6 months of data from

3 different homes in which activities such as “enter-

ing” or “leaving home” were monitored. From the

best experimental settings the authors claimed accu-

racy for “entering” home about 0.80 of f1-score but

around 0.4 for “leaving” home tasks.

In a more extensive work (Cook, 2010) they use

ıve Bayes (NB), Hidden Markov (HMM) models

and Conditional Random Fields (CRF) for the ac-

tivity recognition problem. In this study, 7 smart

environments were used and 11 different data sets

were obtained. Several activities were attempted to

be recognized. Among others, we highlight “enter-

ing” and “leaving home” as relevant for our approach.

Although they did not report speciﬁc accuracies for

these activities, authors claimed an overall recogni-

tion performance on the combined dataset of 0.74 for

the NB classiﬁer, 0.75 for the HMM model, and 0.72

for the CRF using 3-fold cross validation over the set

of annotated activities.

In (Ord

onez et al., 2013), authors proposed a hy-

brid approach to recognize ADLs from home envi-

ronments using a network of binary sensors. Among

the different activities recognized “leaving” was one

of them. The hybrid system proposed was composed

by using an SVM to estimate the emission proba-

bilities of an HMM. The results showed how the

combination of discriminative and generative models

is more accurate than either of the models on their

own. Among the different schemes evaluated, the

SVM/HMM hybrid approach obtains a signiﬁcant 0.7

of f1-score a notable better performance than the rest

of approaches.

Detecting “multiple” people in single room by us-

ing binary sensors was already studied in an early

work (Wilson and Atkeson, 2004). In that work, au-

thors proposed a method based in Expectation Maxi-

mization Montecarlo algorithm. In a more recent ar-

ticle (Nait Aicha et al., 2013), high accuracy (0.85)

were reported on detecting visits at home using bi-

nary sensors. In that approach, they used an HMM al-

gorithm over the room events although not all rooms

of the home were monitored.

6 CONCLUSIONS AND FUTURE

WORK

Community based living, often alone with intermit-

tent care, creates possible scenarios of risks for all in-

dividuals. When cognitive changes are likely to have

taken place it is crucial to understand what the risks

may be and monitor these. To monitor users at home,

we develop a sensor-based system which is able to

gather data and report on the stability and evolution

of the user’s daily life activity. Unfortunately, per-

formance of sensor-based telemonitoring and home

support systems depends, among other issues, on the

reliability of the adopted sensors. It is particularly

true in the case of a wireless and binary sensors are

adopted. To solve this problem, we presented a hier-

archical approach, based on machine learning tech-

niques, aimed at improving the recognition of the

presence of the user at home and, being interested in

monitoring people that live alone, if the user is alone

or received some visits. The system has been devel-

oped under the umbrella of the BackHome project and

is currently running in 2-healthy-users’ home and in

3-end-users’ home. Results clearly show an high dis-

criminative performance improvement of 15% with

respect to a rule-based solution. Let us also stress the

fact that the proposed approach can be used as a pre-

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

166

processing phase in every activity recognition system,

being completely independent from that.

As for the future work, we are currently setting-up

new experiments aimed at comparing the hierarchical

approach with a multi-class classiﬁer and with an en-

semble (not hierarchical) of classiﬁers. Moreover, we

are improving the approach in order to be totally auto-

matic, by using data from Moves as feature instead of

to validate the initial dataset. We are also interested in

studying if we can generalize the proposed approach

to adopt it for all the user of the system, or if we have

to use a personalized approach for each different user

(or a small group of them).

ACKNOWLEDGEMENTS

The research leading to these results has received

funding from the European Community’s, Seventh

Framework Programme FP7/2007-2013, BackHome

project grant agreement n. 288566.

REFERENCES

Carneiro, D., Costa, R., Novais, P., Machado, J., and Neves,

J. (2008). Simulating and monitoring ambient assisted

living. In Proc. ESM.

Casals, E., Cordero, J. A., Dauwalder, S., Fern

andez, J. M.,

Sol

a, M., Vargiu, E., and Miralles, F. (2014). Ambient

intelligence by atml: Rules in backhome. In Emerging

ideas on Information Filtering and Retrieval. DART

2013: Revised and Invited Papers; C. Lai, A. Giuliani

and G. Semeraro (eds.).

Cook, D. J. (2010). Learning setting-generalized activity

models for smart spaces. IEEE intelligent systems,

2010(99):1.

Cook, D. J. and Das, S. K. (2007). How smart are our en-

vironments? an updated look at the state of the art.

Pervasive and mobile computing, 3(2):53–73.

Corchado, J., Bajo, J., Tapia, D., and Abraham, A. (2010).

Using heterogeneous wireless sensor networks in a

telemonitoring system for healthcare. IEEE Trans-

actions on Information Technology in Biomedicine,

14(2):234–240.

Datar, M., Gionis, A., Indyk, P., and Motwani, R. (2002).

Maintaining stream statistics over sliding windows.

SIAM Journal on Computing, 31(6):1794–1813.

Edlinger, G., Hintermller, C., Halder, S., Vargiu, E., Mi-

ralles, F., Lowish, H., Anderson, N., Martin, S., and

Daly, J. (2015). Brain neural computer interface for

everyday home usage. In HCI International 2015.

Fern

andez-Delgado, M., Cernadas, E., Barro, S., and

Amorim, D. (2014). Do we need hundreds of classi-

ﬁers to solve real world classiﬁcation problems? The

Journal of Machine Learning Research, 15(1):3133–

3181.

Jafari, R., Encarnacao, A., Zahoory, A., Dabiri, F., Noshadi,

H., and Sarrafzadeh, M. (2005). Wireless sensor net-

works for health monitoring. In Mobile and Ubiqui-

tous Systems: Networking and Services, 2005. Mo-

biQuitous 2005. The Second Annual International

Conference on, pages 479–481. IEEE.

Krishnan, N. C. and Cook, D. J. (2014). Activity recogni-

tion on streaming sensor data. Pervasive and Mobile

Computing, 10:138–154.

Markou, M. and Singh, S. (2003). Novelty detection: a re-

view?part 1: statistical approaches. Signal processing,

83(12):2481–2497.

Meijer, G. A., Westerterp, K. R., Verhoeven, F. M., Koper,

H. B., and ten Hoor, F. (1991). Methods to assess

physical activity with special reference to motion sen-

sors and accelerometers. Biomedical Engineering,

IEEE Transactions on, 38(3):221–229.

Miralles, F., Vargiu, E., Dauwalder, S., Sol

a, M., Fern

andez,

J., Casals, E., and Cordero, J. (2014). Telemonitoring

and home support in backhome. In Proceedings of the

8th International Workshop on Information Filtering

and Retrieval co-located with XIII AI*IA Symposium

on Artiﬁcial Intelligence (AI*IA 2014).

Mitchell, M., Meyers, C., Wang, A., and Tyson, G.

(2011). Contextprovider: Context awareness for med-

ical monitoring applications. In Conf Proc IEEE Eng

Med Biol Soc.

Nait Aicha, A., Englebienne, G., and Kr

ose, B. (2013).

How lonely is your grandma?: detecting the visits to

assisted living elderly from wireless sensor network

data. In Proceedings of the 2013 ACM conference on

Pervasive and ubiquitous computing adjunct publica-

tion, pages 1285–1294. ACM.

Nugent, C. D., Hong, X., Hallberg, J., Finlay, D., and

Synnes, K. (2008). Assessing the impact of indi-

vidual sensor reliability within smart living environ-

ments. In Automation Science and Engineering, 2008.

CASE 2008. IEEE International Conference on, pages

685–690. IEEE.

Ord

onez, F. J., de Toledo, P., and Sanchis, A. (2013). Activ-

ity recognition using hybrid generative/discriminative

models on home environments using binary sensors.

Sensors, 13(5):5460–5477.

Pitta, F., Troosters, T., Spruit, M. A., Decramer, M., and

Gosselink, R. (2005). Activity monitoring for as-

sessment of physical activities in daily life in patients

with chronic obstructive pulmonary disease. Archives

of physical medicine and rehabilitation, 86(10):1979–

1985.

Ranganathan, A., Al-Muhtadi, J., and Campbell, R. H.

(2004). Reasoning about uncertain contexts in perva-

sive computing environments. IEEE Pervasive Com-

puting, 3(2):62–70.

Scanaill, C. N., Carew, S., Barralon, P., Noury, N., Lyons,

D., and Lyons, G. M. (2006). A review of approaches

to mobility telemonitoring of the elderly in their liv-

ing environment. Annals of Biomedical Engineering,

34(4):547–563.

Sch

olkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J.,

and Williamson, R. C. (2001). Estimating the support

ImprovingActivityMonitoringThroughaHierarchicalApproach

167

of a high-dimensional distribution. Neural computa-

tion, 13(7):1443–1471.

Tapia, E. M., Intille, S. S., and Larson, K. (2004). Activity

recognition in the home using simple and ubiquitous

sensors. Springer.

Van Kasteren, T., Noulas, A., Englebienne, G., and Kr

ose,

B. (2008). Accurate activity recognition in a home

setting. In Proceedings of the 10th international con-

ference on Ubiquitous computing, pages 1–9. ACM.

Vargiu, E., Fern

andez, J. M., and Miralles, F. (2014).

Context-aware based quality of life telemonitoring. In

Distributed Systems and Applications of Information

Filtering and Retrieval. DART 2012: Revised and In-

vited Papers. C. Lai, A. Giuliani and G. Semeraro

(eds.).

Vargiu, E., Miralles, F., Martin, S., and Markey, D. (2012).

BackHome: Assisting and telemonitoring people with

disabilities. In RAatE 2012 - Recent Advances in As-

sistive Technology & Engineering.

Warren, S. (2000). Wearable and wireless: Distributed,

sensor-based telemonitoring systems for state of

health. Canadian Journal of Animal Science, 80:381–

392.

Wilson, D. and Atkeson, C. (2004). Automatic health mon-

itoring using anonymous, binary sensors. In CHI

Workshop on Keeping Elders Connected, pages 1719–

1720. Citeseer.

Wilson, D. H. and Atkeson, C. (2005). Simultaneous

tracking and activity recognition (STAR) using many

anonymous, binary sensors. In Pervasive computing,

pages 62–79. Springer.

Ye, J., Dobson, S., and McKeever, S. (2012). Situation

identiﬁcation techniques in pervasive computing: A

review. Pervasive and mobile computing, 8(1):36–66.

Yohannes, A. M., Baldwin, R. C., and Connolly, M. (2002).

Mortality predictors in disabling chronic obstructive

pulmonary disease in old age. Age and ageing,

31(2):137–140.

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

168