ACHIEVING MODEL COMPLETENESS FOR HIERARCHALLY

STRUCTURED ACTIVITIES OF DAILY LIFE

Usman Naeem, Abdel-Rahman H. Tawil, Rabih Bashroush and Ameer Al-Nemrat

School of Architecture, Computing & Engineering, University of East London

Docklands Campus, University Way, London, U.K.

Keywords: Hierarchal Activities of Daily Life, Alzheimer’s Disease, Task Sequences, Decision Trees, ID3 Algorithm,

Object Usage.

Abstract: Being able to recognise everyday activities of daily life provides the opportunity of tracking functional

decline among elderly people who suffer from Alzheimer’s disease. This paper describes an approach that

has been developed for recognising activities of daily life based on a hierarchal structure of plans. While it

is logical to envisage that the most common activities will be modelled within a library of plans, it can be

impossible to imagine that the library contains plans for every possible hierarchal activity. In order to

generalise the activity recognition capability outside the framework of the core activities constructed to

support recognition, decision trees are constructed using a well-known induction algorithm during a train

period. The motivation of this work is to allow people with Alzheimer’s disease to have additional years of

independent living before the disease reaches a stage where it becomes incurable.

1 INTRODUCTION

Alzheimer’s disease is a progressive disease that

gradually destroys an elderly person’s memory and

their capability to learn, communicate and carry out

everyday activities. Managing people with this

disease incurs high costs for the government, as well

as the people associated with person who has the

disease. The total cost of dementia for the UK in

2006 was an estimated £17 billion, which then

escalated to an approximate £23 billion in 2010

(Alzheimer’s Research Trust, 2010).

In order to provide any form of assistance or to

find out if the elderly person is safe, it is important

to recognise what Activity of Daily Life (ADL) they

are carrying out. Depending on the memory

condition of an elderly people with Alzheimer’s

disease, their brain sometimes does not permit them

to remember what activity they were carrying out.

Usually in these cases, the sufferers are often

prescribed a set of daily activities by visiting carers

in order to deal with forgetfulness as well as giving

the elderly stimulation and a framework for an

independent life (The Alzheimer’s Association,

2005). Nevertheless, there can be still many

instances where the elderly person can forget what

activity they were conducting, which can lead to

anxiety (Feretti et al., 2001) and frustration as they

become aware that they are slowly losing their

independence. Hence, the recognition of activities

not only provides useful information about what

activity the sufferer is carrying out, but it also has

the capability of providing information about what

activity the sufferer is meant to be doing next and

provide assistance accordingly.

This paper describes a hierarchal approach that

has been developed for carrying ADL recognition,

which utilises more knowledge about the structure of

ADLs rather than solely relying on data gathered

from the extensive monitoring.

2 RELATED WORK

Activity recognition in the home can be conducted

in many ways, however the work in this paper

focuses on carrying out activity recognition with

object usage data, as opposed to data generated by

visual based systems. In order to make this possible,

a popular technique has been adopted, which is

known as ‘Dense Sensing’ (Buettner et al., 2009);

(Philipose et al., 2004). This is based around

numerous individual objects such as toasters and

kettles being tagged with wireless battery-free

Naeem U., H. Tawil A., Bashroush R. and Al-Nemrat A..

ACHIEVING MODEL COMPLETENESS FOR HIERARCHALLY STRUCTURED ACTIVITIES OF DAILY LIFE.

DOI: 10.5220/0003802200510057

In Proceedings of the 2nd International Conference on Pervasive Embedded Computing and Communication Systems (PECCS-2012), pages 51-57

ISBN: 978-989-8565-00-6

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

transponders that transmit information to a computer

via an Radio Frequency Identification (RFID) reader

(Kalimeri et al., 2010); (Philipose et al., 2005) when

the object is used or touched. Wearable sensors such

as accelerometers can be seen as more intrusive then

RFID tags, however they are very practical for

capturing data that is concerned with human body

movements, as they provide accurate recognition of

movement (Wang et al., 2007).

Many computational models have been

constructed for recognising activities, typical

examples include Hidden Markov Models (HMM)

and Bayesian Models, whether it is simply

determining the likely sequence of an activity given

the objects (Wilson et al., 2005); (Patterson et al.,

2005) or being used as temporal smoother for

specific classifiers, and classifying likelihoods

(Lester et al., 2005). Dynamic Bayesian Networks

(DBN) have been used to capture relationships

between state variables of interest (Petney et al.,

2006), for example, in the common sense based joint

training approach (Wang et al., 2007), the DBN is

able to represent the state of a system in time slices.

The work in this paper is performing much the

same function of activity recognition via object

usage data. However rather than having complete

dependency on the object data for activity

recognition, we have developed a approach that is

based on hierarchal structured plans (representing

ADLs) where knowledge at different levels of

abstraction is used to determine which activity is

being carried out.

3 HIERARCHAL ACTIVITIES OF

DAILY LIFE

For the work in this paper, ADLs have been

represented in a hierarchal structure, where the

ADLs can correspond from a simple action such as

“switching the kettle on”, to a more complex activity

such as “making breakfast”. In order to

accommodate the different range of activities the

ADLs are modelled as plans. The plans are made up

of sub-plans. Where a plan cannot be decomposed

any further it is then recognised as a task. Task

recognition is based on analysing sensor event data

that is based on the usage of objects that have been

used to perform the activity. While ADL recognition

is based recognising constituent tasks that belong to

a particular ADL (Naeem and Bigham, 2009).

Figure 1 illustrates a structure of a Hierarchal

ADL (HADL), which depicts the ADL “Make

Breakfast”. This ADL contains a simple sequence of

tasks such as “Make Tea” and “Make Toast”. The

lowest tier of this hierarchal structure deals with the

incoming sensor events that have been detected.

These sensor events are then associated with the

tasks. For example in figure 1, kettle sensor event

can be associated with “Make Tea” or “Make

Coffee”. Once the sensor events have been mapped

into the associated tasks, an algorithm is then

applied in order to segment the tasks efficiently. For

the task recognition tier an approach has been

developed, which is responsible for generating a set

of different tasks sequences from a stream of object

usage data that is based on the conjunction of the

disjunction of task possibilities for each sensor

event. This approach is called Generating

Alternative Task Sequences (GATS).

Figure 1: Example of a hierarchal ADL (HADL).

For the higher tier, the number of levels above

the task identification level depends on the

complexity of the task. For example, an ADL may

have a series of nested sub-activities above the

actual task recognition level. Also there is a series of

possibilities that need to be considered when

modeling/ representing ADLs, such as:

 Some ADLs may occur in parallel with other

ADLs.

 ADLs may also have temporal constraints.

 Not all sub- activities need to be executed.

Taking the above into consideration, ADLs have

been represented using a knowledge representation

language called Asbru. This is a task-specific and

intention-oriented plan representation language

which was initially designed to model clinical

guidelines (Fuchsberger et al, 2005). This plan

representation feature allows the capability of being

able to represent ADL and sub-activities within an

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

ADL, for example, “Prepare Lunch” is and ADL,

and a sub-activity of this ADL is to “enter kitchen”.

An ADL 3recognition component for the higher tier

has been developed, which manages the output from

the task recognition component (lower tier) to

determine which activity is going to be conducted

and determine the current and future intentions of

the elderly person. Future intentions are established

by predicting what ADL the subject might conduct

next.

In order to generalise the activity and intention

recognition capability outside the framework of the

core ADLs constructed to support recognition,

decision trees are constructed using a well-known

induction algorithm during a training period. Once

the tree has been developed the trees are used as a

support tool for determining if a correct task or ADL

has been recognised at the current iteration of the

recognition process.

3.1 Task Recognition – Lower Tier

Tasks are considered to be short activities,

essentially atomic. The stream of sensor events from

the different objects will be small, and so an

enumeration based approach is feasible as long as

the combinations are explored in an ordered manner.

Hence the lower tier allows enumeration of the

possibilities, which can be useful when testing the

learning and feedback approaches at the higher tier

of the HADL. An enumeration-based approach is

also necessary for carrying out task segmentation in

this type of task identification. The entire sensor

event stream is segmented into appropriate task

segments. The segmented tasks are then used to

determine which ADL is currently active. There is

range of techniques that can be applied to the task

associated sensor events for segmenting them into

appropriate tasks. However the difference between

the GATS approach and other statistical approaches

(Naeem and Bigham, 2007) is that the GATS

approach employs a simple algorithm that works out

all the possible combinations for each task given the

sensor event. This approach therefore mitigates the

chances of not being able to recognise tasks that

have been conducted via different variations (Naeem

and Bigham, 2009). The execution of this approach

may seem computationally expensive when

performed, however a best first identification in

synchronisation with the ADL recognition in the

higher tier could prove a simple but effective

approach, particularly as each task will not be

associated with a large number of distinct sensor

events.

3.2 ADL Recognition – Higher Tier

The higher tier of the hierarchal approach gives an

overview of the possible ADLs that can occur within

a specified time frame. Additionally, the higher tier

has the capability of taking into account any

overlapping ADLs, which can be useful when trying

to determine the ADL that is currently active from

the tasks that are discovered in the lower tier task

recognition. The input for the higher tier recognition

components are task sequences generated by the

lower tier, while the output is a list of alternative

ADL sets, which are sequences of the possible

ADLs that could occur given the tasks sequences

that have generated from the lower tier. Each of the

ADLs sets has an associated utility, which is based

on the cost of each segmented task sequence. Hence

it is imperative to recognise as many tasks as

possible within a window of events, which in return

will lead to accurate activity recognition. The

generated utilities for the ADL sets are based on

ADL schedules within a certain time frame (e.g.

10.00am to 10.15am). This allows a more

manageable and accurate recognition process, as it

eliminates any unlikely possibilities from the initial

stages of the recognition process. The inspiration for

ADL schedules that are used for the hierarchal

approach originates from real life prescribed

activities that have been constructed by the

Alzheimer’s Association. The ADL schedules are

developed for helping people suffering from

dementia by planning their day with a prescribed set

of ADLs (The Alzheimer’s Association, 2005).

These set of activities are based on an interval based

structure, where the activities are grouped according

to different time segments throughout the course of

the day. However, there is always the possibility that

a number of ADLs can occur at any given time, e.g.

a phone ringing leads to the activity ‘engaged in a

phone call’. In the proposed hierarchal approach

these ADLs are referred to as interruption ADLs and

therefore these are modelled within every ADL

schedule in the ADL library.

4 RECOGNITION OF ADLS

SUPPORTED BY DECISION

TREES

Given the nature of the prescribed activity schedules

for people suffering from dementia and the

hierarchal recognition approach, it can be logical to

envisage that the most frequent ADLs will be

ACHIEVING MODEL COMPLETENESS FOR HIERARCHALLY STRUCTURED ACTIVITIES OF DAILY LIFE

modelled in the library of plans. However it can be

an audacious and near impossible task of making

sure that the library contains plans modelled for

every possible hierarchical ADL. Hence, extensive

use of decision trees has been made for constructing

trees using a well-known induction algorithm during

a training period that will support the recognition

capability outside the framework of the core ADLs.

The trees are used to support recognition of the ADL

at each iteration of the recognition process. For

example, every time a new task is recognised by the

lower task recognition tier, an ADL recognition

iteration is performed at the higher tier, which is also

used to predict the next ADL. This capability sits on

top of the hierarchal recognition process that finds

the best match in the kernel of ADLs. It is

instinctively obvious that if the ADL to be

recognised is in fact one of the core ADLs within the

library of plans, then recognition and prediction

could be fine tuned further.

For the recognition process, a decision tree is

generated for each ADL schedule, which is used to

classify the correct task/ADL that is being conducted

within the current ADL schedule given the current

instance and taking into account the training data.

The decision tree has to be learned during a training

phase. The data needed for this training phase can be

generated in two ways. In the first case, the data

generated can be based on subjects performing

ADLs from the core ADLs only, where the

information used is based on the tasks and sub-

activities actually undertaken by the subject. In the

second case, the subject may follow other plans, not

necessarily one of the core ADLs during training

and the information used in the training instance is

based on tasks actually observed and the best match

to ADLs in the core ADL library. Even though none

of the core plans are necessarily being followed, the

system will find a nearest match to use in the

training instance. In both cases the training is done

using information taken from the core ADLs.

A learning instance is created when each task is

labelled during training. The objective of the

decision tree is to act as a classifier that is used to

predict the class label for all labelled instances. In

order to determine an outcome for an instance a

decision tree needs to find an appropriate node to

split in order to form the branches and leaves of the

tree, which will lead to a predicted outcome.

Information theory is used to split the sets of training

instances associated with each node in the tree,

which leads to small and consistent nodes being

generated. The algorithm used is ID3.

4.1 Information Gain Split Decision

Trees

Figure 2 shows an ADL schedule modelled for the

time interval 9.00- 10.00. This ADL schedule also

incorporates the location of where each task is

conducted.

Figure 2: ADL schedule 1 modelled for decision trees.

When a task is recognised in the lower tier, the

location of where the task was conducted does get

recognised, however we make full use of this

information when constructing a decision tree based

on the ADLs within the ADL schedule that this task

belongs to.

Figure 3: Decision tree (ID3 Splitting) based on ADL

schedule 1.

Typically the decision tree learning algorithm

computes the quality of each possible split that can

be produced by each attribute and chooses the

attribute that has the highest utility based on the

quality of the split. The ID3 algorithm has been

adopted and illustrated in figure 3.

The entropy formula (1) is an idea formulated in

information theory that is used to measure the

amount of information in an attribute. Given a

collection S (entire sample set) of m outcomes:





ppSEntropy

log)(

(1)

where

is the proportion of S belonging to class i,

while ∑ is over the m labels. Note that a entropy

formula normally uses log base 2, however on this

occasion we use log base 10 as we are simply

looking to get to a classification point where the

lowest entropy, rather than an absolute value.

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

This is then followed by computing the expected

entropy for each attribute to see which attribute has

the highest gain so that it can be used as a split to

build the tree further. The gain for each attribute is

determined is a follows (2):



 )_()_()( setschildSsetcurrentSAGain

(2)

The gains for each of the attributes are shown in

table 1, which shows that attribute ‘Previous Task’

has the highest gain value, hence in figure 3 it is

chosen as the node which is split.

Table 1: Gains for all of the attributes to determine where

to split node.

Attributes Gain

Room 1.457

Time Frame 1.128

ADL 1.903

Previous Task 2.165

Previous ADL 1.276

This splitting process continues until a situation

is reached were the remaining entropy is equal to 0.

Given the following instance after a task has

been identified, we can identify by looking at the

decision tree (figure 3) that the task that has been

conducted is task ‘c’.

{Room of Observed Task = Kitchen, Time Frame

of Observed Task= 9.15-9.30, Parent ADL of the

Observed Task =1, Grandparent ADL of the

Observed Task = Root, Previously Observed

Task=a, ADL of Previously Observed Task=1}

We can see that information gain is good as a quality

measure for the decision trees that we have

constructed for correctly classifying a task within the

ADL schedule. However only one attribute is tested

at time for making a decision, therefore it cannot

take into consideration other future child nodes, as

its priority is to split the attribute it is currently at. In

addition it can also be computationally expensive

when classifying continuous data.

4.2 Gain Ratio Split Trees

Another method that can be used as splitting criteria

is gain ratio, which is a way of compensating for a

large number of attributes by normalising. This is

done by computing the information gain for an

attribute, which is then followed by dividing the gain

for the attribute by the information associated with

that attribute that is based only on the set of values

for that attribute. Figure 4 shows a tree constructed

based on the labelled data generated by figure 2.

Figure 4: Decision tree (Gain Ratio Splitting) based on

ADL schedule 1.

It can be seen that both of the trees generated via

two different splitting methods are different,

however both of the generated trees are correct in

terms of current training data that we have and we

already know. It is important to evaluate both sets of

trees to see which would be best suited for carrying

out classification if an unlabeled instance occurred.

5 EXPERIMENTS AND RESULTS

The objective of these experiments is to see which

splitting criteria is best suited to construct the

decision trees and to assess the potential of the

decision tree approach in predicting the next task or

ADL in a context where the performed activities do

not match any of the plans associated within the core

ADLs. Both of the splitting methods have been

tested with different combination ranges of labelled

and sample holdout instances.

The training instances for these experiments are

based on activities that have been carried out using a

wide range of objects (e.g. Kettle, Mug) that were

tagged with RFID transponders. Whenever these

objects were used or touched the object data was

captured by an RFID reader, which is a size of

matchbox and was worn on the finger of the subject

conducting the experiment. The subjects carried out

these experiments in a range of rooms such as

kitchen, bathroom and living room.

The activities carried out were based on two

ADL schedules, ‘Morning’ and ‘Afternoon’

activities. Both ADL schedules are similar to the

ADL schedule in figure 2, as they take into

consideration the location of where the tasks have

been conducted. For both of the schedules, two sets

of decision trees have been constructed from two

sets of training data, one is used to classify the

outcome of the next task, while the other tree is

classifying the parent ADL of the next task being

conducted. Both ADL schedules for morning and

ACHIEVING MODEL COMPLETENESS FOR HIERARCHALLY STRUCTURED ACTIVITIES OF DAILY LIFE

afternoon will also incorporate Interruption ADLs,

such as a phone call, someone at the door or going to

the toilet. Each of the ADL Schedules used for these

experiments has different training data sets used to

build its decision tree. As well as having instances

which correspond to the different timings of the day

(e.g. morning and afternoon), each of these decision

trees built from the training data also have different

characteristics that imposed to validate different

types of schedules. For example, training data for

morning ADL schedule has incorporated instances

that have an outcome of an interruption ADL

differently to the way the instances are incorporated

in the training data for afternoon ADL schedule.

Table 2: Holdout samples for splitting criteria

experiments.

Holdout Sample

[%]

Training

Data

Holdout

Sample

Morning ADL

schedule

20 176 46

Morning ADL

schedule

50 111 111

Morning ADL

schedule

90 22 200

Afternoon ADL

schedule

20 162 40

Afternoon ADL

schedule

50 101 101

Afternoon ADL

schedule

90 20 182

Using different size variations of the labelled

data as holdout samples has been used to see how

well the splitting approaches work with different

sizes of holdout samples. Table 2 shows the

variations of holdout samples that were used for

these experiments. Three variations of holdout

sample have been used, these are 20%, 50% and

90% of the complete training data size, which is 222

instances for morning ADL schedule and 202

instances for afternoon ADL schedule.

The results in table 3 indicate that for both ADL

schedules, gain ratio was more efficient way of

splitting the attributes for constructing a decision

trees as it had higher percentage of classification

results for the holdout samples. One of the reasons

why gain ratio performed better as a splitting

approach than the ID3 is because in contrast to the

gain ratio splitting approach, the ID3 tends to learn

the training set too well when attributes have a large

number of distinct values, which can also be its

downfall when trying to classify instances that have

not occurred before.

In relation to the task being carried out, the

attribute with the highest gain might be the previous

task within the current ADL schedule, as this will

also be able to uniquely identify a task given the

previous task. However this is not always suitable,

as a tree that focuses its classification based on

previous tasks is unlikely to recognise a task that has

not been witnessed before.

Table 3: Results of holdout samples correctly classified.

Holdout

Sample

[%]

Morning ADL Schedule

Afternoon ADL Schedule

ID3

[%]

Gain Ratio

[%]

ID3

[%]

Gain Ratio

[%]

20 91 93 98 99

50 75 82 96 98

90 62 71 78 86

The results in table 3 reiterate the fact that the

gain ratio splitting is better at considering unknown

tasks or unlabelled instances, as gain ratio splitting

performed better with all holdout samples for the

morning ADL schedule, which consisted of tasks

from interrupted ADLs occurring at random

junctures within the constructed training data.

Another observation is that both of the splitting

methods classified the holdout samples better for the

afternoon ADL schedule than the morning ADL

schedule. This was expected as the morning ADL

schedule was intentionally constructed with

infrequent and inconsistent appearance of tasks with

no particular order. However, this does not imply

that training data constructed for the afternoon

schedule was simply easy for classification, as it was

constructed keeping in mind the general slower

pattern of how activities and tasks would normally

be conducted by Alzheimer’s patients.

6 CONCLUSIONS

The work described in this paper looked at how

decision trees can be utilised for generalising a

hierarchal approach for activity recognition. The

integration of decision trees gives the potential of

being able to carry out activity recognition, with the

intention of being able to learn and predict the

likelihood of what task within an activity may be

conducted next. Out of the two splitting methods

that were used for constructing the decision trees it

can be seen that the gain ratio method performed

better whilst trying to classify instances that have

not occurred before. However, the interaction of

these approaches is only successful when consistent

and cohesive training data is available.

Further work is being carried out that is

exploring ways of using the ADL recognition

process that has been described in this paper for

hygiene related activities that can help stop

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

spreading of diseases amongst Alzheimer’s patients.

In addition, privacy is an area of prime importance,

as assistive technologies should not be needlessly

intrusive or the elderly community will simply

refuse to use them, despite their potential benefits.

Hence the work in this paper did not make use of

any visual surveillance equipment. Nevertheless

even RFID sensors can be intrusive to a certain

extent and once such approach that will be

investigated is the integration of privacy policies

into our current hierarchal approach. A person may

want to switch some or all of the sensors off from

time to time, or may opt for a programmed approach

where more sensors can be used at certain times of

the day, or if the system believes that the person is in

need of help. The question of accuracy is a difficult

one as increased detection usually means false

positives and a trade off between the two is

necessary. However policies for when more

information is needed could be used to mitigate this

problem.

REFERENCES

The Alzheimer’s Association, 2005. Activities at Home,

Planning the day for a person with dementia,

http://www.alz.org/national/documents/brochure_activ

ities.pdf

Alzheimer’s Research Trust., 2010. Dementia 2010, The

prevalence, economic cost and research funding of

dementia compared with other major diseases.

Executive summary, Health Economics Research

Centre, University of Oxford for Alzheimer’s

Research Trust. http://www.dementia2010.org/reports/

Dementia2010ExecSummary.pdf

Buettner, M., Prasad, R., Philipose, M., Wetherall, D.,

2009. Recognizing Daily Activities with RFID-Based

Sensors. In Proceedings of the 11

International

Conference on Ubiqitous Computing, Orlando

Ferretti, L., McCurry, S. M., Logsdon, R., Gibbons, L. E.,

Teri, L., 2001. Anxiety and Alzheimer’s Disease. In

Journal of Geriatric Psychiatry and Neurology, vol.

14, no. 1, pp.52-58

Fuchsberger, C., Hunter, J., McCue, P., 2005. Testing

Asbru Guidelines and Protocols for Neonatal Intensive

Care. In Proceedings of the Tenth European

Conference on Artificial Intelligence, Aberdeen,

United Kingdom, pp.101-110

Kalimeri, K., Matic, A., Cappelletti, A., 2010. RFID:

Recognizing failures in dressing activity. In

Proceedings of the 4

International Conference on

Pervasive Computing Technlogies for Healthcare

(PervasiveHealth), Munich

Lester, J., Choudhury, T., Kern, N., Borriello, G.,

Hannaford, B., 2005. A Hybrid

Discriminative/Generative Approach for Modelling

Human Activities. In Proceedings of the 19

International Joint Conference on Artificial

Intelligence, pp. 766-772

Naeem, U., Bigham, J., Wang, J., 2007. Recognising

Activities of Daily Life Using Hierarchical Plans. In

Proceedings of the 2nd European Conference on

Smart Sensing and Context, LNCS 4793, Lake

District, UK, pp. 175-189

Naeem, U., Bigham, J., 2009. Activity Recognition in the

Home using a Hierarchal Framework with Object

Usage Data. In Journal of Ambient Intelligence and

Smart Environments, IOS Press

Patterson, D. J., Fox, D., Kautz, H., Philipose, M., 2005.

Fine-Grained Activity Recognition by Aggregating

Abstract Object Usage. In Proceedings of the 9

IEEE

International Symposium on Wearable Computers,

Osaka, Japan, pp.44-51

Petney, W., Popescu, A., Wang, S., Kautz, H., Philipose,

M., 2006. Sensor-Based Understanding of Daily Life

via Large-Scale Use of Common Sense. In

Proceedings of the 21

AAAI Conference on Artificial

Intelligence, Boston, USA

Philipose, M., Fishkin, K. P., Perkowitz, M., Patterson, D.

J., Kautz, H., Hahnel, D., 2004. Inferring Activities

from Interactions with Objects. In IEEE pervasive

Computing Magazine, vol.3, no.4, pp.50-57

Philipose, M., Smith, J. R., Jiang, B., Mamishev, A., Roy,

S., Sundara-Rajan, K., 2005. Battery-Free Wireless

Identification and Sensing. In: Pervasive Computing

IEEE Journal, vol. 4, no.1, pp.33-45

Wang, S., Petney, W., Popescu, A., Choudhury, T.,

Philipose, M., 2007. Common Sense Based Joint

Training of Human Activity Recognizers. In

Proceedings of the 20

International Joint Conference

on Artificial Intelligence, Hyderabad, India, pp.2237-

2243

Wilson, D. H., Long, A. C., Atkeson, C., 2005. A Context-

Aware Recognition Survey for Data Collection Using

Ubiquitous Sensors in the Home. In Proceeding of the

International Conference for Human-Computer

Interaction, Portland, Oregon, USA, pp.1865-1868

ACHIEVING MODEL COMPLETENESS FOR HIERARCHALLY STRUCTURED ACTIVITIES OF DAILY LIFE