Citation
Mohamed, Raihani
(2018)
Improving multi-resident activity recognition in smart home using multi label classification with adaptive profiling.
Doctoral thesis, Universiti Putra Malaysia.
Abstract
“Smart Home” services offer to improve living conditions and levels of
independence for the elderly that require support with both physical and cognitive
functions via Activities of Daily Living (ADL). Due to human ethics and privacy
concern, ambient-based sensor technologies are preferred and deployed in the
environment. Nevertheless, as human activities gradually becoming complex and
thus complicate the inferences of activities especially involving multi-resident
within the same home premises that deploy solely ambient-based sensor
technology.
Existing works and solutions focused on separate models for recognizing the
residents, activities and interactions. On top of that, data association and algorithm
modification inherit drawbacks on recognizing the residents and interactions of
multi-resident complex activities. When the data are induced with the lower quality
model, the performance is also truncated. Furthermore, there is tendency that multi
label classifications used instead of traditional single label classification technique.
Consequently, this could cater the simple and complex activity recognition of
multi-resident in a separate model. Moreover, with the incremental numbers of
resident living together in the same smart home environment, the class-overlapping
sensor event sequence could occur and might share the same features for subsequences
that correspond to each individual activity. At the same time, the sensor
events are always uncertain and intricate in nature led to conflict occurs at its
interaction layer.
In accordance to the mentioned problem, Label Combination (LC) of multi label
classification is introduced because of its ability to transform the multi label
problem into 2ᶫ multi-class problem and exploit the correlation between the class
labels. On top of that, the label correlation can be solved with the Random Forest
(RF) as a base classifier due to its capability to produce the most probable class
from its majority-voting task as output. Nevertheless, the learning complexity of
classification is increased due to the increment number of residences and activities
are also intricate. Therefore, Adaptive Profiling (AP) for multi-resident involving
context information includes temporal and spatial information is proposed to
address the class-overlapping using Expectation-Maximization (EM) clustering.
The clusters parameter is adaptively generated from the active labelset from the
real-world data. The multi label relation method using Two-Stage Label
Construction (TSLC) is presented, resolve the conflicts in complex activity of
multi-resident is also outlined in this research.
Two publicly available datasets; WSU’s CASAS and ARAS Dataset are selected
and experimented to evaluate the proposed framework. About 26 pairs of volunteer
performing 15 scripted activities collected over four months’ time with almost
17,500 instances from CASAS. In addition, three days of house A from ARAS
dataset is also selected to evaluate its effectiveness. LC-RF is tested with other base
classifiers such as k-NN, SVM and HMM. However, LC-RF showed the most
promising results among others. Furthermore, its performance is also benchmarked
with previous work that used single label classification. Consequently, the obtained
results demonstrate the improvement of 2.4% increment in Hamming score as
compare with the highest results from the previous work. Experimental results have
significantly promised an improvement level in multi-resident simple and complex
activity recognition simultaneously, capable to cater the problems mentioned
specifically when the number of resident increase and reside together in the same
smart home environment.
Download File
Additional Metadata
Actions (login required)
![View Item View Item](/style/images/action_view.png) |
View Item |