Database Reference
In-Depth Information
esophageal reflux, along with congestive heart failure, coronary artery
disease, and chronic obstructive pulmonary disease.
Sensor mining, on data collected from a combination of body sen-
sors and smart environments, has been used successfully for automatic
assessment of ADL-IADL activities. In [102] RFID tags are attached
to different key objects with which a person interacts for a specific set
of activities. The data from these tags is augmented by accelerome-
ters placed at diffrent strategic locations on the person (such as wrist,
hip, and thigh). The combined dataset is analyzed using different fea-
ture extraction and mining and classification techniques. The computed
features include statistical properties such mean, variance, energy, spec-
tral entropy, pairwise correlation between the three axes, and the first
ten FFT coecients and exponential FFT bands, computed over sliding
windows shifted in increments of 0.5 seconds. For classification of ac-
tivities the authors use three different approaches, namely Naive Bayes,
Hidden Markov Models (HMMs) and Joint Boosting. They show that
Naive Bayes and HMM classifiers are well-suited for low-level activities
such as sitting, standing and walking or wood workshop activities. The
Joint Boosting method is successfully applied to overcome limitations
of the sensing and feature extraction. The results show that combined
recognition helps in cases when tagged objects are being shared among
the activities, as well as in periods when the RFID reader can not detect
interactions with objects due to its short range. The authors also con-
sider extensions of this work to include techniques for accurate activity
recognition with reduced supervision.
Researchers from the Imperial College [80] have developed an ear-
based Activity Recognition (e-AR) sensor that identifies four different
levels of activity ranging from almost no activity (during sleeping or
sitting for example) to activities involving a lot of movement (running,
exercising). The activity level is continuously detected using a classi-
fier applied to the accelerometer measurements and streamed from the
e-AR device every 4 seconds. While some activities may be described
by a single activity level, many activities produce a sequence of activity
levels. The work in [81] uses the output of the e-AR sensor to e-
ciently mine and update a concise variable-resolution synopsis routine
for ecient behavior profiling in a home healthcare environment. The
authors use the FP-Stream [82] and Closet+ [83] mining algorithms to
describe behavior patterns using a routine tree data structure. The au-
thors demonstrate that using this technique they can identify frequent
patterns to describe the structure present in an individuals daily activity,
and can then analyze both routine behavior as well as deviations.
