Signal Processing - March 2016 - 86

are used as features for neural network models, while the outTwo broad categories in machine learning are generacomes of the neural networks are fused under an HMM.
tive and discriminative models, where the former is modelFor the data representation in activity recognition and ADL
ing the joint probability distribution of the samples and the
classification scenarios, the bag-of-words (BoW) approach has
labels and the latter is modeling the conditional probability
proven to be convenient and successful. Originating in natural
of the labels given the samples. The standard HMM is a typilanguage processing, the BoW approach represents a text (such
cal algorithm of the first category, with several of its extenas a sentence or a document) as the bag (multiset) of its words,
sions also falling into the same group. In the discriminative
disregarding grammar and even word order but preserving
group, the basic models are CRFs and their extensions [for
multiplicity. An analogous bag-of-visual-words also has been
example, latent-dynamic CRFs (LDCRFs) and semi-Markov
successfully used for general image classification [79] and
CRFs (SMCRFs)] as well as certain types of artificial neural
later for human action recognition and classification in video
networks (ANNs), with the most prominent ones being the
sequences [80]. Recent studies on human activity recognition
recurrent neural networks (RNNs).
show that the BoW representation allows achievement of highFinally, a multitude of hybrid methods, aiming to combine
performance action recognition [81], [82].
the advantages of discriminative and generative models, are
also available. These include, for example, approaches relying
on kernel metric distances such as the Fisher kernel and variGenerative models
ous combinations of HMMs with discriminant models such as
Generative models estimate the joint probability distribution
random forests and ANNs.
of observation samples, which can be used to predict the
While most work in ADL classification is performed using
most likely class to which a new sample belongs. They are
one of the aforementioned machine-learning techniques,
called generative, because the model can be used to generheuristic methods also were successfully applied. Short-term
ate samples given the joint probability distribution. HMMs
activities and data sets with sufficiently
are a popular generative model that can
redundant sensor setups (to suppress false
deal with structured data where the IID
Because HMMs are suitable
alarms) are especially suitable for heuristic
assumption does not hold. In the context
to model sequential data, it
methods. One successfully applied heuristic
of traditional HMMs (having a finite
is a popular classification
is the circadian activity rhythms [23], [77],
number of discrete states), three impormethod in activity
which describe the measurement of home
tant questions are asked as part of the
recognition.
rhythmic behavioral activity as the resident
model learning and its application on
engages in the habitat. In some cases, these
unseen data [83].
simple heuristics are either fused together or used as features
1) Likelihood: Given a model and a sequence of observations,
for a second-level machine-learning algorithm. For example,
how likely is it that this sequence was generated by the
in [78], simple heuristics measures like means and variances
given model? The answer to this problem is given by the
forward-backward algorithm.
2) Decoding: What is the most likely sequence of model
states that generated a sequence of observations? The
answer to this question is given by the Viterbi algorithm.
3)
Learning: How should transition and emission probabiliGenerative
Discriminative
ties be learned from observed sequences? The answer is
given by the Baum-Welch algorithm, which can be seen as
a special case of the expectation maximization algorithm
and tries to optimize the model parameters to best describe
FKL
HMM
CRF
the observation sequence, while using also the results of
TOP
the two previous problems.
HHMM
LDCRF
Because HMMs are suitable to model sequential data, it
is a popular classification method in activity recognition. A
RF-HMM
Spectral
variety of HMM-based variants is presented in a compreSMCRF
HMM
hensive survey by Turage et al. [84]. Also, the recognition
ANN-HMM
of human motion data can be modeled with HMMs. Li [85]
RNN
RBM
proposed a straightforward and effective motion descriptor
CRF-HMM
based on oriented histograms of optical flow field sequences. Following dimensionality reduction performed by principal component analysis, the method was applied to human
action recognition using the HMM approach. Yamato et al.,
in [86], used HMMs in their simplest form: training a set
of HMMs, one for each action and modeling the observafigure 3. The taxonomy of algorithms for structured learning.
tion probability function as a discrete distribution, adopting
86

IEEE SIgnal ProcESSIng MagazInE

March 2016

Table of Contents for the Digital Edition of Signal Processing - March 2016