Signal Processing - March 2016 - 32

from the action research arm test, which is designed to test
recovery of upper-limb function [29].
Instance-based learning methods classify an instance
based on the similarity between the instance under test, and
the labeled instances in the training data set. This method
does not need to train a model in the training phase. However,
it is computationally expensive in the testing phase because it
needs to calculate the similarity between each testing instance
and all of the instances in the training set. The k-nearest
Classification
neighbors (kNN) algorithm is one example of an instanceClassification is widely used in applications of assisted living.
based classification algorithm. It performs well in activity recClassification can be used to detect falls and prefalls, to disognition tasks, and it is used to determine the different types
tinguish between healthy and unhealthy motor function, and
of the ADLs [16], [24].
to detect ADLs. A variety of machine-learning and pattern
Neural networks are a family of statistical algorithms
recognition algorithms are explored in the area of the IMUinspired by biological neural networks (i.e., the human
based assisted living. Table 2 shows some of the commonly
brain). It consists of a large number of nodes acting as neuused classification algorithms.
rons in a network and the weighted conThresholding-based decision making is
nections between different neurons. With
a popular classification scheme in assisted
feature selection provides a large enough set of training data and
living applications. This approach is
a way to select the most
parameter tuning, it can provide high
straightforward to use and is often used for
classification performance. A very large
binary classification tasks. When the value
suitable feature subset
data set is often required for training, and
of a feature is above a threshold, it is clasfor certain tasks from the
this is not usually available for IMUsified as one of the two states and when the
available features.
based applications. Moreover, the trained
value is below a threshold, it is recognized
model is not interpretable for users. In
as the other. When the designer finds a feaIMU-based assisted living applications, the training data is
ture that can discriminate between two possible states, the
usually small, and, in most cases, the user wants to underthresholding technique is a good candidate due to its simplicistand the models. These two factors make neural networks
ty and because it can be easily interpreted. The thresholding
less attractive in this area. The authors explored the classifitechnique is applied to classify walking versus running [10].
cation performance of a neural network while varying the
If the variance of the accelerometer is below a defined threshsize of the training data set for a physical movement moniold, the activity is recognized as walking, and recognized as
toring application [16]. Four transition movements were
running if the accelerometer variance is larger than a defined
detected using the neural networks and kNN for an average
threshold. Based on this decision, an adaptive step length estiaccuracy of 84%.
mation algorithm is derived. A thresholding technique is
SVM is one of the most popular discriminative classificaapplied to the inertial frame's vertical velocity magnitude to
tion algorithms in different areas in recent years. SVM tries to
detect the occurrence of falls before impact [28]. To deterfind the margins that will maximize the separation between
mine the posture transition time for sit-to-stand, a threshold is
different classes. In the training phase, the margins are deterapplied to determine the beginning and ending of the transimined and it is computationally efficient in the testing phase
tion movement [8]. A threshold based on the maximum meabased on the trained model. It is similar to neural networks in
sured vertical velocity from ADLs and the minimum
that it will be difficult to interpret by users. However, it does
measured vertical velocity from falls is used to distinguish
not require a very thorough training or a very large training
falls and normal activities [15]. Thresholding is used to distindata set. A preimpact fall detection system is discussed based
guish tremor motions from nontremor motions in a movement
on the SVM classifier [14]. A SVM is applied for monitoring
motor fluctuations in patients with Parkinson's disease and the
optimal kernel is analyzed [23].
Table 2. Classification algorithms.
The HMM is a statistical Markov model in which the sysClassification Type
Classifier
tem is assumed to be a Markov process with unobserved
states. HMM is well studied and is often used in temporal
Thresholding
[6], [8], [10], [15], [28], [29]
pattern recognition such as speech recognition and gesture
Instance based
KNN [16], [24]
recognition. It is widely used to recognize different activities
Neural networks
Multilayer perceptron [16]
based on IMU time series sensor data and is also good at recLinear kernel SVM [14], polynomial kernel
ognizing a sequence of movements. Human intention recogniSVM
SVM [23]
tion in smart assisted living systems is presented using a
Hidden Markov models
Hierarchical HMM [7], continuous HMM [27],
hierarchical HMM [7]. The HMM is first used to recognize
(HMMs)
HMM [4], [17]
the low-level hand gestures with a finger-worn inertial sensor
elderly and patients with chronic diseases [24]. Three different feature selection algorithms are tested for 13 different features for five different groups of ADLs. Accelerometer based
balance parameters are determined and compared during the
sit-to-stand movement and the results show the area under the
curve (AUC) and RMS are useful features and AUC appeared
to be more sensitive than RMS [18].

IEEE SIgnal ProcESSIng MagazInE

March 2016

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