Signal Processing - March 2016 - 30

Therefore, a fall detection system is necessary in assisted living. There are two major considerations for this application:
1) speed of detection and 2) accuracy of detection. The first is
important because the nature of this application is to prevent
injuries caused by falling if possible. For example, having a system that can detect falls before the user hits the ground
becomes very promising if the system can activate a protective
device (e.g., air bags attached to the user's hip). If no protective
device is available, it is critical to get help to the user as soon as
possible. The second consideration is necessary to ensure that
falling events are properly detected and there are no false or
missed detections. Several methods have been proposed using
IMU sensors to detect falls before the target hits the ground
[14], [15]. In these methodologies, detection speed is very
important and challenging. Another approach reliably detects
falls while also finding other common human activities that
may share similar attributes (e.g., standing to lying in bed) [16].
Many elderly people cannot stand after falling due to injuries
sustained, and an emergency message is needed to inform the
hospital or care center [8].

Rehabilitation
There are many cases in which clinicians do not need to keep
patients in the hospital to observe the recovery process after
an operation or treatment. IMU-based systems can play an
important role here and can allow patients to live independently in their homes while being monitored remotely by clinicians. The idea behind such systems is to provide general
information about the effect of certain behavioral recommendations without having the patient admitted to a rehabilitation
center or a laboratory for observation [17]. IMU sensors are
able to measure the muscle strength and power by detecting
high-frequency body sway [18] and the speed with which
muscular forces produce movement of body segments [19].
Estimating knee joint flexion or extension angles can be used
to infer activity type or intensity, muscle activity, and gait
events [20]. Monitoring ADLs is also a key for evaluating
changes in physical and behavioral profiles of the elderly and
other patients, including obese people [21]. For example,
increasing activity levels after surgery can be used to indicate
overall improvement as well as efficacy of therapeutic procedures [22].

Signal processing techniques
Signal processing techniques translate the physical signals
sensed from wearable IMU sensors into useful information
required by target applications. In this article, our goal is to
review the signal processing techniques from various perspectives, including preprocessing, feature extraction, feature
selection, classification, and measurement models.

Preprocessing
For IMU-based assisted living applications, the raw sensor
data usually gets preprocessed to remove noise from the signal and to determine the segments of interest. These tasks are
called filtering and segmentation. Filtering techniques retain
30

the useful information in a signal while rejecting unwanted
information based on the application. Segmentation techniques are used to determine the duration of the movements
or events of interest.
Three different types of filters are used: low-pass filters, highpass filters, and band-pass filters. A low-pass filter is used to
remove high-frequency noise for a recognition task of five hand
gestures [7] and for physical activity monitoring for assisted living [16]. A 17-Hz low-pass filter is used to reject electronic noise
in gyroscope data for sit-to-stand and stand-to-sit measurements
[8]. Based on the walking frequency of test subjects, a 3-Hz lowpass filter is applied to remove noise from walking signals [12].
A 6-Hz low-pass filter is applied for balance control measurements during sit-to-stand movements [18], while a low-pass filter
with a cut-off frequency of 3 Hz is used to preprocess raw data
for sit-to-stand parameter measurement [19]. The accelerometer
measurement consists of gravitational acceleration and dynamic
acceleration caused by motion. In some applications, only one
part of the acceleration is used, and filtering techniques are
applied to reject the other one. A 1-Hz low-pass filter is used to
remove the dynamic acceleration, and thus the direction of the
gravity vector is found during quasi-static activities [15]. A 1-Hz
high-pass filter is used to reject the gravitational acceleration,
which, in turn, removes the effect of the gross changes in the orientation of the body segment where the sensor is placed [23].
Some applications may only look at signals within a certain frequency range, and the band-pass filter can be used to preprocess
the data. A 3-11-Hz band-pass filter is used to clean the accelerometer signal for detecting sleep and awakening phases [5]. For
motor fluctuation monitoring in Parkinson's disease patients, a
3-8-Hz band-pass filter is used for the analysis of tremors, and a
1-3-Hz filter is applied for analysis of bradykinesia and dyskinesia [23]. The sliding window segmentation technique is simple
and effective and is often used in the reviewed literature for segmentation [6], [7], [14], [17], [24], [25].

Feature extraction
Features are normally extracted from the sensor data depending
on their effectiveness in a particular application. Feature extraction starts with the preprocessed sensor data and generates
derived values that are intended to be informative and nonredundant while enabling subsequent learning and generalizing
the data, which will lead to better human interpretation. The
features are divided here into four categories: time domain features, frequency domain features, time-frequency domain features, and others. The time domain features are the general
statistical measurements that can represent the generalization
of the data. The frequency domain features analyze the frequency performance of the sensor signals, which is usually
the periodicity of the signal over a long duration (i.e., periodicity of the walking). The time-frequency domain features
refer to features that contain both time and frequency information simultaneously with different time-frequency representations (e.g., short-time Fourier transform, wavelets) that
are useful for nonstationary signals (e.g., postural transitions).
The other features refer to the features that have specific

IEEE SIgnal ProcESSIng MagazInE

March 2016

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