Human Fall Detection using Built-in Smartphone Accelerometer

Abstract

Falls are serious health hazard issue among the aged people around the world. It’s a common accident for the elderly people living alone. Obviously, this accident can be timely reduced by using accurate fall detection method in order to reduce injuries and loss of life. For this purpose, we used smartphone-based fall detection method using the features of triaxial acceleration values of x, y and z which is obtained from the built-in accelerometer sensor embedded on our smartphones. We do a lot of daily activities like sitting, walking, standing, lying, and running. These were collected through accelerometer data. An app was used called Physics Toolbox Sensor Suite to take the data values which consist of accelerometer. The data values were taken through two positions one in chest pocket and another in pant pocket for both falls and non-falls. Also, intentional falls were also taken like fall -forward, fall-backward, right lateral fall, left lateral fall and so on. All these data were collected together to distinguish between fall and non-fall. These falls and non-falls were submerged together in a given time set keeping its frequency fixed along 6000samples from each data set through MATLAB. Then by using the Neural Net Pattern Recognition app leads us solving data classification problem using two-layer feed forward network. Using our data, we trained, validated and test the data through Neural Network Pattern Recognition, and achieved our classification accuracy to 90.6%. Using 67 data consisting of 26 falls and 41 non-falls. Basically, we classified and predict the data’s through offline activity recognition. Once the falling victim is detected his positions along its locations will be tracked. And instantly will send an alert to the caregivers for immediate assistance.

Full text

2020 IEEE International Women in Engineering (WIE) Conference on Electrical and Computer Engineering (WIECON-ECE)
978-0-7381-3144-3/20/$31.00 ©2020 IEEE
Human Fall Detection using Built-in Smartphone
Accelerometer
Chowdhury Sayef Abdullah
Dept. of Electrical & Electronic
Engineering
Independent University, Bangladesh
Dhaka, Bangladesh
sayef.384@gmail.com
Tasnuva Faruk
Dept. of Public Health
Independent University, Bangladesh
Dhaka, Bangladesh
tasnuva.faruk@iub.edu.bd
Masud Kawser
Dept. of Electrical & Electronic
Engineering
Independent University, Bangladesh
Dhaka, Bangladesh
kawsermasud35@gmail.com
Md Kafiul Islam
Dept. of Electrical & Electronic
Engineering
Independent University, Bangladesh
Dhaka, Bangladesh
kafiul_islam@iub.edu.bd
Md Tawhid Islam Opu
Dept. of Electrical & Electronic
Engineering
Independent University, Bangladesh
Dhaka, Bangladesh
1931471@iub.edu.bd
Abstract— Falls are serious health hazard issue among the
aged people around the world. It’s a common accident for the
elderly people living alone. Obviously, this accident can be
timely reduced by using accurate fall detection method in order
to reduce injuries and loss of life. For this purpose, we used
smartphone-based fall detection method using the features of
triaxial acceleration values of x, y and z which is obtained from
the built-in accelerometer sensor embedded on our
smartphones. We do a lot of daily activities like sitting, walking,
standing, lying, and running. These were collected through
accelerometer data. An app was used called Physics Toolbox
Sensor Suite to take the data values which consist of
accelerometer. The data values were taken through two
positions one in chest pocket and another in pant pocket for both
falls and non-falls. Also, intentional falls were also taken like fall
-forward, fall-backward, right lateral fall, left lateral fall and so
on. All these data were collected together to distinguish between
fall and non-fall. These falls and non-falls were submerged
together in a given time set keeping its frequency fixed along
6000samples from each data set through MATLAB. Then by
using the Neural Net Pattern Recognition app leads us solving
data classification problem using two-layer feed forward
network. Using our data, we trained, validated and test the data
through Neural Network Pattern Recognition, and achieved our
classification accuracy to 90.6%. Using 67 data consisting of 26
falls and 41 non-falls. Basically, we classified and predict the
data’s through offline activity recognition. Once the falling
victim is detected his positions along its locations will be tracked.
And instantly will send an alert to the caregivers for immediate
assistance.
Keywords—fall detection, accelerometer, smartphone sensor,
classification, neural network
I. INTRODUCTION
A. Overview
Most of the times fall may be the leading cause of
accidental or unintentional injury throughout the world.
According to WHO every year about 6,46,000 individuals die
due to various falls globally of which around 80 % cases are
from low and middle-income countries [1]. The aged or
elderly people in a population inspired the development of
various research based on healthcare. In today’s world elderly
people might live alone which increases the chance of
different fall events. To reduce the events related to fall and
deaths caused by fall many researchers devoted to the research
on the field of fall detection. In order to detect fall mobile
phone sensor are being used now-a-days. In the past detection
of fall wasn’t so easy as this day. There were many
machineries and various method were involved in detection of
fall. In a mobile device there are many sensors like
acceleration sensor, magnetic field sensor, light sensor,
proximity sensor, gyroscope sensor, gravity sensor etc.
Modern phone or termed as “Smartphone” contains tri-axial
accelerometers that measures acceleration in x, y and z -axis.
In this report, we explore the use of acceleration sensor and
obtaining various data which we classified using feature
extraction.
B. Background
Falls speak to a huge danger to the wellbeing and
autonomy of grown-ups 65 years old and more seasoned. As
a wide assortment and huge measure of detached checking
frameworks are as of now and progressively accessible to
distinguish when an individual has fallen, there is a need to
dissect and orchestrate the proof with respect to their capacity
to precisely recognize tumbles to figure out which frameworks
are best [2].
Insights and realities related with falls in older individuals
is to some degree stressing. For example, around one in each
three individuals, beyond sixty-five a year old, a fall, at any
rate once per year, and these are the main source of
hospitalization for this age bunch. Another very concerning
part of falls, among the older, is their hesitance to looking for
treatment in the wake of enduring a physical issue. In addition,
the monetary effect of falls was assessed in 2000 to be $US19
billion in the US just. The entirety of this is considerably
increasingly pertinent when one thinks about that the quantity
of elderly individuals (over 60 years of age) on the planet is
required to increment from 841 million out of 2013 to in
excess of 2 billion of every 2050. The past discoveries, stress
the need for social insurance suppliers to concentrate on
measures to lessen the hazard and seriousness of falls-related
wounds. Programmed fall recognition frameworks are a
significant part in this exertion and are a momentum
significant examination theme [3].
In light of a portion of these issues, numerous specialists
are concentrating their endeavors on cell phone-based
applications. Truth be told, the expanding fame of cell phones
makes them an appealing stage for the improvement of new
fall discovery frameworks. In addition, cell phones are very
much acknowledged, even among the older populace, and the

effectively implicit correspondence offices, including, e.g.,
SMS (short message administration) and GPS (worldwide
position framework), makes them an ideal possibility for a
programmed fall recognition framework that covers the
location and correspondence stages. The expanding number of
inherent sensors, for example, accelerometer, spinner, and
magnetometer, is likewise exceptionally beneficial to analysts
[4].
A fall discovery framework must have the option to
classify, or distinguish a fall occasion with ordinary conduct
to diminish the bogus constructive alert troubling the older
individuals. Simultaneously, this framework has the capacity
of covering all fall for security prerequisite. Thus, how to
structure a discovery framework which can adjust these two
necessities is a difficult strategic. A fall recognition
framework is first planned not to diminish the event of fallen
however means to alarm when a fall occasion occurs. In any
case, fall locators have been shown to coordinate effect on the
decrease of fall dread. Truth be told, falls and dread of falling
isn't autonomous. A person who is as often as possible falls
gives off an impression of being trepidation of falling and this
dread a short time later may build the danger of experiencing
a fall. Dread of fall significantly negative effects on the
existence nature of old which can cause melancholy, exercises
constraint, social association diminishing, falling, lower life
quality. So. Smartphone based accelerometer sensor is used to
detect the fall which can definitely improve the safety of elder
people [5, 6].
C. Literature Review
After careful reviewing existing literature, we found and
short-listed the most relevant works with top accuracy as
shown in Table I.
TABLE I. EXISTING WORKS ON SMARTPHONE BASED FALL
DETECTION
Ref Type of
Sensor
Classifier
Used
Performance
(Accuracy) Limitation
[7]
3-axis
Acceleromet
er
SVM
(RBF) 90-98%
External (Non-
smartphone
based)
accelerometer
used
[8]
3-axis
Acceleromet
er
SVM 99.38%
Threshold based
detection is
often not robust
[9]
3-axis
Acceleromet
er + Micro-
controller
HMM 94.8%
Additional
microcontroller
circuitry used
[10]
Acceleromet
er +
Gyroscope
CNN 90%
EEG recordings
are used for
classification
D. Objectives
The purpose of this work is to study the current state of
design and implementation of fall detection methods.
Exploring current implementation of fall detection methods to
assess the results in order to achieve more accuracy on fall
detection by implementing similar methods. By archiving
more accuracy on fall detection means reducing more frequent
fall events on adults, children, specialized people in need of
treatment and sick people. The main aim of this report can be
summarized as:
II. MATERIALS AND METHODS
Two smartphones were used for recording the data for
both fall and non-fall. One is Xiaomi Redmi Note-5 and
another is Xiaomi Redmi Note-6 Pro. Xiaomi Redmi Note-5
contains the feature of Fingerprint, accelerometer, gyro-
scope, proximity sensor, etc. while Xiaomi Redmi note 6 pro
contains the feature of accelerometer, gyroscope,
magnetometer, temperature, etc.
A. Experiment and Dataset Collection
The raw data was taken at sampling frequency of 200 Hz.
The data were recorded for each unit along three perpendicular
axes. That is x-axis, y-axis and z-axis. The dataset contained
6000 samples collected from one subject which is male. Only
two males were involved in recording data. The data were
taken for approx. one minute. The age is around 24-25 years.
One of the males weight is 55 kilogram, height 5’3 inch and
another one is 66 kilogram,5’8 inch. First person has taken 14
falls and 20 non-falls and second person recorded 12 falls and
21 non-falls which in total combination gives 67 data for both
falls and non-fall. First person took the data reading for
approx. 1.5 minute which gave 10000 samples. The target was
to achieve the data for 6000 samples in 30 secs. So, the first
person cropped the data by keeping sampling frequency fixed
and acquired 6000 samples. Again, in the same manner the
second person took the data reading for one minute and
obtained 12000 samples. And it was cropped to achieve 6000
samples for each data set.
TABLE II. LIST OF EXPERIMENT TO RECORD FALL AND NON-FALL
Fall Patterns
Non
-
Fall Patterns
1.Upstair fall
2. Run and fall
3. Lean forward fall
4.Lean backward fall
5. Lateral right fall
6. Lateral left fall
7. Fall lateral right and sit up
from floor
8. Fall lateral left and sit up
from floor
9. Fall from a chair straight
10. Downstairs fall
11. Chair right fall
12. Chair left fall
13. Bed right fall
14. Bed left fall
1. Chair lean left
2. Chair lean right
3. Jumping
4. Lie down and stand up from bed
5. Lying down left
6. Lying down right
7. Lying down straight
8. Lying to standing position
9. Run
10. Sitting down on a chair
11. Sitting on a chair
12. Squat down
13. Standing (chest)
14. Standing (pant pocket)
15. Standing to lying position
16. Standing up from a chair
17. Standing up from a floor
18. Walk
19. Walking downstairs & Upstairs
Fig. 1. Steps used in the proposed model for fall and non-fall classification
using MATLAB pattern recognition toolbox.

B. Proposed Model
We have followed a simple step-by-step procedure to collect,
process and classify acquired data with the built-in
smartphone accelerometers. The model is depicted in figure 1.
C. Data Preprocessing
After taking the data sample of 67 events both fall and non-
fall. We have cropped/selected 6000 samples out of 10,000+
samples within a duration of 120 seconds + which was also
eventually cropped to 30 seconds corresponds to the 1st 6000
samples. There was another scenario where we took 10,000+
sample within a duration of 120 seconds which we selected
6000 sample corresponds to 30 seconds, but we did take 1st
6000 samples rather we took samples according to our most
important activities within 10,000+ sample with a difference
of 6000 sample example. If our event took place from 2000
sample to 8000 sample, we selected that part only.
D. Feature Extraction
The procedure of feature extraction is helpful when we
need to diminish the quantity of assets required for handling
without losing significant or applicable data. Feature
extraction can likewise decrease the measure of excess
information for a given examination. We used four feature
extraction for classifying the data: maximum, minimum,
average and variance.
E. Classification
For classification neural network pattern recognition app is
used in MATLAB. The app classifies the inputs into set of
target categories. the algorithm tries to label input into two
distinct classes: fall and non-fall; hence it is called binary
classification. Neural network can learn from the data so that
patterns can be recognized, and classified. Further it is tested
to evaluate its performance and overall accuracy.
III. RESULTS AND DISCUSSION
A. Raw Data Recordings
Table-II shows an example of the difference between fall and
non-fall events for both subjects in terms of minimum and
maximum accelerometer readings in . Fig. 2 shows few
sample recordings of fall and non-fall for both subjects.
During fall, there is a sudden spike occurs in the acceleration
which is an important criterion to differentiate fall from non-
fall.
TABLE III. SAMPLE RECORDING OF ACCELEROMETER (RAW DATA)
WITH MINIMUM AND MAXIMUM VALUES
Events Minimum Maximum
Fall Events
Sub 01: Walk and Fall Forward -39.4827 58.8619
Sub 01: Fall Left and Stand Up -37.3401 20.8167
Sub 01: Walk and Fall Left -13.4729 34.9576
Sub 02: Run and Fall -63.4928 20.8167
Sub 02: Lateral Left Fall -22.9578 13.3670
Sub 02: Lean Forward Fall -40.8042 11.3642
Non-Fall Events
Sub 01: Stand Up from floor -5.6869 8.4864
Sub 01: Upstairs -9.1777 16.5592
Sub 01: Rickshaw -5.5465 6.6131
Sub 02: Stand to lying position -4.2165 2.1870
Sub 02: Jumping -48.8785 11.7002
Sub 02: Squat down -3.0502 1.5669
However, simply detecting high amplitude sudden spikes
may be confused with artifacts and intentional sudden
movement of the subjects. Therefore, instead of a simple
threshold based approach, a neural network based machine
learning classifier has been used with six statistical features
in order to detect fall from non-fall events. As the network is
trained with supervised mode, the reliability of the
classification is supposed to be higher than simple amplitude
threshold based detection of fall.
Fig. 2. Smartphone embedded accelerometer readings during different
experiments of fall and non-fall for two subjects (1st Subject’s recordings in
the upper trace and 2nd Subject recordings in the lower trace).
Fig. 3. Confusion matrices for training, validation, testing, and finally
combined with all.
B. Confusion Matrix and ROC Curve
The classification performances for training, validation,
testing and combined one have been shown in terms of a
confusion matrix in Fig. 3. In addition, Receiver Operating
Characteristic (ROC) curve displays true positive rate
(sensitivity) w.r.t false positive rate (1-specificity) of a
classifier with different thresholds. An ideal classifier should
have a curve at the top-left portion of the graph (well above
the line from a random predictor). Our classifier’s
performance is shown in terms of such ROC curve in figure 4.

C. Cross Entropy Performance
The final mean-square error was found small while the test set
error and the validation set error have similar characteristics.
Referring to figure 5, it is seen that no significant overfitting
has occurred at iteration 12 (where the best validation
performance occurs).
Fig. 4. ROC curves for training, validation, test and finally combined with
all.
Fig. 5. Performance of the classifer in terms of cross-entropy w.r.t epoch
numbers during training, validation and testing. The best performance
achieved at 12th epoch with minimum cross-entropy value.
D. Discussion
A total of 6000 samples of 67 events were selected with six
features (as input) and it consists of two classes where binary
01 (class-1) represents fall and 10 (class-2) represents Non-
fall. For validation and testing, by default 70% data were used
for training, 15% data for validation and rest 15% data were
used for testing while a default number of hidden neurons (i.e.
ten) was selected for the network architecture in the toolbox.
Training stops automatically when generalization stops
improving , as indicated by an increase of cross-entropy error
of the validation samples and training multiple times will
generate different results due to different initial conditions and
sampling. The limitations of this work are as follows: only two
subjects participated in fall events, the classification was done
offline although in real application online processing is
required for real-time fall detection, we could not compare our
method with others as the data were generated by us and yet
not shared in the research community, advanced deep learning
or CNN classification have not been applied, etc.
IV. C
ONCLUSION
A. Summary
Elderly people, children and sick people are among the
group of people where they are risk of any type of fall which
could be life threatening. Nowadays elderly people are mostly
living alone which increases the events related to fall as there
are no people around them to help in case of emergency.
Accidents can be reduced by using accurate fall detection
method which eventually helps to reduce injuries and loss of
life. In this work, we used smartphone-based fall detection
method using the features of tri-axial acceleration values
obtained from the built-in smartphone accelerometer. We
have recorded 67 activities like sitting, walking, standing,
lying, running, etc. using the app Physics Toolbox Sensor
Suite. Taking 67 events and 4 features as input, 01 as fall and
10 (binary) as non-fall as the target output for the Neural
network, we have analyzed Confusion matrix, ROC,
Performance, Error histogram etc. Our Overall accuracy is
about 93% which is quite acceptable for such purpose.
B. Future Works
As our project is offline based where we took live data’s
of falls and non-falls and distinguish between them by using
tri-axial accelerometer and classify them correctly as a result
to reduce false alarm. So, we want to work further on online
base process where we will create our own Fall Detection app
using tri-axial accelerometer. For instance, if a person falls,
he/she can confirm the fall by using the app where the
emergency center or caregivers can trace the positions of that
particular person by a text message for immediate assistance.
R
EFERENCES
[1] Shomir Chaudhuri, Hilaire Thompson, and George Demiris. Fall
Detection Devices and their Use with Older Adults: A Systematic
Review. J Geriatr Phys Ther. 2014 Oct-Dec; 37(4): 178-196.
[2] Mohammad Ashfak Habib, Mas S. Mokhtar, Shahrul Bahyah
Kamruzzaman, Kheng Seang Lim, Tan Maw Pin, and Fatimah
Ibrahim. Smartphone-Based Solutions for Fall Detection and
Prevention: Challenges and Open Issues. Sensors (Basel). 2014 Apr;
14(4): 7181–7208.
[3] Sabel N. Figueiredo1, Carlos Leal1, Luís Pinto1, Jason Bolito2 and
André Lemos2. Exploring smartphone sensors for fall detection.
Figueiredo et al. mUX: The Journal of Mobile User Experience (2016)
5:2.
[4] Tran Tri Dang, Hai Truong, Tran Khanh Dang. Automatic Fall
Detection using Smartphone Acceleration Sensor. (IJACSA)
International Journal of Advanced Computer Science and
Applications, Vol. 7, No. 12, 2016.
[5] fossbytes.com/which-smartphone-sensors-how-work (accessd on 19
th
June, 2020)
[6] https://gizmodo.com/all-the-sensors-in-your-smartphone-and-how-
they-work-1797121002 (accessd on 19th June, 2020)
[7] Aziz O, Klenk J, Schwickert L, Chiari L, Becker C, Park EJ, et al.
(2017), ‘Validation of accuracy of SVM-based fall detection system
using real-world fall and non-fall datasets.’, PLoS ONE 12(7):
e0180318. https://doi.org/10.1371/journal.pone.0180318
[8] Jin-Shyan Lee, ‘Development of an Enhanced Threshold-Based Fall
Detection System Using Smartphones with Built-In Accelerometers’,
IEEE Sensors J., vol.19, no. 12, pp 8295-8297, Sep. 2019.
[9] Jin Wanga, Ronghua Chena, Xiangping Suna, Mary F.H Shea,
Yuchuan Wub, ‘Recognizing Human Daily Activities From
Accelerometer Signal’, Procedia Engineering, vol. 15, pp. 1783-1785,
Aug. 2011.
[10] Shalu Chaudhary, Sachin Taran , Varun Bajaj, ‘Convolutional Neural
Network Based Approach Towards Motor Imagery Tasks EEG Signals
Classification’, IEEE Sensors J., vol. 19, no. 12, pp 4496, Jun. 2019.