See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/356206581

Human Fall Detection with Wearable Sensors Using ML Algorithms

Conference Paper · October 2021

DOI: 10.1109/ICOSEC51865.2021.9591912

CITATIONS READS

7 45

2 authors, including:

Anju Pillai
Amrita Vishwa Vidyapeetham
74 PUBLICATIONS 533 CITATIONS

SEE PROFILE

All content following this page was uploaded by Anju Pillai on 16 June 2022.

The user has requested enhancement of the downloaded file.

 Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6

Human Fall Detection with Wearable Sensors Using

ML Algorithms
2021 2nd International Conference on Smart Electronics and Communication (ICOSEC) | 978-1-6654-3368-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICOSEC51865.2021.9591912

Aki Sowmya Anju S. Pillai

Department of Electrical and Electronics Engineering Department of Electrical and Electronics Engineering
Amrita School of Engineering, Coimbatore Amrita School of Engineering, Coimbatore
Amrita Vishwa Vidyapeetham, India. Amrita Vishwa Vidyapeetham, India.
[email protected] [email protected]

Abstract—In the healthcare systems, wearable issue has a major disadvantage of camera-based systems,
technology is enhanced by continuously monitoring particularly in the case of fall detection at the home user [5].
people's physical activities and behaviors in daily life. A
Another technique that can be used is to observe floor
wearable system has two technologies, one based on
vibration and abnormal noise, while the system attempts to
sensors and the other based on cameras. The use of
determine by comparing the received data with the previously
machine learning (ML) in health care applications, such as
gathered vibration and noise signals [6]. Using the light-weight
human fall detection, has become an active area of
hand sensors like tri-axial accelerometer and gyroscope, are
research. This research work has proposed a sensor based
utilized to detect falls, body position, hand movement, and
human fall detection system, which detect a fall by using
gesture recognition which improve the accuracy and energy
sensor-based technology other than the vision-based
productivity can achieve more reliable fall detection [7][8].
technology because most people are concerned about their
privacy very seriously. The overall performance of the A fall happens when an individual loses their equilibrium
proposed scheme was examined by using four machine and can't stand upstanding. When a youngster loses their
learning algorithms namely, Support Vector Machine equilibrium, then they have the solidarity to recapture their
(SVM), k-nearest neighbour’s (KNNs), Decision Tree (DT), equilibrium. Yet, when an older individual loses balance, it is
and Random Forest (RF). The proposed methodology hard for him/her to recover because he/she is physically fragile
achieves the best accuracy for human fall detection, i.e., at that age. This research work has proposed a human fall
98% , by using ensemble technique as RF classifier. detection system (FDS) with a wearable tri-axial accelerometer
and gyroscope sensor using four machine learning algorithms.
Keywords—Fall Detection, Wearable Sensors, Machine The main aim of the project is to compare ensemble, eager and
Learning, Health care applications lazy algorithms and resulting in a suitable feature extraction.

I. INT RODUCT ION The rest of the paper organized as follows. Section 2
discuss the previous studies in the area of fall detection system.
As humans age, the human body will degenerate due to Section 3 describes the methodology of the system and in
overwork, which means that the jobs and activities those section 4, the result and analysis of fall detection system using
people do every day can damage their physical function and machine learning algorithms is presented. Finally, section 5
cause falls. Most people may think that accidental falls are concludes the paper.
common [1]. But for the elderly, the accidental fall is
incredibly dangerous as it causes serious harm to the elderly. II. LIT ERAT URE SURVEY
The annual probability of falls for people over 65 has A human can fall in any type of situation while
increased from 28% to 40% [2]. A human fall detection walking, sleeping, or even when people are in a conscious
system plays a significant role in improving the quality of life situation where they may trip. There were many types of
in older people and providing them with life assistance. sensors used to detect the fall, including accelerometer, tilt
sensor, magnetometer, and gyroscope sensors. Most of the
Fall detection has been widely researched through many studies used acceleration measurements based on fall
precise methods. One method is computer vision, visual-based properties [9][13]. The accelerometer and gyroscope
systems like video surveillance are a traditional non-portable combining, which changes the client's body direction and
fall detection system [3]. Difficulty in collecting data from body position, is analyzed throughout weight and effect
cameras, a three-dimensional method based on neural condition, which might improve the accuracy.
convolutional networks has developed to detect falls to build a
training data set which is limited [4]. Accordingly, the privacy

978-1-6654-3368-6/21/$31.00 ©2021 IEEE 1092

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on June 16,2022 at 06:10:39 UTC from IEEE Xplore. Restrictions app
 Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6

Krupitzer et al. showed that when the sensors were publicly available dataset for testing and training the system to
put along the longitudinal hub of body (chest, waist, head) can differentiate fall or not-fall actions [20]. This methodology
obtain better results, and they compared it to other positions consists of four main steps: data acquisition, data processing,
(wrist, thighs) as well in their work [10]. The movement of the feature extraction and fall detection shown in Fig.1.
shaft during the fall was more stable, but dedicated devices
need to be used on unusual body parts, which causes
inconvenience. Hence, different investigations utilized items
(such as smartphones in thighs, smartwatches on wrists) that
generally do not disturb the users because they are already
using it [12][2]. Nonetheless, individuals will in general try to
remove these devices at home, making the FDS unu sable.
Valcourt et al. used sensors of a cell phone such as, gyroscope
Fig.1. Methodology for Fall Detection System
and accelerometer to determine the area of phone on the
subject’s body and find patterns related to falls. To detect a
fall, they used three main locations, texting with an accuracy A. Data Acquisition
of 95.8%, shirt front pockets with an accuracy of 83.3%, a side The initial step of the planned methodology is to
trouser pocket with an accuracy of 87.5% and the overall gather data which will be utilized in additional stages once the
accuracy was 81.3%[11]. data processing is completed. It is a demanding task to collect
the real-time data from people. Various datasets are
Generally, a fall detection system algorithm has two accessible, yet most of these activities that are only performed
approaches, a threshold-based and machine learning-based by either elder people or young people. For an effective
algorithms. In threshold-based algorithms, calculating features human fall detection, the dataset should contain daily activities
based on acceleration measurements and compared them with and falls. Hence, we chose the Sisfall dataset, which has
thresholds to identify whether a fall has existed or not [14]. activities for both young and elderly people.
They are usually sufficient, but they tend to generate a false
positive result, especially when falling, such as when sitting The Sisfall dataset has high quality information and
down suddenly. To follow this, a simple pattern recognition we used several criteria to assess this quality, specifically the
algorithm and postures are added to these studies to recognize scale of the dataset and therefore the diversity of subjects in
changes in body posture and activity level, which increases the terms of gender, range of falls and age per subject and ADL.
reliability of recognition and has low computational The SisFall dataset utilizes three-axis accelerometers and a
complexity [15][16]. Nonetheless, it might in any case come three-axis gyroscope which is situated at the waist of the
up short during explicit falls and activities of daily living human body with a sampling rate of 200 Hz.
(ADL).
The dataset contains 23 young people (ages 19-30)
Typically, most of the studies used the supervised experienced 15 kinds of falls and 19 sorts of ADLs including
machine learning algorithms, they were common algorithms falls.15 older people performs the similar ADLs on real-time.
such artificial neural network (ANN) [17], as support vector Five trails were carried out for each activity apart from
machines (SVM) [19][16], K-nearest neighbour (KNN), jogging and walking activities, which includes a single
Hidden Markov model (HMM) [18] etc. Whereas, the attempt. So SisFall [20] contains a complete of 4505 records,
unsupervised learning algorithm uses a cluster technique to includes 1798 falls and 2707 ADLs. In this work, we excluded
automatically group the actual data, rather than training the the data provided by the older people since they did not
predecessors to label the data. perform the fall activities and also eliminated the data from
three young people due to lack of records, leaving us with data
In literature, the fall detection can be done using non- from 20 subjects. There are 3400 recorded data, composed of
wearable sensors and the vision-based sensors. And the 1500 falls and 1900 ADL, which makes the data most
features are calculated by the threshold-based approach which adjustable
produces false alarms. In this proposed work, wearable
sensors such as, tri-axial accelerometer and gyroscope are B. Data Processing
used, and the data is taken from the publicly available dataset
Sisfall dataset [20]. From the data, the most relevant features After collecting the data, the subsequent step is to pre-
process the data to eliminate unwanted noise from the signal
are calculated. The calculated features were given to the eager,
lazy, ensemble machine learning algorithms which can reduce so that the machine learning algorithm can perform better
the false alarms. classification. We use a fourth-order Butterworth filter with
infinite impulse response (IIR) detailed [20], and therefore the
cut-off frequency Fc = 5 Hz since it is basic and requires little
III. M ET HODOLOGY
computation.
The proposed human fall detection system follows an
effective methodology to distinguish the falls. Sisfall is a

978-1-6654-3368-6/21/$31.00 ©2021 IEEE 1093

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on June 16,2022 at 06:10:39 UTC from IEEE Xplore. Restrictions app
 Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6

C. Feature Extraction creates model in the case of hyperplane to call it as

After data processing on the original signal, the consequent eager learning algorithm. And it finds the best
hyperplane that will maximize the margins between
step helps to extract features for classification. We have
each class. It is normally used to distinguish the fall or
tendency to extract seven features from the data processing.
non-fall by using wearable devices.
These attributes incorporate the mean, minimum amplitude,
standard deviation, sum of vector magnitude, maximum  Decision tree: It has the shape of tree, where every node
amplitude, peak difference and orientation of trunk. These could be a decision, resulting in additional precise data
features are listed in Table I, along with the mathematical for classification. Throughout the training process the
expressions. Every component was extracted from sensors tree was created which leads to overfitting and it is also
which includes gyroscope and two accelerometers. For every called as eager learning algorithms.
sensor which have three axes x, y, z axes, each feature will be  Random Forest Classifier: It uses several decision trees
multiple with number of axes the vector size of these features in an equal manner. To determine the likely category,
were 21 for a sample. Thus, we got a final feature with three each decision tree was trained on subset of the data,
sensors with a vector size of 63. which can reduce the overfitting caused by the decision
tree. As it was using many other algorithms, referred as
ensemble learning algorithms.
TABLE I. LIST OF EXTRACTED FEATUES
Fe ature s Equation IV. RESULT S
To evaluate the model and the performance of the
Mean proposed algorithm, analysis based on the following metrics,
[ ]
were carried out: sensitivity, confusion matrix, specificity, and
accuracy, which are defined below.
Maximum Amplitude [ ]

Sensitivity measures the capability of the system to

Minimum Amplitude [ ] detect a fall. The specificity is utilized to recognize a fall just
when they happen and accuracy is the capacity of the
Standard Deviation √ [ ] [ ] [ ] framework to separate among falls and non-falls.

TABLE II. RESULTS OF FALL DETECTION SYSTEM USING MACHINE

Peak Difference [ ] [ ] LEARNING ALGORITHM
Algorithms Sensitivity Specificity Accuracy
Sum of Vector Magnitude [ ] [ ] [ ] [ ]

SVM 96.4% 87.1% 92.76%

Orientation of Person [ ] [ ]
T runk ( ( )) DT 96.6% 97.26% 97%
[ ]

KNN 97.6% 99.1% 97.46%

D. Fall Detection Algorithm

RF 98.2% 99.5% 98.2%
After extracting the features, the subsequent step was to
categorize the activities as a fall or non-fall. A human fall
detection using a binary form either 0 or 1 to classify the
activity as falling or not falling is adopted. Therefore, we Table II shows the analysis of the human fall
divided the entire data set into two classifications: class -1 for detection system among the four chosen machine learning
non-fall actions and class-2 for fall actions. Feature extraction algorithms and results of implementation of an effective fall
was carried out for fall activity and ADL activity. After the detection system. For the random classifier algorithm, the
features have been labelled, four machine learning algorithms metrics specificity and sensitivity both exceed 98%.
were implemented. Generally, ensemble techniques are superior to other
algorithms since they utilized various machine learning
 K-Nearest Neighbors: It is an effective and simple algorithms, even though the overall performance development
algorithm. The data is sorted by highest decision and comes from a high rate of additional assets. Support vector
has the strongest representative class among the k machines are harder to differentiate between actions, even if
nearest neighbours. This algorithm was often called as changes in certain hyper parameters are done.
lazy learning algorithms since most of the work was
V. CONCLUSION
done before making the predictions.
In this work, an implementation of the human fall
 Support Vector Machines: These are widely utilized in detection system using sensor-based technology is carried out.
several task. It plays a big role in the learning phase and The system uses simple and computational features extracted

978-1-6654-3368-6/21/$31.00 ©2021 IEEE 1094

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on June 16,2022 at 06:10:39 UTC from IEEE Xplore. Restrictions app
 Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6

from a publicly available dataset: Sisfall. And it was a reliable Communications and Informatics (ICACCI), 2018, pp. 638-643, doi:
methodology because the test was carried out in an outsized 10.1109/ICACCI.2018.8554582.
[15] S. Abdelhedi, R. Bourguiba, J. Mouine and M. Baklouti, "Development
dataset which consists of thousands of daily activities and of a two-threshold-based fall detection algorithm for elderly health
falls. The results are obtained by comparing four machine monitoring", Proc. IEEE Res. Challenges Inf. Sci. (RCIS), pp. 1-5, Jun.
learning algorithms and the ensemble learning algorithm is 2016.
found to have the highest accuracy of 98%. It confirms that [16] A. Kurniawan, A. R. Hermawan and I. K. E. Purnama, "A wearable
the ensemble technique outperforms both eager and lazy device for fall detection elderly people using tri dimensional
accelerometer", Proc. IEEE Int. Seminar Intell. Technol. Appl. (ISITIA),
algorithms. Future work may include integration of the deep pp. 671-674, Jul. 2016.
learning algorithms to reduce the number of steps and features [17] A. Purushothaman, "Fall detection system using Artificial Neural
extracted. Network", 2nd International Conference on Inventive Communication
and Computational T echnologies (ICICCT 2018), 2018.
REFERENCES [18] L. T ong, Q. Song and Y. Ge, "HMM-Based Human Fall Detection and
Prediction Method Using T ri-Axial Accelerometer", IEEE Sensors
Journal, vol. 13, pp. 1849-1856, May 2013
[1] P. Pierleoni, A. Belli, L. Palma, M. Pellegrini, L. Pernini and S. Valenti, [19] S. Badgujar and A. S. Pillai, "Fall Detection for Elderly People using
"A high reliability wearable device for elderly fall detection ", IEEE Machine Learning," 2020 11th International Conference on Computing,
Sensors J, vol. 15, no. 8, pp. 4544-4553, Aug. 2015. Communication and Networking T echnologies (ICCCNT), 2020, pp. 1-
[2] M. A. Habib, M. S. Mohktar, S. B. Kamaruzzaman, K. S. Lim, T. M. Pin 4, doi: 10.1109/ICCCNT 49239.2020.9225494.
and F. Ibrahim, "Smartphone-based solutions for fall detection and [20] Angela Sucerquia, Jose David Lopez and Jesus Francisco Vargas-
prevention: Challenges and open issues", Sensors., vol. 14, no. 4, pp. Bonilla, "SisFall: A Fall and Movement Dataset", Sensors, vol. 17, no.
7181-7208, 2014. 3390/s17010198, 2017.
[3] K. Vinaya Sree and Dr. Jeyakumar G., “A Computer Vision Based Fall
Detection T echnique for Home Surveillance”, In the Proceedings of the
International Conference on ISMAC in Computational Vision and Bio -
Engineering (ISMAC-CVB),Advances in Intelligent Systems and
Computing, vol. 1108., Cham, 2019.
[4] Stone E, Skubic M (2014) Fall detection in homes of older adults using
the microsoft kinect. IEEE J Biomed Health Inform 19(1):290 –301.
https://doi.org/10.1109/JBHI.2014.2312180
[5] L. Van Zoonen, "Privacy concerns in smart cities", Government
Information Quarterly, vol. 33, no. 3, pp. 472-480, 2016.
[6] Y. Zigel, D. Litvak, I. Gannot, "A Method for Automatic Fall Detection
of Elderly People Using Floor Vibrations and Sound-Proof of Concept
on Human Mimicking Doll Falls," IEEE T ransactions on Biomedical
Engineering, vol. 56, no. 12, pp. 2858-2867, Dec. 2009
[7] Q. Li, J. Stankovic, M. Hanson, A. Barth, J. Lach, G. Zhou, "Accurate,
Fast Fall Detection Using Gryoscopes and Accelerometer-Derived
Posture Information," Sixth International Workshop on Wearable and
Implantable Body Sensor Networks 2009, pp. 138-143, Jun. 3-5, 2009
[8] K. Nguyen, I. Chen, Z. Luo, S. Yeo, H. Duh, "A Wearable Sensing
System for T racking and Monitoring of Functional Arm Movement,"
IEEE/ASME T ransactions on Mechatronics, vol. 16, no. 2, pp. 213-220,
Apr. 2011
[9] R.M. Gibson, A. Amira and N. Ramzan, "Multiple comparator classifier
framework for accelerometer-based fall detection and
diagnostic", Applied Soft Computing, vol. 39, pp. 94-103, 2016
[10] C. Krupitzer, T. Sztyler, J. Edinger, M. Breitbach, H. Stuckenschmidt
and C. Becker, "Hips Do Lie! A Position-Aware Mobile Fall Detection
System", 2018 IEEE International Conference on Pervasive Computing
and Communications (PerCom), pp. 1-10, March 2018.
[11] L. Valcourt, Y. De La Hoz and M. Labrador, "Smartphone-based human
fall detection system", IEEE Latin AmericaTransactions, vol. 14, no. 2,
pp. 1011-1017, 2016.
[12] T . Vilarinho et al., "A Combined Smartphone and Smartwatch Fall
Detection System," 2015 IEEE International Conference on Computer
and Information T echnology; Ubiquitous Computing and
Communications; Dependable, Autonomic and Secure Computing;
Pervasive Intelligence and Computing, 2015, pp. 1443 -1448, doi:
10.1109/CIT /IUCC/DASC/PICOM.2015.216.
[13] Yanyan Zhuang, J. Baldwin, L. Antunna, Y. O. Yazir, S. Ganti and Y.
Coady, "T radeoffs in cross platform solutions for mobile assistive
technology," 2013 IEEE Pacific Rim Conference on Communications,
Computers and Signal Processing (PACRIM), 2013, pp. 330-335, doi:
10.1109/PACRIM.2013.6625498.
[14] M. R. S. Sai, S. V. K. Bondada, V. S. S. Arjun and V. Bhavana,
"Implementation of Compact Wearable Fall Detector for the
Elderly," 2018 International Conference on Advances in Computing,

978-1-6654-3368-6/21/$31.00 ©2021 IEEE 1095

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on June 16,2022 at 06:10:39 UTC from IEEE Xplore. Restrictions app
View publication stats