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

net/publication/346238191

Fall Detection for Elderly People using Machine Learning

Conference Paper · July 2020

DOI: 10.1109/ICCCNT49239.2020.9225494

CITATIONS READS
33 2,991

2 authors, including:

Anju Pillai
Amrita Vishwa Vidyapeetham
74 PUBLICATIONS 557 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.

 IEEE - 49239

FALL DETECTION FOR ELDERLY PEOPLE

USING MACHINE LEARNING
Sejal Badgujar 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 VishwaVidyapeetham, India. Amrita VishwaVidyapeetham, India.
[email protected] [email protected]

and waist are used to get the data input. Features are
Abstract—Health is the major worrisome point whose
impalpability increases with increase in the age. Thus, taking calculated by using sensor readings to get some meaningful
care of elders is very important responsibility. In such data out of the raw data. Falls also can be detected by using
scenario, technology is helping people by providing living ambient sensors like IR sensors or movement based sensors.
assistance. One of the major causes of health degradation or
death of elders is ‘fall’. In this paper, a fall detection system is Major hurdle to develop precise fall detection
proposed based on machine learning. The system detects falls system is false alarms i.e. alerting fall when there is a fall like
by classifying different activities into fall and non-fall actions activity but not exactly a fall. Most of the researches are
and alert the relative or care taker of the elderly person in case focused on reducing false alarms and improving accuracy of
of emergency. The dataset SisFall with variety of activities of
multiple participants is used to calculate features. Machine the fall detection system. The main aim of studies is to detect
learning algorithms SVM and decision tree are used to detect fall in the daily life activity situations with high accuracy.To
the falls on the basis of calculated features. The system classify the activity into fall or not fall criteria, threshold
acquires accuracy up to 96% by using decision tree algorithm. based algorithms also can be used. But if sensor detects an
Index Terms—Fall detection, Gait analysis, Machine unusual activity, then the probability of false alarm increase
learning, Activities of daily living, Elderly care, Health care unlike machine learning based approach.
systems.
In this work, a fall detection system is proposed
I. I NTRODUCTION which monitors elderly people in real-time. The system uses
open source available dataset SisFall which has recorded Gait
Now-a-days elderly people live alone at home because of
data by using Tri-axial accelerometer [15]. By using machine
poor conditions, different working culture of people and due
learning algorithms, falls are detected after calculating
to many other reasons. According to World Health
various features. Two different machine learning algorithms,
Organization (WHO) studies, falls causes many accidental
SVM and decision tree are implemented and compared for
deaths. Greatest number of fatal falls is seen in adults older
better accuracy and performance.
than 65 leading to health problems or injuries [1]. Thus,
elderly people require an attention at the times of The rest of the paper is organized as follows: section
emergencies at their residence because they cannot call for 2 discusses about the literature survey in the area of fall
help due to lack of technology access in rural areas or due to detection followed by methodology of implementation in
their physical conditions. To improve quality of life of old section 3. The results are presented in section 4 and the paper
aged people and to provide living assistance to them, is concluded in section 5.
automatic fall detection systems are in place. The risk of II. L ITERATURE SURVEY
falling is high among older people, individuals with
There are no standard sets for fall detection, to get perfect
Parkinson's disease or patients in rehabilitation units [3].
fall detection system, in terms of sensors used, features
Main reasons for occurrence of falls are physical factors like
extracted, and machine learning algorithm with better
of muscle weakness, posture, Gait balance, vision, due to old
performance. Fall detection system can be implemented by
age, or psychological factors or environmental factors etc.
using cameras, wearable sensors or ambient sensors. When
Falls are the major cause of injuries and hip fractures [4]. If
cameras are used in the system, image processing
immediate aid is not provided, it may lead to death. To
algorithms are used but it has been seen that 24% of the
ensure proper treatment and care of the elderly people, fall
falls are not getting detected [3]. Ambient based sensors
detection system plays important role.
like motion detector or passive infrared (PIR) sensors also
Various researches have been done on the fall can be used in fall detection system detection [7]. But these
detection system. There are no fixed criteria on basis of are limited to particular area in which sensors are
sensors used or on calculated features or on algorithms used implemented.
to classify. Image processing techniques are used after
Wearable sensor based fall detection systems are more
capturing images from camera to detect falls. Wearable suitable for elderly people because it can detect the fall any
sensors like accelerometer, gyroscope at knee, wrist, neck,
time and any place unlike vision based and ambient based
11th ICCCNT 2020
July 1-3, 2020 - IIT - Kharagpur
Authorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL Kharagpur, India Downloaded on June 16,2022 at 05:57:55 UTC from IEEE Xplore. Restrictions apply.
OF ENGINEERING.
 IEEE - 49239

fall detection which are restricted to the house or particular A. Data Acquisition
indoor environment. Also wearable sensor is lesser in cost
The SisFall [15] dataset is considered which consists
than that of camera or PIR sensors. In literature, sensors are
of data collected with the help of ADXL345 accelerometer
used in necklace [8], subjects’ head band, chest band, waist
from the waist band. Tri- axial accelerometer ADXL345 is
band, right wrist band, right thigh band, and right ankle
used to get the acceleration values along three axes x, y and z.
band [9], waist band [10], etc. The pendant connected to
It is energy efficient as it can work with low power mode.
mobile phone via Bluetooth and phone communicates with
concerned person. The problem with this system is, phone
should be in 100 meters range of pendant; and activities
like front bending or front falls are not considered due to
false interpretation of fall [8]. Too many sensors also lead
to miss prediction of the accurate activity. Wearable sensor
accelerometer is better than camera based and ambient
based techniques.
Fall can be detected using two techniques after
collection of data from sensors and feature calculation. One
is threshold based, if reading of the sensor is above
particular threshold, it can be categorized as fall. In this Figure 2 Waist belt with sensor position
technique, threshold for each calculated feature is different
and many false alarms may be generated. If machine In this work, six young adults and two elder’s
learning classification is used, the calculated features can acceleration data is considered. Adult participants have
be tested on a pre-trained model with high accuracy and performed 19 daily life activities and 15 fall activities, only
guarantee lesser false alarms. one elder performed all the activities. 1 lakh data is considered
In the current work, accelerometer data from a wearable for training as partial dataset and 40,000 acceleration data
sensor is used, which is already measured for different samples are used for testing purpose. (Young participant’s
activities in SisFall dataset [15]. Most relevant features are code- SA02, SA05, SA10, SA15, SA20, SA23, Adult
calculated. Machine learning models of SVM and decision participant’s code- SE06,SE10 ).
tree are trained and tested.
In literature, these features are used for threshold based
fall detection, and the proposed work implements machine
learning approach on same features and provides improved
accuracy. Also, none other paper have used decision tree
algorithm for training and testing this dataset. The current
work implements SVM and decision tree algorithm and
compares its performance
III. METHODOLOGY

Fall detection system consists of the following steps as

in Fig. 1 and is explained below:

Figure 1 System flow

Figure 3 SisFall Dataset

11th ICCCNT 2020

July 1-3, 2020 - IIT - Kharagpur
Authorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL Kharagpur, India Downloaded on June 16,2022 at 05:57:55 UTC from IEEE Xplore. Restrictions apply.
OF ENGINEERING.
 IEEE - 49239

D. Fall Detection Algorithms

B. Data Processing The machine learning algorithms used are Support
Data processing is done by using 4th Order butter- Vector Machines (SVM) and Decision Tree. SVM is an
worth filter with cutoff of 5Hz to the sensor data to remove algorithm that can be used for classification as well as
noise and unwanted glitches. This filter is used because it regression. SVM finds the hyper plane such that it has largest
has given similar results as that of more elaborated IIR and distance from nearest training data point of any class to
FIR filters at different frequencies [15]. Figure 3a) and 3b) provide good separation. For non-linear dataset, it can be
shows original and filtered data respectively. possible to find hyper plane to determine the classification by
using kernel function. Data points falling on each side of the
plane belongs to different classes.
In decision tree algorithm, data is represented as tree
like model with nodes and edges. Trees basically have the root
node, internal nodes and leaf nodes, but with real datasets,
different features get added. The decision tree is a
classification technique with the finite number of classes. Once
tree is ready with training data, decision rules are drawn, on
basis of that decisions for classification is taken. The reason
behind decision tree is classification model popularity is they
are easy to interpret, the results obtained can be more easily
understood. With decision tree algorithm, complex decision-
making process becomes simpler and solution of the problem
can be interpreted by decision makers.
These algorithms are compared on basis of fall
detection accuracy. The models are tested on 40,000 data
samples and accuracy is calculated. The most accurate model
is saved for further testing. It has been seen that,sum vector
Figure 3 a) Original data of X, Y, Z axes b) Filtered data magnitude (C1), sum vector magnitude horizontal plane (C2),
standard deviation magnitude horizontal plane (C8) and
C. Feature Calculation
standard deviation magnitude (C9) features gives higher
There are many features that can be calculated using
accuracy than all other features considered together.
accelerometer data, and which are significant. In this work,
we have considered the same features SisFall dataset [15]. But E. Alert
they have used threshold technique and this paper uses In case of fall, concerned person should be informed.
machine learning technique to cover all non-listed activity When real time data is collected and tested for fall, message
classification as well. The features with formula are or call should be initiated on fall detection.
mentioned in the Table 1. For one acceleration sample i.e.
IV. ALGORITHM IMPLEMENTATION
acceleration along all the three axes, all above features are
calculated. For 1 lakh data samples, 1 lakh features are The flowchart of implementation of algorithm is
calculated. shown in Fig. 4. Software implementation is done using
Table 1: Calculated features Python 3.7 and sklearn.

Figure 4 System implementation

11th ICCCNT 2020

July 1-3, 2020 - IIT - Kharagpur
Authorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL Kharagpur, India Downloaded on June 16,2022 at 05:57:55 UTC from IEEE Xplore. Restrictions apply.
OF ENGINEERING.
 IEEE - 49239

Organization, Geneva, 2008.

[3] De Backere, F., Ongenae, F., Van den Abeele, F., Nelis, J., Bonte, P.,
IV. OUTCOME ANDDISCUSSION Clement, E., Philpott, M., Hoebeke, J.Verstichel, S., et al.”Towards a social
To evaluate various trained models and to compute the and context-aware multi-sensor fall detection and risk assessment platform”,
Computers in biology and medicine 64,307/320, 2015.
efficiency of various algorithms, different parameters such as: [4] Khan, M.S, Yu, M, Feng, P, Wang, L, Chambers,”An unsupervised
confusion matrix, sensitivity, specificity, accuracy, training acoustic fall detection system using source separation for sound interference
time and prediction time are computed. suppression”,Signal processing 110, 199/210, 2015.
[5] Glen Debard, Greet Baldewijns, ToonGoedem, TinneTuytelaars and Bart
Sensitivity represents the actual positive cases which got Vanrumste, “Camera-based fall detection using a particle filter Glen”,IEEE,
predicted as positive (or true positive), Specificity represents 978/4244/9270, 2015.
[6] Rita Cucchiara,AndreaPrati,RobertoVezzani, "A multi-camera vision
the actual negatives, which got predicted as the negative (or system for fall detection and alarm generation“, Expert Systems, 24, 5, 2007.
true negative). Accuracy is ratio of predictions our model got [7] AhmetYazar, FatihErden, A. Enis Cetin,” Multi-sensor ambient assisted
right which is calculated by using average of sensitivity and living system for fall detection”, ResearchGate Conference Paper ,May 2014
specificity. Also confusion matrix is used to get true positive [8] Joseph Santiago, Eric Cotto, Luis G. Jaimesy, IdalidesVergara-Laurens,
“Fall Detection System for the Elderly”, IEEE, 978-1-5090-4228-9/17, 2017.
and true negative values. The accuracy is calculated with [9] AhmetTuranÖzdemirandBillurBarshan,” Detecting Falls with Wearable
selected features and is shown in Table II. Sensors Using Machine Learning Techniques” , Sensors, 14,
10.3390/s140610691, 2014.
[10] FalinWu, Hengyang Zhao, Yan Zhao, HaiboZhong,” Development of a
Wearable-Sensor-Based Fall Detection System”, International Journal of
Table II: Performance comparison of ML algorithms
Telemedicine and Applications, Volume 2015, Article ID 576364, 2014.
Algorithms Accuracy Training time Prediction [11]Diana Yacchirema, JaraSuárez de Puga, Carlos Palaua, Manuel
Esteve,”Fall detection system for elderly people using IoT and Big Data ”, The
time
9th International Conference on Ambient Systems, Networks and
SVM 84.17% 294.95 sec 84.71 sec Technologies , 2018.
[12]Tuan Nguyen Gia, Victor KathanSarkera, Igor Tcarenkoa, et.al., “Energy
Decision 95.87% 2.741 sec 0.02 sec efficient wearable sensor node for IoT-based fall detection systems”,
tree Microprocessors and Microsystems 56,34–46, 2018.
[13]HristijanGjoreski, MitjaLuštrek, Matjaž Gams, ” Accelerometer
Placement for Posture Recognition and Fall Detection”, Seventh International
Conference on Intelligent Environments, 2011.
The performance of decision tree algorithm is found to be [14]Yueng Santiago Delahoz , Miguel Angel Labrador,” Survey on Fall
better than SVM algorithm for the considered data. Several Detection and Fall Prevention Using Wearable and External Sensors”,
Sensors, 14, 19806-19842, 2014.
factors were considered for comparison and one among them [15]Angela Sucerquia, José David López ,Jesús Francisco Vargas-Bonilla,
is the ability to define and classify each attribute to each class. “SisFall: A Fall and Movement Dataset”, Sensors, 17, 3390/s17010198, 2017.
The computing time of decision tree is found to be less than [16] S. Vadivelu, Ganesan, S., Murthy, O. V. R., and Dhall, A., “Thermal
SVM. SVM works better out of the box, but decision trees imaging based elderly fall detection”, Lecture Notes in Computer Science, vol.
10118, 541-553, 2017
gives more insight into how the model works. Decision trees [17] A. Purushothaman, “Fall detection system using Artificial Neural
are great for their simplicity and interpretation, but it has Network”, in 2nd International Conference on Inventive Communication and
limitations in learning complicated rules and to scale to large Computational Technologies(ICICCT 2018), Hotel Arcadia, Coimbatore,
data sets. 2018.

V. CONCLUSION
This paper presents wearable sensor based fall detection
system, which are suitable for elderly people. The proposed
method uses machine learning algorithms to detect falls from
a set of daily living activities. Machine learning technique are
found better than the threshold method, as it gives less false
alarms due to pre-trained Gait patterns. The decision tree
gives higher accuracy than SVM as decision tree has the
ability to define and classify each attribute to each class
precisely. Also prediction time of SVM is greater than
decision tree which leads to a slower system. The models are
evaluated by using parameters such as: sensitivity, specificity,
accuracy and confusion matrix. Falls are appropriately
detected using decision tree algorithm with an accuracy of
96%. Further improvement in accuracy can be obtained by
training the models with large dataset and by identifying
optimal features.

REFERENCES
[1]World Health Organization, World Report on Ageing and Health, World
Health Organization, Geneva, 2015.
[2]WHO Global Report on Falls Prevention in Older Age, World Health

11th ICCCNT 2020

July 1-3, 2020 - IIT - Kharagpur
Authorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL
View publication stats
Kharagpur, India Downloaded on June 16,2022 at 05:57:55 UTC from IEEE Xplore. Restrictions apply.
OF ENGINEERING.