Fall Detection System for Elderly based on

2D LiDAR : A Preliminary Study of Fall Incident
and Activities of Daily Living (ADL) Detection
Herti Miawarni Tri Arief Sardjono Eko Setijadi
Dept. of Electrical Engineering Dept. of Electrical Engineering Dept. of Electrical Engineering
Faculty of Intelligent Electrical and Dept. of Biomedical Engineering Faculty of Intelligent Electrical and
Informatics Technology (F-Electics) Faculty of Intelligent Electrical and Informatics Technology (F-Electics)
Institut Teknologi Sepuluh Nopember Informatics Technology (F-Electics) Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia Institut Teknologi Sepuluh Nopember Surabaya, Indonesia
hertimiawarni.19071@mhs.its.ac.id Surabaya, Indonesia ekoset@ee.its.ac.id
sardjono@bme.its.ac.id

Wijayanti Dwi Arraziqi Agustinus Bimo Gumelar

Dept. of Arcitecture Dept. of Electrical Engineering Dept. of Electrical Engineering
Faculty of Engineering Faculty of Intelligent Electrical and Faculty of Intelligent Electrical and
Universitas Diponegoro Informatics Technology (F-Electics) Informatics Technology (F-Electics)
Semarang, Indonesia Institut Teknologi Sepuluh Nopember Institut Teknologi Sepuluh Nopember
wijayanti@lecturer.undip.ac.id Surabaya, Indonesia Surabaya, Indonesia
dwiarraziqi.19071@mhs.its.ac.id bimogumelar@ieee.org

Mauridhi Hery Purnomo

Dept. of Electrical Engineering
Dept. of Computer Engineering
Faculty of Intelligent Electrical and
Informatics Technology (F-Electics)
Institut Teknologi Sepuluh Nopember
The Science and Technology Center of
Artificial Intelligent for Healthcare and
Society (PUI AIHeS)
Surabaya, Indonesia
hery@ee.its.ac.id

Abstract—FDS (Fall Detection System) is a technology that from fall incidents [2]. The elderly generally live alone
is very essential for the elderly, in order to immediately get whether temporarily or daily. Thus, it is necessary to
help when the fall incident happens. This paper aims to build a establish a real-time fall incident detection system. The
FDS dedicated to the elderly. We propose 2D LiDAR as the results of this detection system will later be sent to family
main sensor in FDS. In this case, 2D LiDAR has the duty to members via smartphone. This paper focuses on the process
obtain information data in a room. FDS is demanded to be able of detecting fall incidents activities.
to distinguish between fall incidents and ADL (Activities of
Daily Living). This paper presents trials on various positions of Researches on FDS (Fall Detection System) have been
the human body that can be detected as fall incidents or ADL. done in various ways: visionary / camera-based FDS,
The trials aim to produce a dataset that will later be processed wearable-based FDS and ambient-based FDS [1]. In FDS
using K-NN and RF as a fall detection algorithm. From the technology, Kinect camera sensors have been used to
results of the trials, 2D LiDAR sensor data can describe two distinguish between fall incident and other activities that are
information as detection points. RF produces accuracy up to not included in fall incident [3]. The use of a camera allows
94% and K-NN produces maximum accuracy of 100%. fall incidents to be detected based on analysis of body
movements and body position [4]. The use of Pi Camera in
Keywords—Fall Detection System, Elderly, 2D LiDAR, K- FDS has advantages, namely it has a good performance and
Nearest Neighbor, Random Forest it is also low cost [5]. In further development, RGB-D (Red,
Green, Blue - Depth) image processing has been developed
I. INTRODUCTION to address privacy issues [6], [7]. Even though, generally the
Beside illnesses, falls are a cause of death from injury in elderly will feel uncomfortable with the camera recording
the elderly. According to WHO, 33% of elderly people have process, especially in a private room such as bathroom [8].
experienced at least one fall and the rest have fallen more The main problem in camera-based FDS is that there is no
than once during their lifetime. Some areas that falls research that proves that camera can work in a very dark and
commonly occur are hallways, bathrooms, kitchens and no light environment.
bedrooms [1]. Elderly people have a risk of falling due to In a wearable-based FDS, detection of fall incidents is
many factors including decreased physical and psychological done by installing a number of sensors located around the
conditions, loss of consciousness, vision problems, joint body. The 3D accelerometer sensor placed on stomach was
problems, respiratory problems, heart failure, and other once used as a fall detection [9], [10]. An accelerometer is
health problems. The older we get, the more the risk of death

 !"!
Authorized licensed use limited London School of Economics & Political Science. 1
"! to:#!$%%&#"'( Downloaded on May 16,2021 at 23:34:48 UTC from IEEE Xplore. Restrictions apply.
 also placed on the user's hand or thigh to detect fall incidents various considerations from the previous papers, then in this
[11], [12]. The main problem of wearable sensors is that it paper, RF and K-NN are used as an algorithm for detection
can be a trouble for users. Elderly people tend to become of fall incidents in LiDAR-based FDS.
forgetful in doing things. Therefore, there is high possibility
that the sensor is not installed. III. 2D LIDAR-BASED FDS
In an ambient-based FDS, various sensors are deployed A. Our Work
ranging from ultrasonic, motion sensors, sound sensors, Our work in this paper is to build an ambient-based FDS
radars and others to detect fall incidents in a room [13], [14]. using 2D LiDAR to detect fall incidents in the elderly. The
FDS has been proposed by using the radio frequency wave FDS that is built must be able to detect fall incidents by
propagation model [15]. However, this method requires recognizing differences in sensor data, both during daily
complicated preparation. FDS was once proposed by activities (ADL) and when the incident occurs. To perform
installing sensor arrays on a carpet to monitor the activities the proposed FDS, the system design is arranged as shown in
of elderly [16]. The problem with ambient sensor-based FDS Figure 1.
in general is a complicated setup process [1].
1) 2D LiDAR functions as the main sensor that has
To overcome this problem, we propose 2D LiDAR (Two the task to collect information in a room.
Dimension Light Detection and Ranging) to be used as FDS. 2) The results of the trials in the ADL and fall
The reason for proposing LiDAR in FDS is, when compared activities produce two trial data then it combined
to camera based FDS, LiDAR is not affected by changes or into a single dataset.
lack of light intensity. Compared to wearable based FDS, the
3) A single dataset will later be split into data training
application of LiDAR does not limit the comfort of the
elderly in doing their daily activities. The set up on LiDAR is and data testing. The percentage split between
very simple and just to be placed in one corner of the room to training and testing will be searched until obtain
scan objects. Thus, there is no need for a combination of optimal accuracy
sensor types, and there are no sensor placement problems as 4) The dataset is processed using K-NN and RF for
in the existing ambient based technology. Although LiDAR detection of fall incidents. The output of the two
is more popular to be applied in terms of navigation such as algorithms is ready to use fall incident information.
UAV [17], autonomous mobile robot [18], autonomous 5) The detection accuracy of the two algorithms are
vehicle [19], we assume that 2D LiDAR can be used to compared.
detect fall incidents. Fall incidents detection can be done by
analyzing the difference in data taken when the room is
empty and when there are people doing activities in the
room.
As general, this paper is still in preliminary stages. This
paper does not discuss the comparison with previous
research [1], [3 – 16]. This paper only proposes and tries to
develop other ways to build ambient-based FDS. We
consider the classical K-NN and RF algorithms to be very Fig. 1. System Design
suitable for preliminary stage of our work. Furthermore, this
paper is organized as follows. Section II describes previous B. Sensor and Environment
researches related to machine learning methods used in FDS,
specifically in detection. We discuss the concept LiDAR- In the proposed FDS, we use the RP LiDAR A1 as the
based FDS in section III. This section also discusses sensors, main sensor. This sensor works with the ability to scan two
environments for trials, datasets and features. In section IV, dimensional objects on 360 degrees. In scanning, the sensor
we tested the two methods we used, compared them and emits infrared laser light with a wavelength of 785 nm with a
analyzed the results. Finally, Section V contains conclusion maximum scanning distance of 12 meters.
from all of the works that we do. In this study, the sensor is not installed in the center of
the room. Instead, it is installed on the side of one wall of the
II. RELATED WORK room. Thus, the sensor scanning angle is assumed not 360
Our work requires learning algorithms to perform degrees but rather 180 degrees. Even though the sensor's
detection tasks. K-NN is one of ten detection algorithms that ability can detect 360 degrees. In addition, the sensor is
are very simple and easy to used [20]. In this study, we need placed 10 cm above the floor (calculated from the surface of
a comparison algorithm which also has good detection the floor to the sensor’s eye). Fig. 2 shows sensor
performance. One of them is a decision tree algorithm that is configuration. Fig. 3 shows the room plan used for the trial.
not too complex [21]. In this case, we chose Random Forest
(RF) as a comparison algorithm which derived from the
development of a decision tree algorithm. For this reason, we
use these two algorithms. K-Nearest Neighbor (K-NN) and
Random Forest (RF) were proposed for ambient-based FDS
[1] and wearable-based FDS [11]. In that research, RF and
K-NN are successfully implemented and have high accuracy
in the dataset produced by the gyroscope and accelerometer.
RF and K-NN are also well implemented in FDS-based
cameras using analysis of key points variation of human
skeleton [22]. In another studies, RF and K-NN have also Fig. 2. Sensor Placement
been implemented in FDS especially fall from bed [23]. With

Authorized licensed use limited to: London School of Economics & Political Science. 2
Downloaded on May 16,2021 at 23:34:48 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. SENSOR TRIALS ON ADL

Activities
An Empty Room
Fig. 3. Room Spesification

Data Sensor
C. Dataset and Feature
When the LiDAR sensor does the scanning, the
information is obtained in every 1º distance of the maximum
scanning of 360°. If it is assumed that the sensor is placed on
one wall of the room, then the sensor only produces
information at every 1º of 180º distances. Thus, the features
processed in the detection algorithm are 180 features.
Then, ten male volunteers with weight, height, average
age (75 kg, 170 cm, 30 years old). In this case, we do not

Activities
involve the elderly when conducting trials. Each volunteer
demonstrates six activities consisting of three ADL (An
Empty Room, Standing Up / Walking, Sitting Down) and
Standing Up / Walking

three falling activities (Falling Transversal, Falling

Longitudinal, Falling Oblique). These six activities are
demonstrated when the room is sufficiently lit. From the
results of the trials, 60 lines of dataset are obtained. Then the
trials are conducted again in the dark room. From the results
of the trials, we get 60 more lines of dataset. Thus, the lines
of dataset become 120. Because the feature used in this study
Data Sensor

is 180° distance information, the dataset in this study is a

matrix with a size of 120x180. Because the features obtained
are the results of scanning sensors, the dataset in this study is
referred to as primary dataset. In this study, the dataset are
used for detection tasks. The purpose of the detection is to
clearly describe the presence or absence of a fall incident.
There are only two information, ADL or Fall.
IV. RESULT AND DISCUSSION
A. Sensor Trials on ADL
Activities

These trials were conducted to find out various activities

except falling that can be recognized by the sensors. Various
activities performed by a volunteer. There are three activities
that have a significant difference in sensor data namely, an
empty room, standing up or walking, and sitting down as
Sitting Down

shown in Table I. The trial is conducted in a dark room even

darker than the picture in Table I to prove that the LiDAR
sensor can work well in a dark room. This is shown from the
sensor data when detecting objects and home furniture in a
dark room. Sensors can detect objects such as when the
Data Sensor

object is standing up, sitting, or not in the room.

B. Sensor Trials on Fall Activities
Same as previous trial, these trials were conducted to find
out various falling activities that can be recognized by the
sensors. There are three activities that have a significant
difference in sensor data namely, falling transversal,
longitudinal and oblique towards the sensor as shown in
Table II. This trial is also conducted in a dark room, the same
as in the previous trial.

Authorized licensed use limited to: London School of Economics & Political Science. 3
Downloaded on May 16,2021 at 23:34:48 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. SENSOR TRIALS ON FALL ACTIVITIES C. K-NN Accuracy
The K parameter in the K-NN algorithm is the parameter
that its optimal value must be found. In addition, the
percentage of data training and testing is also a parameter
that its optimal value must be found. The optimal value of
Activities
Falling Transversal Towards the Sensor

the two parameters will produce maximum accuracy. For this

reason, the purpose of this trial is to find the optimal K value
and percentage split (training%, testing%). By using the
standard K-NN library on the python platform, the results are
as shown in Fig. 4. The trial results show that the maximum
accuracy of K-NN in this trials is 1.0 from the range (0.0 to
1.0) or 100% in range percent. Maximum accuracy is
achieved when K = 2 and percentage split (90%, 10%).
Data Sensor
Activities
Falling Longitudinal Towards the Sensor

Fig. 4. K-NN Accuracy in Detection Mode

Data Sensor
Activities
Falling Oblique Towards the Sensor

Fig. 5. RF Accuracy in Detection Mode

D. RF Accuracy in Detection Mode

The purpose of this trial is to get an optimal accuracy by
finding the optimal value of the percentage of data training
and testing. In contrast to KNN, the RF algorithm does not
recognize the parameter K. Thus, the parameters sought for
their optimal value are only percentage split (training%,
Data Sensor

testing%). However, in these trials, accuracy values vary at

various percentage split values (training%, testing%). For
this reason, the accuracy value is taken from the average
accuracy of 20x trials. Fig. 6 shows the results of 20x trials
on various percentage split values (training%, testing%).
The trial results show that the maximum RF accuracy in this
trial is 0.94 from the range (0.0 to 1.0) or 94% in range
percent. This maximum accuracy is the average value of 20x
trials when percentage split (80%, 20%).

Authorized licensed use limited to: London School of Economics & Political Science. 4
Downloaded on May 16,2021 at 23:34:48 UTC from IEEE Xplore. Restrictions apply.
 E. Discussion and Methods Comparison International Seminar on Sensors, Instrumentation, Measurement and
Metrology (ISSIMM), Surabaya, 2017, pp. 26-29.
From the results of trials III A and III B, using a single [8] M. Koutli, N. Theologou, A. Tryferidis and D. Tzovaras, "Abnormal
2D LiDAR sensor, there are six activities that have a Behavior Detection for Elderly People Living Alone Leveraging IoT
significant difference in data, namely, an empty room, Sensors," 2019 IEEE 19th International Conference on Bioinformatics
standing up/walking, sitting down, falling transversal, falling and Bioengineering (BIBE), Athens, Greece, 2019, pp. 922-926,
longitudinal and falling oblique toward the sensor. Then, [9] A. Kurniawan, A. R. Hermawan and I. K. E. Purnama, "A wearable
from six activities categorized into two information, Fall and device for fall detection elderly people using tri dimensional
ADL. The proposed FDS must also be able to detect two accelerometer," 2016 International Seminar on Intelligent Technology
and Its Applications (ISITIA), Lombok, 2016, pp. 671-674.
information namely Fall and ADL. RF algorithm produces
an detection accuracy up to 94%. Meanwhile, the K-NN [10] T. de Quadros, A. E. Lazzaretti and F. K. Schneider, "A Movement
Decomposition and Machine Learning-Based Fall Detection System
algorithm shows better results up to 100%. Thus, the number Using Wrist Wearable Device," in IEEE Sensors Journal, vol. 18, no.
of datasets used in this study are enough to perform the 12, pp. 5082-5089, 15 June15, 2018.
detection task. However, this paper is still in the preliminary [11] W. Saadeh, S. A. Butt and M. A. B. Altaf, "A Patient-Specific Single
stages of work. Need to improvement of feature number and Sensor IoT-Based Wearable Fall Prediction and Detection System," in
size of the dataset. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, vol. 27, no. 5, pp. 995-1003, May 2019.
V. CONCLUSION [12] N. Zurbuchen, P. Bruegger and A. Wilde, "A Comparison of Machine
Learning Algorithms for Fall Detection using Wearable Sensors,"
This paper proposes to build an FDS dedicated to the 2020 International Conference on Artificial Intelligence in
elderly by using the 2D LiDAR sensor. This proposal is Information and Communication (ICAIIC), Fukuoka, Japan, 2020,
alternative option in building an FDS, with the advantage of [13] E. Cippitelli, F. Fioranelli, E. Gambi and S. Spinsante, "Radar and
having a LiDAR sensor implementation that is not affected RGB-Depth Sensors for Fall Detection: A Review," in IEEE Sensors
by light intensity, so that it can still detect well in dark room. Journal, vol. 17, no. 12, pp. 3585-3604, 15 June15, 2017.
Overall both methods show good performance. K-NN shows [14] F. De Backere et al., “Towards a social and context-aware multi-
better performance than RF This is indicated by the sensor fall detection and risk assessment platform,” Comput. Biol.
maximum accuracy value that reach. Med., vol. 64, pp. 307–320, Sep. 2015.
[15] Y. Wang, K. Wu and L. M. Ni, "WiFall: Device-Free Fall Detection
ACKNOWLEDGMENT by Wireless Networks," in IEEE Transactions on Mobile Computing,
vol. 16, no. 2, pp. 581-594, 1 Feb. 2017.
We would like to offer our special thanks to LPDP
[16] [15] F. Muheidat, L. Tawalbeh and H. Tyrer, "Context-Aware,
scholarship program of the Ministry of Finance, Republic of Accurate, and Real Time Fall Detection System for Elderly People,"
Indonesia with the BUDI-DN scheme. 2018 IEEE 12th International Conference on Semantic Computing
(ICSC), Laguna Hills, CA, 2018, pp. 329-333.
REFERENCES [17] L. Zheng, P. Zhang, J. Tan and F. Li, "The Obstacle Detection
Method of UAV Based on 2D LiDAR," in IEEE Access, vol. 7, pp.
[1] F. Hussain, F. Hussain, M. Ehatisham-ul-Haq and M. A. Azam, 163437-163448, 2019.
"Activity-Aware Fall Detection and Recognition Based on Wearable [18] D. Ghorpade, A. D. Thakare and S. Doiphode, "Obstacle Detection
Sensors," in IEEE Sensors Journal, vol. 19, no. 12, pp. 4528-4536, 15 and Avoidance Algorithm for Autonomous Mobile Robot using 2D
June15, 2019. LiDAR," 2017 International Conference on Computing,
[2] M. Cheffena, "Fall Detection Using Smartphone Audio Features," in Communication, Control and Automation (ICCUBEA), Pune, 2017,
IEEE Journal of Biomedical and Health Informatics, vol. 20, no. 4, pp. 1-6.
pp. 1073-1080, July 2016. [19] A. N. Catapang and M. Ramos, "Obstacle detection using a 2D
[3] S. Kasturi and K. Jo, "Human fall classification system for ceiling- LIDAR system for an Autonomous Vehicle," 2016 6th IEEE
mounted kinect depth images," 2017 17th International Conference on International Conference on Control System, Computing and
Control, Automation and Systems (ICCAS), Jeju, 2017, pp. 1346- Engineering (ICCSCE), Batu Ferringhi, 2016, pp. 441-445.
1349. [20] Okfalisa, I. Gazalba, Mustakim and N. G. I. Reza, "Comparative
[4] A. Lotfi, S. Albawendi, H. Powell, K. Appiah and C. Langensiepen, analysis of k-nearest neighbor and modified k-nearest neighbor
"Supporting Independent Living for Older Adults; Employing a algorithm for data classification," 2017 2nd International conferences
Visual Based Fall Detection Through Analysing the Motion and on Information Technology, Information Systems and Electrical
Shape of the Human Body," in IEEE Access, vol. 6, pp. 70272-70282, Engineering (ICITISEE), Yogyakarta, 2017, pp. 294-298.
2018. [21] R. Agrawal. “K-Nearest Neighborn for Uncertain Data”. International
[5] S. A. Waheed and P. Sheik Abdul Khader, "A Novel Approach for Journal of Computer Applications (0975-8887). 2014. Vol. 105 No.
Smart and Cost Effective IoT Based Elderly Fall Detection System 11. pp 13-16.
Using Pi Camera," 2017 IEEE International Conference on [22] A. Y. Alaoui, S. El Fkihi and R. O. H. Thami, "Fall Detection for
Computational Intelligence and Computing Research (ICCIC), Elderly People Using the Variation of Key Points of Human
Coimbatore, 2017, pp. 1-4. Skeleton," in IEEE Access, vol. 7, pp. 154786-154795, 2019.
[6] C. Sun, M. Sheu and Y. Syu, "A new fall detection algorithm based [23] F. Zhao, Z. Cao, Y. Xiao, J. Mao and J. Yuan, "Real-Time Detection
on depth information using RGB-D camera," 2017 International of Fall From Bed Using a Single Depth Camera," in IEEE
Symposium on Intelligent Signal Processing and Communication Transactions on Automation Science and Engineering, vol. 16, no. 3,
Systems (ISPACS), Xiamen, 2017, pp. 413-416. pp. 1018-1032, July 2019.G. Eason, B. Noble, and I. N. Sneddon, “On
[7] Muhtadin, A. Kurniawan, A. A. N. S. Laksamana and I. K. E. certain integrals of Lipschitz-Hankel type involving products of
Purnama, "Fall detector implementation in a robot service," 2017 Bessel functions,” Phil. Trans. Roy. Soc. London, vol. A247, pp. 529–
551, April 1955.

Authorized licensed use limited to: London School of Economics & Political Science. 5
Downloaded on May 16,2021 at 23:34:48 UTC from IEEE Xplore. Restrictions apply.