Final Project Presentation

DESIGN AND IMPLEMENTATION OF

FALL DETECTION FOR ELDERLY USING
CONVOLUTIONAL NEURAL NETWORKS
Gusta Wacika Gede Juniarta
NRP. 1110171022

Supervisors:
Hary Oktavianto
Akhmad Hendriawan
Elly Purwantini

Bachelor Degree Program of Electrical Engineering Department

Politeknik Elektronika Negeri Surabaya
June 19, 2024
Outlines

• Chapter 1. Introduction
• Chapter 2. Literature review
• Chapter 3. System design
• Chapter 4. Experimental result
• Chapter 5. Conclusion

2
Chapter 1. Introduction
Motivation
Falling is the most common of events that threaten the elderly. In this case, falling detection and
ongoing monitoring technology is needed to warn others to help victims get timely treatment.

Problem definition
 How to measure the movement of the human body?
 How to detect the falling movement of human body

Objectives
 Making a circuit using accelerometer on the belly to measure human
body movement.
 Designing a classification algorithm using CNN to detect the falling
movement
Contributions
 Monitoring elderly who lives alone.
3
Chapter 2. Literature review
Song-Mi Lee, Sang Min Yoon and Heeryon Cho, "Human activity recognition from accelerometer
data using Convolutional Neural Network," 2017 IEEE International Conference on Big Data and
Smart Computing (BigComp), Jeju, Korea (South), 2017, pp. 131-134, doi:
10.1109/BIGCOMP.2017.7881728.
 This paper explains about Convolutional Neural Network (CNN) which classify human body movements by
collecting data from the accelerometer and then changing it to vector quantities.
 This method is used to classify between daily activities and falling events

From the picture it is explained that

the workings of this method are by
collecting raw data from the
accelerometer then transform it
into a vector quantity and finally
classifying it from the data taken.

4
Chapter 2. Literature review
P. Pierleoni, A. Belli, L. Palma, M. Pellegrini, L. Pernini and S. Valenti, "A High Reliability
Wearable Device for Elderly Fall Detection," in IEEE Sensors Journal, vol. 15, no. 8, pp. 4544-
4553, Aug. 2015, doi: 10.1109/JSEN.2015.2423562
 This paper explains about the fall detection algorithm through a simulation test fall, fall
simulation with recovery and ADL.
 The fall simulation is such as :
 Start: Attracted by the force of gravity, the body accelerates reaching maximum speed
just before impact to the ground. During the free fall, the root mean square (RMS) of
acceleration will tend to 0 g or otherwise will be substantially less than 1 g.
 Impact: After experiencing weightlessness, the subject impacts the ground or other
objects. In this phase the RMS rapidly increases until reaching a peak greater than 2 g.
 Aftermath: In this phase the RMS presents a flat trend. After this period of immobility,
the conscious subject starts to move trying to get up.
 Posture : After a fall, the subject’s body will be in a different orientation
 than before the impact. Therefore the acceleration values of the individual axes will be From the picture there are differences
different from those measured before the fall in the value of acceleration and gravity
 The method used in this paper is the Support Vector Machine (SVM) of each event that occurs
 method to classify between daily activities and falling events
5
Chapter 2. Literature review
MPU-6050 Sensor
The definition of MPU-6050 sensor
The MPU6050 module is a Micro Electro-Mechanical
Systems (MEMS) which consists of a 3-axis
Accelerometer and 3-axis Gyroscope inside it. This helps
us to measure acceleration, velocity, orientation,
displacement and many other motion related parameter
of a system or object.

How it works?
All accelerometers work on the principle of a mass on a
spring, when the thing they are attached to accelerates,
then the mass wants to remain stationary due to its
inertia and therefore the spring is stretched or
compressed, creating a force which is detected and
corresponds to the applied acceleration.
6
Chapter 3. System design
Block diagram system

7
 Chapter 3. System design
Software design

• Collect the data from accelerometer and

send the data accelerometer to python
using UDP communication
• Convert the acceleration data into the
resultant acceleration vector, then save
the data in the form of a csv.file
• Read the csv.file and convert it into
spectrogram image
• Collect spectrogram image for training
dataset.
• Training the dataset
• Get the training model to distinguish 3
class of walking, idle, falling
• Test the CNN prediction of 3 class
8
 Chapter 3. System design
CNN structure
•Input Layer: there is 3 input classes walking, idle, falling with
503 x 250 pixel size

•6 Hidden Layers: Consist of 2 Conv2D layers, 2 MaxPooling2D

layers, 1 Flatten layer, and 1 Dense layer.
Hidden
•Output Layer: Dense Layer 2 (Output Layer) and Number of layer
Neurons: 3 (each corresponding to one of the classes: walking,
falling, idle)

•Batch size : 32 epoch : 10

𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑚𝑎𝑔𝑒𝑠
•𝑆𝑡𝑒𝑝 𝑝𝑒𝑟 𝑒𝑝𝑜𝑐ℎ =
𝐵𝑎𝑡𝑐ℎ 𝑠𝑖𝑧𝑒

300
•Step per epoch= = 9.375 = 10
32 Output
•Total Steps: Since I have 10 steps per epoch and I use 10 epochs, layer
the total number of steps is 10 steps/epoch×10 epochs= 100 steps 9
Chapter 3. System design
Hardware design

• NodeMCU 8266 is for connect wifi and send the data

accelerometer to laptop that installed with python
software
• MPU-6050 to get the 3 axis accelerometer data in X-
axis,Y-axis,Z-axis.
• Ppowerbank with 5V dc output with 5000mAH
capacity for the power supply
• USB cable Micro to USB to connect Powerbank with
ESP8266

10
Chapter 3. System design
Mechanical design

Design with Sketchup Actual design

Sensor box dimensions 10.5 cm x 8 cm x 5 cm Sensor box dimensions 10.6 cm x 8.1 cm x 5 cm

11
Chapter 4. Experimental result
Converting dataset into spectrogram

Idle dataset before spectrogram Idle dataset after spectrogram

12
Chapter 4. Experimental result
Converting dataset into spectrogram

walking dataset before spectrogram walking dataset after spectrogram

13
Chapter 4. Experimental result
Converting dataset into spectrogram

falling dataset before spectrogram falling dataset after spectrogram

14
Chapter 4. Experimental result
Classification CNN
The ouput of Classification and
identification class

15
Chapter 4. Experimental result
Classification CNN
The ouput of Classification and
identification class

16
Chapter 4. Experimental result
Classification CNN
The ouput of Classification and
identification class

17
Chapter 4. Experimental result
Sending alert via Whatsapp

18
Chapter 4. Experimental result
Power consumption

• ESP8266 The current measurement test was carried out using an avometer and the results obtained
were a working current of 127.8 mA
• To calculate the power supply resistance, it can be calculated using the following formula.

𝐵𝑎𝑡𝑡𝑒𝑟𝑦 𝑐𝑎𝑝𝑎𝑐𝑖𝑡𝑦 (𝑚𝐴𝐻)

• 𝑊𝑜𝑟𝑘 𝑂𝑝𝑒𝑟𝑎𝑡𝑖𝑜𝑛 (ℎ𝑜𝑢𝑟) =
𝑃𝑜𝑤𝑒𝑟 𝑐𝑜𝑛𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛(𝑚𝐴)

5000 (𝑚𝐴𝐻)
𝑊𝑜𝑟𝑘 𝑂𝑝𝑒𝑟𝑎𝑡𝑖𝑜𝑛 (ℎ𝑜𝑢𝑟) =
127.8(𝑚𝐴)

𝑊𝑜𝑟𝑘 𝑂𝑝𝑒𝑟𝑎𝑡𝑖𝑜𝑛 ℎ𝑜𝑢𝑟 ≈ 39.10 hours

• From the calculation results, it was found that the battery life was up to 39 hours or 1 day 15 hours,
but the results of this test were that the battery could only last 29 hours or 1 day 5 hours.

19
 Probability Measurement Target Actual
0.98 idle idle idle
In this test, 10 0.97 idle idle idle
idle trials were 0.98 idle idle idle
carried out and 0.99 idle idle idle
the error 1.00 idle idle idle
percentage result 1.00 idle idle idle
was 0%. 0.98 idle idle idle
0.98 idle idle idle
0.97 idle idle idle
0.99 idle idle idle

System Performance
20
 Probability Measurement Target Actual
0.98 walking walking walking
In this test, 10 0.97 walking walking walking
walking trials 0.95 falling walking walking
were carried out 0.98 walking walking walking
and the error 0.97 walking walking walking
percentage result 1.00 walking walking walking
was 11.11%. 0.97 walking walking walking
1.00 walking walking walking
0.96 walking walking walking
0.97 walking walking walking

System Performance
21
 Probability Measurement Target Actual
0.98 falling falling falling
In this test, 10 0.98 falling falling falling
walking trials 0.97 falling falling falling
were carried out 0.97 falling falling falling
and the error 0.97 falling falling falling
percentage result 0.98 falling falling falling
was 11.11%. 0.97 falling falling falling
0.95 walking falling falling
0.98 falling falling falling
0.98 falling falling falling

System Performance
22
Chapter 4. Experimental result
Analysis and discussion
• CNN identification testing showed that walking, standing still and falling
experiments had different spectrogram shapes.

• The still spectrogram results have a blue color, which means they have a small
magnitude value of -70dB, and the falling spectrogram results have the largest
magnitude value of -20dB.

• The idle test had the smallest error percentage compared to the walking and
falling experiment, this could be due to the lack of CNN training data and the
noise during data collection.

• From the power consumption test the battery could only last 29 hours or 1 day 5
hours. This can be caused by wasted heat energy and a mismatch in the capacity
of the power bank or battery.
23
Chapter 5. Conclusion
• The box sensor dimension is 10.6 cm x 8.1 cm x 5 cm
• The accelerometer values have a minimum of 1 G when idle and a maximum
of 2 G up when falling
• In a spectrogram, there are typically three main parameters such a Time in x-
axis, Frequency represented on the y-axis indicated content of the signal, and
magnitude/ amplitude indicates the strength or magnitude of each frequency
component in the signal
• The system using a power bank with a capacity of 5000mAH and the system
can run for 29 hours.
• We can use WhatsApp to send “fall detection” alerts

24
Thank you
Gusta Wacika Gede Juniarta
+62 812-3528-4339