FALL DETECTION FOR
ELDERLY PEOPLE
Final Year Project TEAM MEMBERS
• AKHIL TOM ZACHARIAH
Department of Electronics and • ALLEN ZACHARIA MATHEW
Communication
• ALOSHI JOSPEH
Project Guide: Ms Shyama Sreekumar • DENNIS BIJU
CONTENTS
• INTRODUCTION
• OBJECTIVE
• LITERATURE SURVEY
• BLOCK DIAGRAM
• METHODOLOGY
• WORK DONE(50%)
• OUTPUT
• REFERENCE
INTRODUCTION
• In today’s world elderly people are left alone at home
and no one around them to help.
• Taking care of the elderly people is a very important
responsibility. In such a scenario, technology helps
people by providing life support.
• Falling down is a major cause for injuries and accidents
in elderly people and it can lead to serious injuries and
even death when it is unnoticed.
• To overcome this situation we are introducing a
wearable that can be worn in the hand which detect
falling in elderly people and alert their caregivers or
their family.
OBJECTIVE
• To design and implement a fall detection
system using sensors and to collect a
diverse dataset of fall and non-harmful
activities of humans.
• To test the employability of CNN technique
for real time fall monitoring and
classification to achieve maximum accuracy.
LITERATURE REVIEW
PAPER1:Accelerometer-Based Human Fall
Detection Using
Convolutional Neural Networks
• Develop a CNN-based model for detecting human falls using
accelerometer data in an IoT and fog computing environment.
• A Convolutional Neural Network (CNN-3B3Conv) with three
convolutional layers, maxpooling, and fully connected layers
was used to process accelerometer data for fall detection.
• Small dataset sizes led to the necessity of using data
augmentation techniques.
• Using data augmentation significantly improved the model's
performance, particularly in terms of accuracy and precision.
• Simpler models like CNN-1Conv performed better in some
cases due to less overfitting compared to more complex
models.
PAPER 2:A Reliable Fall Detection
System Based Analysing The Physical
Activities Of Older Adults Living In
Long-Term Care Facilities
• Develop a machine learning-based fall detection system using
real-world data from older adults in long-term care facilities.
• Data from wrist-worn and neck-worn devices were recorded
over 400 days using accelerometers and barometers.
• Existing datasets often use simulated ADLs and falls, which do
not reflect real-world conditions, leading to reduced
sensitivity.
• High computational complexity for real-time processing on
wearable devices.
• Training on real-world ADLs significantly improved the
system's sensitivity, achieving 100% on both real and
simulated falls.
PAPER 3:SmartFall: A Smartwatch-
Based Fall Detection
System Using Deep Learning
• Develop a fall detection system using smartwatches with
accelerometer data and deep learning techniques, comparing
traditional models (SVM, Naive Bayes) and deep learning models
(RNN with GRU layers).
• A recurrent neural network (RNN) with gated recurrent units
(GRUs) was implemented to capture temporal patterns in
accelerometer data.
• The models showed a tendency for false positives, particularly in
non-fall activities with abrupt movements like jogging or waving.
• The deep learning model generalized better to new users,
indicating its potential for real-world application despite false
positives in activities with sudden motions.
PAPER 4:Machine Learning
Techniques Applied to the
Development of a Fall Risk Index
• for Older
Extracting Adults
and analyzing a comprehensive dataset of older adults'
profiles, clinical information, and fall history using natural
language processing and machine learning techniques to develop
a predictive fall risk index.
• Reliance on natural language processing for data extraction may
introduce biases or inaccuracies due to inconsistent formatting or
overlooked critical information.
• The retrospective nature of data collection limits the
generalizability of findings to broader populations or different
clinical settings.
• Potential variability in practitioner documentation practices could
affect the completeness and reliability of the dataset.
BLOCK DIAGRAM
METHODOLOGY
The MPU6050 sensor which is accelerometer gyroscope
6 axis sensor take sensor values continuously.
The Raspberry Pi 4 microprocessor is loaded with the
CNN algorithm.
The data set that is taken from the “Fall vs Normal
Activities” from Kaggle is being pre processed, trained
with CNN algorithm tested and validated.
This is loaded to the microprocessor.
Now when the person falls the caretaker will receive a
message that the person has been fallen via warning
signal.
WORK DONE(50%)
Data preprocessing done to the dataset.
Dataset was trained with different CNN algorithm.
The best CNN algorithm was found with 97%
accuracy.
3D designing was done with AutoCAD 3D for the
belt.
The 3D model has been given for 3D printing.
DATA PREPROCESSING
SPLIT DATA:
• Data is split into 80% training and 20% testing using stratified sampling
to maintain class distribution.
• Splitting is done before SMOTE to prevent data leakage and ensure a
realistic evaluation.
RESHAPE DATA:
• The data is reshaped into a 3D format (samples, 400, 6), where 400
represents time steps and 6 represents sensor features.
• This format is necessary for deep learning models like CNNs to process
sequential sensor data effectively.
SCALING DATA:
• StandardScaler is applied to normalize the data by removing the mean
and scaling to unit variance.
• The scaler is fitted on the training data and then applied to the test data
to prevent data leakage.
CNN ALGORITHMS USED TO
COMPARE THE ACCURACY
The given CNN model is a sequential deep learning model used for
human activity and fall classification. It processes triaxial
accelerometer and gyroscope data through convolutional layers,
pooling, and fully connected layers.
Total Layers 6 (including input and output).
Feature Extraction Done using Conv1D and
MaxPooling1D.
ReLU(Rectified Linear Unit) In hidden layers to introduce non-
linearity.
Sigmoid In the output layer for class
probabilities.
Accuracy 94%
CNN-HE(Heuristic Enhancements)
The CNN-HE neural network is designed for the classification of human
daily activities and fall behaviors. It utilizes convolutional layers and
heuristic enhancements to improve accuracy in fall detection. The
network structure consists of the following layers:
Total Layers 8 (including input and output).
Feature Extraction Done using Conv1D and
MaxPooling1D.
ReLU(Rectified Linear Unit) Hidden layers(avoids vanishing
gradients).
Softmax Output layer (for multi-class
probability distribution).
Accuracy 95%
ADVANTAGES OVER FIRST
CNN
· More filters (better feature extraction).
· Larger dense layer (512 neurons) for complex
pattern recognition.
· More training epochs (70) for better convergence
CNN-3B3Conv Neural Network
Architecture
It uses three blocks of convolutional layers to extract features,
followed by fully connected layers for classification.
Total Layers 12 (including input and output).
Feature Extraction Conv1D and MaxPooling1D.
ReLU(Rectified Linear Unit) Hidden layers(Used in
convolutional and fully connected
layers for non-linearity).
Softmax Output layer (for multi-class
probability distribution).
Accuracy 96%
ADVANATGES OVER CNN-HE
· Uses three blocks of convolutional layers for deeper feature
extraction.
· The three blocks of convolutional layers enable hierarchical
feature learning:
· Dropout Layer for Regularization Dropout improves the
model's generalization ability, making it perform better on
unseen data.
CNN-EDU Neural Network
Architecture
It uses multiple convolutional layers and dropout regularization to improve
performance and reduce overfitting.
Total Layers 14 (including input and output).
Feature Extraction Two Conv1D and MaxPooling1D.
ReLU(Rectified Linear Unit) Hidden layers(Used in
convolutional and fully connected
layers for non-linearity.)
Softmax Output layer(Used in output layer).
Accuracy 97%
ADVANTAGES OVER CNN-3B3
• It has a deeper and progressive architecture for better
feature extraction.
• It uses dropout to prevent overfitting.
• It is trained for more epochs with a smaller batch size,
improving performance.
• It incorporates educational enhancements for robustness
and accuracy
CBAM (Convolutional-Bidirectional
LSTM with Attention Mechanism)
It combines convolutional layers, bidirectional LSTM layers, and dropout
regularization to enhance feature extraction and temporal modeling.
Total Layers 6 (including input and output).
Feature Extraction Two Conv1D and MaxPooling1D.
ReLU(Rectified Linear Unit) Hidden layers(Used in
convolutional and fully connected
layers for non-linearity.)
CBAM (Combined) Gets the best of both worlds by
detecting local patterns and
understanding how they change
over time.
Accuracy 97%
ADVANTAGES OVER CNN-EDU
· Better at handling sequential data: Processes data in both
directions, understanding time-dependent patterns.
· Attention mechanism: Focuses on critical moments in the
data.
· Improved feature extraction: Extracts both local and global
features.
CBAM-EDU (Convolutional-
Bidirectional LSTM with
It combinesEducational Enhancements)
convolutional layers, bidirectional LSTM layers, and educational
enhancements (like the GELU activation function) to improve feature
extraction and temporal modeling.
Total Layers 14 (including input and output).
Feature Extraction Two Conv1D and MaxPooling1D.
GeLU(Gaussian Linear Unit) Provides smoother non-linearity
than ReLU, improving gradient flow
and reducing saturation effects.
CBAM (EDU) Hybrid architecture that
combines convolutional layers for
spatial feature extraction
and bidirectional LSTM layers for
temporal modeling
Accuracy 97%
ADVANTAGES OF CBAM-EDU
· It uses GELU activation, which is smoother and more
efficient than ReLU.
· It includes dropout to prevent overfitting.
· It is trained for 50 epochs with a batch size of 32,
balancing training efficiency and performance.
OUTPUTS
ACCURACY AND LOSS GRAPH
Accuracy improves rapidly in the first few epochs.
Loss decreases sharply in the first few epochs.
The model reaches peak validation performance at ~35
epochs.
After 35 epochs, training might not yield much improvement.
No major overfitting is observed, but training beyond this
point may cause diminishing returns.
3D MODEL OF THE BELT
WORK PLAN
WORK DIVISION
Data Acquisation & Data Allen and Aloshi
Preprocessing
Sensor Integration & Akhil and Dennis
Cloud Creating
Training Model with CNN Allen and Aloshi
algorithm
Designing wearable Akhil and Dennis
device
CONCLUSION
Designed 3D model for wearable device using
Autocad 3D
Given 3D model for printing.
Data preprocessing done to the dataset.
Trained the model with CNN algorithm for fall
detection with “Fall vs Normal activities dataset”.
REFERENCES
Badgujar, Sejal, and Anju S. Pillai. "Fall detection for elderly people using machine
learning." 2020 11th international conference on computing, communication and
networking technologies (ICCCNT). IEEE, 2020.
M. J. A. Nahian et al., "Towards an Accelerometer-Based Elderly Fall
Detection System Using Cross-Disciplinary Time Series Features," in IEEE
Access, vol. 9, pp. 39413-39431, 2021, doi:
10.1109/ACCESS.2021.3056441.
Vaiyapuri, Thavavel, et al. "Internet of things and deep learning enabled elderly fall
detection model for smart homecare." IEEE Access 9 (2021): 113879-113888.