Department of Computer Science and Engineering

Fall Detection In the Elderly

By
Under the Guidance of
P. Nitya Reddy 1MS22CS098 Uzma Sulthana
Joann Joseph 1MS22CS072 Assistant Professor, Dept.
of CSE, RIT
 INTRODUCTION
This project focuses on developing a fall detection system for elderly individuals using sensor
data and machine learning algorithms. This system aims to provide an efficient, real-time
solution for fall detection, using machine learning algorithms to accurately distinguish falls from
normal activities, ensuring quick response and better elderly care.
1. Dataset:
The dataset used for training and testing the fall detection model contains sensor data, which
includes measurements of acceleration in multiple directions (e.g., x, y, z axes).
2. Data Preprocessing:
The raw data is first filtered to remove noise and standardize the features. This includes
normalizing the sensor data and splitting it into training and testing sets.
3. Feature Extraction:
Features such as mean, variance, and standard deviation are extracted from the raw sensor
data. These features capture the important characteristics of the data that help distinguish
between normal activities and falls.
INTRODUCTION
•Model Training & Evaluation:
Uses LSTM-based neural networks trained on diverse datasets, including features
from accelerometers and gyroscopes, with advanced preprocessing techniques (e.g.,
sliding window, normalization).

•Fall Detection:
The trained LSTM model processes incoming sensor data in real-time to detect falls
by identifying temporal patterns and transitions between stable and unstable phases.
About the Fall Detection System Project
Project area:
The fall detection system is an essential area of research, especially for elderly care. Falls are
a leading cause of injury in older adults, often leading to severe consequences. This project
focuses on developing an automatic fall detection system using sensor data and machine
learning algorithms to enhance safety and provide timely assistance.
Importance:
Falls among the elderly are a significant health risk, often leading to long-term injuries or
fatalities. Detecting falls in real-time allows for quicker medical intervention, reducing the
risk of complications and improving the quality of life for elderly individuals.
 About the Fall Detection System Project
Main Goals:
1. To develop an accurate fall detection system using machine learning classifiers.
2. To process and analyze sensor data, extracting meaningful features to distinguish
between falls and normal activities.
3. To provide real-time fall detection with a mechanism for alerting caregivers or
emergency services.
This project aims to contribute to the growing field of elderly care and healthcare
technology, ensuring better safety and prompt response in case of falls.
OBJECTIVES OF LITERATURE SURVEY
•Key Trends: Analyze advancements in sensor tech, data collection, and ML algorithms
for fall detection.

•Methodologies: Review approaches using accelerometers, gyroscopes, and

classification techniques like Decision Trees and Neural Networks.

•Research Gaps: Identify areas to improve accuracy, reduce errors, and enhance real-
time detection.

•Current Solutions: Evaluate the effectiveness of fall detection systems in elderly care.

•Emerging Tech: Explore wearable devices, deep learning, and IoT integrations for
better detection.
 Literature Survey
Title, Author and Year Technique/Methodology used Result Remark
“Support Vector Machine Classifiers Datasets: SisFall, FARSEEING • All classifiers (SVM, kNN, RF) • High generalizability, robustness
Show High Generalizability in achieved high performance metrics on • FARSEEING has limited dataset.
Automatic Fall Detection in Older Algo: SVM(Support Vector the SisFall test set. • 6% of detections were
Adults“ Machine), kNN(K-Nearest • SVM and RF performed better than misclassified as falls (false alarms
Neighbors), RF (Random Forest) kNN in sensitivity. should be reduced)
Jalal Alizadeh, Martin Bogdan, Joseph • SVM maintained high generalizability • Temporal alignment of falls (not
Classen Statistical Analysis: Kruskal– to real-world accidental falls than RF, be feasible in a real-time
Sensors - MDPI Wallis test, Mann–Whitney-U kNN continuous data stream)
2021 test
“An artificial neural network–based Algo: Artificial Neural • Sigmoid activation+random • Wrist-based acceleration sensors
fall detection” Network (ANN) initialization=99.98% accurate more accurate (100%) than waist
Three hidden layers of size: 500- detection of falls from the SMV (Signal or chest sensors
SunGil Yoo, Dongik Oh 500-2000 neurons Magnitude Vector) data. • SMV data proved more efficient
Input layer size: 525 neurons • The system also showed relatively than raw value data
International Journal of Engineering Output layer: Two neurons (fall short training times, particularly for • Zero-fill compensation for
Business Management ([0, 1]) or non-fall ([1, 0])) the case where zero-fill was used for missing data better than scaling
2019 Activation Functions: Sigmoid, data compensation. • Limited fall types and sensor
ReLU placement
“An Elderly Fall Detection System Dataset : Samples (767 falls, Accuracy: 99% Data Diversity: Future work should
Using Enhanced Random Forest in 1017 non-falls (9 activities)) Precision: High focus on testing the system with a
Machine Learning” Algo: Enhanced Random Recall: High larger and more diverse dataset,
Ravikumar Subburam, Vijay Forest Classifier (ERF) F1 Score: High including data from different
Kandasamy ERF - more effective than RF, AdaBoost, demographics.
Engineering Proceedings KNN, Gradient Boosting)
2023

7
 Literature Survey
Title, Author and Year Technique/Methodology used Result Remarks
“Fall Detection System for Elderly Dataset : UR Fall Dataset (URFD), Performance Metrics: • Falls were detected when the
People using Vision-Based Analysis” Fall Detection Dataset (FDD) person’s centroid moved quickly
Sensitivity/Recall: 91.17% relative to the ground plane, with a
Thathupara Subramanyan, Young-Min Texture segmentation with Gabor Specificity: 96.00% significant change in angle
JANG, Sang-Bock filter(sinusoidal wave*Gaussian Accuracy: 90.53% observed.
function) • The system was able to detect falls
Romanian Journal of Information Angle change calculation+Object It works with RGB data (no depth maps from videos taken in various
Science and Technology Tracking with Kalman Filter or accelerometer data), making it both settings (indoor environments) with
2020 (tracking centroid) practical and cost-effective. different actors.
• The dataset used for testing was
Fall Detection: limited to videos with only one
Key Metric: Rate of change of person per video (not practical)
angle between tracked object point
and ground plane

“An empirical study on fall-detection ML Algo: RF, kNN, DT • D’ size < D size (after clustering) • Dataset (D) partitioning into 6
using K-means based training data and hence low cost, faster training clusters to reduce dataset size(D’)
subset selection” Algo for dataset size reduction: K- times using k-means, then the traditional
means clustering algo • RF consistently performed the best algos (RF, kNN, DT) are performed
Dr. Kouichi Sakurai, Dr. Lopa Mandal, 6 Clusters : Standing, Walking, compared to kNN, DT. on D’.
Dr. Baisakhi Das, R. Jothi Sitting, Falling, Cramps, Running. • The reduced dataset (D') led to a • Some clusters did not align
(Then core points are selected from slight decrease in performance but perfectly with activity class labels.
American Journal of Advanced each cluster to form a reduced did not cause a significant drop in • Effectiveness could be further
Computing (AJAC), Vol. 1, Iss. 1, Art. 2 training set D′) detection rates. optimized by using mathematical
2024 model to compute core points.
8
 Literature Survey
Title, Author and Year Technique/Methodology used Result Remarks

“Design and Analysis for Fall Detection Sensors: infrared, • 94% accuracy, 99.6% specificity, • Combos of 7 sensor modalities + 3
System Simplification” electroencephalograph (EEG) 70.44% F1 Score window frames were analyzed
helmet, cameras • Random Forest (RF) produced the • 10-fold cross-validation(70%
Lourdes Martinez-Villaseñor, Hiram Dataset : Multimodal (UP-Fall best results. training 30% testing)
Ponce Detection21) • 3-second window+50% overlap • Fall detection system took samples
Algo: RF, SVM, kNN, Multi-Layer was best comb. from young, healthy subjects only
Journal of Visualized Experiments Perceptron (MLP) • The best sensor placements -waist, • Tripping or near-fall events were
2020 neck, right pocket not considered.

“Sensor Location Analysis and Minimal Camera placing: video-based, Experiment 1: Single IMU Sensor • RF consistently the top performers
Deployment for Fall Detection System” OpenPose, Kinect cameras method RF and SVM: 3 seconds window best • The lateral camera view and
MLP and KNN: 2 seconds waist-mounted IMU sensor was
Hiram ponce, Lourdes martínez- Sensor placing : Temporal Experiment 2: Single Camera most effective.
villaseñor, José nuñez-martínez segmentation, Sensor modality Lateral view (camera 1) best • Lack of handling class imbalance
RF was the most effective (requires synthetic data)
Special section on behavioral biometrics Algo: Decision trees (DT), random Experiment 3: Sensor Fusion • Uses synthetic data
for ehealth and well-being committee (RC), NaiveBayes (NB) IMU sensors +with lateral-view camera
2020 learning model, KNN, RF, SVM best, RF dominates still

9
 Research gaps
 Variety of Fall Types:
Current research mainly focuses on rear falls. Future studies should include other fall types,
like side or forward falls, to improve model robustness and generalization.
 Real-world Testing:
The system requires validation in real-world scenarios, considering varying sensor
placements, user demographics, and diverse activity types, to ensure reliable performance
outside controlled environments.
 Data Diversity:
Testing the system on a more diverse dataset, including various demographics (age, gender,
health conditions), will help enhance the model’s generalizability.
 Multimodal Data:
Incorporating multimodal data (e.g., video, pressure sensors, heart rate) alongside
accelerometer and gyroscope data can improve detection accuracy and robustness.
 Research gaps
 Synthetic Data and Simulations:
Techniques to generate synthetic datasets for simulating rare fall events are still
underdeveloped. Future research should explore the generation of realistic synthetic
datasets.
 Class Alignment and Clustering:
Challenges with class alignment during clustering need to be addressed, particularly in
distinguishing between similar activities such as sitting, standing, and walking.
 Handling Complex Scenarios:
Fall detection in complex, dynamic outdoor environments or under varying lighting
conditions should be further explored.
Key findings-
Common Machine Learning Algorithms in Fall Detection
 Decision Trees:
Used for classification and regression, creating a tree structure. Effective for large datasets
but computationally expensive.
 Random Forest:
Builds multiple Decision Trees and aggregates results for improved accuracy, requiring
more time and power.
 Support Vector Machine (SVM):
Creates a decision boundary between classes. Needs significant processing power but uses
less memory.
• K-Nearest Neighbours (KNN):
Classifies based on similarity to nearby data points. Simple and efficient for small
datasets, but sensitive to noise.
 Key findings-
Common fall detection datasets (public)
 UR Dataset: Multimodal dataset with accelerometer and Kinect camera sensors.
 UP-Fall Detection Dataset: A comprehensive multimodal dataset with wearable IMU sensors,
EEG headsets, and cameras.
 SisFall Dataset: A wearable dataset with accelerometer and gyroscope data.
 UmaFall Dataset: A wearable dataset with IMU sensors and smartphones.
 These datasets contain a mix of activities, falls, and participant ages, providing a diverse range
of data for training fall detection models. same with this
 Key findings-
Common tools and technologies
 Python:
Widely used in machine learning for its simplicity and the availability of powerful libraries
such as Pandas, Scikit-Learn, and SciPy for data manipulation, modeling, and analysis.
 MATLAB:
A standard tool for scientific computing, often used for testing and implementing machine
learning algorithms.
 SUMMARY
• The fall detection system utilizes various machine learning algorithms, including
Decision Trees (DT), K-Nearest Neighbors (KNN), and Random Forest (RF), to
classify falls based on sensor data such as accelerometer and gyroscope readings.

• Key methodologies include window length optimization, handling class imbalances,

and feature extraction from time and frequency domains.

• The system uses multiple classifiers to improve detection accuracy and robustness.
• Research gaps include the need for more diverse fall types, real-world testing,
integration with IoT, better sensor placement, and more efficient models for improved
scalability and performance.
 Proposed Solution:
A Multi-Activity Fall Detection System using
 Random Forest (RF)
 Artificial Neural Network (ANN) with LSTM Layers
 Data Preprocessing, Visualization
aimed at improving accuracy and scalability in real-world scenarios.
 How our approach addresses these gaps:
1. Variety of Fall Types
• Solution: Trained on diverse datasets (e.g., WEDA) using a sliding window approach for
sequence labeling.
• Gap Addressed: Captures temporal patterns, distinguishing falls from intentional movements.
2. Real-World Testing
 Solution: Processes sensor data from wearable devices with noise filtering and normalization.
 Gap Addressed: Ensures reliable performance in unpredictable, noisy real-world settings
3. Data Diversity
 Solution: Includes multisensor data from varied age groups and conditions.
 Gap Addressed: Reduces bias and improves accuracy across diverse populations and scenarios.
 How our approach addresses these gaps:
4. Handling Multimodal Data
Solution: Integrates accelerometer, gyroscope, and orientation data for comprehensive input.
Gap Addressed: Detects subtle changes in motion and balance for reliable fall detection.
5. Synthetic Data
Solution: Generates rare fall scenarios using GANs or physics-based simulations.
Gap Addressed: Prepares the model to handle edge cases effectively in real-world
deployments.
6. Class Alignment and Clustering
Solution: Applies class weights and feature selection to manage imbalanced datasets.
Gap Addressed: Improves differentiation between falls and non-falls, even with imbalanced
data.
Next Steps
1. Expand Activity Dataset
Extending the dataset to include additional activities (e.g., sitting, standing, walking)
and more fall types. Multi-activity classification can be better achieved by having a
more diverse and well-labeled dataset.
2. Optimize Feature Extraction
Refining and optimizing the feature extraction process by adding more features or
adjusting the existing ones (e.g., peak detection, spectral features). Better features can
significantly improve model performance.
3. Handle Class Imbalance More Effectively
Techniques like SMOTE (Synthetic Minority Over-sampling Technique) or cost-
sensitive learning.
Next Steps
4. Deploy on Wearable Hardware
Deploying our fall detection system on a physical wearable device, like a smartwatch or
wristband. This step is key to ensuring that our solution works outside of the development.
5. Test Real-World Data Collection
Conducting data collection experiments using real wearable devices (e.g., smartwatches) with
multiple participants in different environments (home, outdoor, etc.). Testing the system in
real-world scenarios helps ensure that the system can handle diverse conditions, environmental
factors, user behavior.
OUTPUT
OUTPUT
OUTPUT
OUTPUT
 CONCLUSION-
 Our fall detection project for the elderly aims to provide a critical solution for
identifying falls in real-time and notifying caregivers or medical
professionals. By leveraging sensors and advanced algorithms, the system
ensures timely intervention, preventing further injuries and providing peace of
mind to both the elderly and their families.
 The project demonstrates the integration of technology with healthcare,
emphasizing safety and prompt assistance. Its potential to improve the quality
of life for elderly individuals highlights the importance of innovation in
addressing age-related challenges. With further testing and optimization, our
system can significantly contribute to elderly care.
CITATIONS
•S. Usmani, A. Saboor, M. Haris, M. A. Khan, and H. Park, "Latest Research Trends in
Fall Detection and Prevention Using Machine Learning: A Systematic Review," Sensors,
vol. 21, no. 15, p. 5134, Jul. 2021, doi: 10.3390/s21155134.

•S. Almeraikhi, "Machine Learning Algorithms for Early Fall Detection of Elderly
People," J. Inf. Technol. Manag..

•R. Jothi, "An Empirical Study on Fall-Detection Using K-Means," Am. J. Adv. Comput.,
2020, doi: 10.1109/ACCESS.2020.3022971.

•H. Ponce, L. Martínez-Villasenor, and J. Nuñez-Martínez, "Sensor Location and

Minimal Deployment for Fall Detection," IEEE Access, vol. 8, pp. 1-12, 2020, doi:
10.1109/ACCESS.2020.3022971.
CITATIONS
•L. Martínez-Villasenor and H. Ponce, "Design and Analysis for Fall Detection
Simplification," J. Vis. Exp., vol. 158, e60361, Jun. 2020, doi: 10.3791/60361.

•S. G. Yoo and D. Oh, "An Artificial Neural Network-Based Fall Detection," Int. J. Eng.
Bus. Manag., vol. 10, pp. 1-7, 2018, doi: 10.1177/1847979018787905.

•R. Subburam, E. Chandralekha, and V. Kandasamy, "Elderly Fall Detection Using

Enhanced Random Forest," presented at the Int. Conf. Recent Advances in Science and
Eng., Dubai, UAE, Oct. 2023.

•T. S. KAVYA, Y.-M. Jang, E. Tsogtbaatar, and S.-B. Cho, "Fall Detection for Elderly
Using Vision-Based Analysis," Romanian J. Inf. Sci. Technol., vol. 23, no. 1, pp. 69–83,
2020.
Thank you.

28