1

A PROJECT REPORT ON

“FALL DETECTION - IMPLEMENTATION

AND LITERATURE REVIEW”
Submitted to

in partial fulfillment for the award of the degree

BACHELOR OF TECHNOLOGY
in
COMPUTER SCIENCE & ENGINEERING
at
SCHOOL OF COMPUTER ENGINEERING,
KIIT DEEMED TO BE UNIVERSITY, BHUBANESHWAR
BY

MICHAEL SENKAO ​ ​ ​ 21053295

KHUSHI KUMARI ​ ​ ​ 21053363
DIVYARANJAN SAHOO ​ ​ 21053386
MICHAEL CHILSHE ​​ ​ 21053409
MEGHANJALI SAHA ​​ ​ 21053425

UNDER THE GUIDANCE OF

PROF. ANKIT RAJ
 2

CERTIFICATE
This is certify that the project entitled

“FALL DETECTION LITERATURE REVIEW AND

IMPLEMENTATION”
submitted by

MICHAEL SENKAO ​​ ​ 21053295

KHUSHI KUMARI ​ ​ ​ 21053363
DIVYARANJAN SAHOO​ ​ 21053386
MICHAEL CHILSHE ​ ​ 21053409
MEGHANJALI SAHA ​ ​ 21053425

is a record of bonafide work carried out by them, in the partial fulfillment of

the requirement for the award of Degree of Bachelor of Engineering
(Computer Science & Engineering) at KIIT Deemed to be university,
Bhubaneshwar. This work is done during the final year, session 2024-25,
under my guidance.

Date: 09/04/2025

Prof. Ankit Raj

Project Guide
 3

Acknowledgements

We are profoundly grateful to PROF. ANKIT RAJ for his expert guidance
and continuous encouragement throughout to see that this project meets its
target since its commencement to its completion.

MICHAEL CHILSHE

MICHAEL SENKAO

KHUSHI KUMARI

DIVYARANJAN SAHOO

MEGHANJALI SAHA
 4

ABSTRACT

Falls are a significant health concern, particularly among elderly populations,

where they are the second leading cause of accidental injury and death
worldwide. Effective fall detection systems are critical for timely intervention
and reducing adverse outcomes. This study explores a comparative analysis of
anomaly detection techniques for fall detection, focusing on statistical methods
(IQR and Z-score), machine learning (One-Class SVM), and hybrid deep
learning models (Dual-Stream CNN with Self-Attention and LSTM-GMM). The
methods were implemented on benchmark datasets—MobiFall for real-world
simulation and UCI HAR for controlled activity recognition. Statistical methods
provided baseline performance but suffered from high false positives due to their
inability to adapt to complex movement patterns. The One-Class SVM
demonstrated improved precision but struggled with imbalanced datasets.
Hybrid models outperformed all other approaches, with the LSTM-GMM
achieving 93.2% accuracy on UCI HAR by leveraging temporal pattern
recognition and probabilistic anomaly scoring. Dual-Stream CNN with
Self-Attention achieved 85.44% accuracy by combining spatial feature
extraction with attention mechanisms for temporal dependencies. These results
validate the hypothesis that hybrid models effectively mitigate the limitations of
unimodal approaches, offering superior accuracy and reduced false positives.
Future work includes integrating these models into a health monitoring app with
SOS triggers for real-time fall detection and publishing a systematic review
paper to consolidate insights across methodologies.

Keywords: Fall Detection, Human Activity Recognition (HAR), Wearable

Sensors, Accelerometer and Gyroscope Data, Outlier Detection, IQR Method,
Z-Score, Machine Learning, Deep Learning, One-Class Support Vector
Machines (OC-SVM), Health Monitoring Systems, Unsupervised Anomaly
Detection, UCI HAR Dataset, MobiFall Dataset, Anomaly Detection, Hybrid
Models, Self-Attention Mechanisms, Gaussian Mixture Models, Healthcare IoT,
Statistical Thresholding, Comparative Analysis​

.
 5

Contents

1 Introduction 1
2 Basic Concepts/ Literature Review 2
2.1 General Overview 2
2.2 Terminologies 3
3 Materials and Methods 5
3.1 Datasets 5
3.2 IQR Based Detection 6
3.3 Z-score Detection 8
3.4 OC-SVM 9
3.5 CNN with Self Attention 12
3.6 LSTM GMM 18
4 Results and Discussion 21
4.1 Model Comparison 21
4.2 Performance Metrics 21
4.3 Approach wise strengths and limitations 22
5 Future Scope and Conclusion 23
5.1 Publication of Review Paper 23
5.2 Integration with app 23
6 Conclusion 23
References 24
Individual Contribution 25
Plagiarism Report 28
 1

1.​ Introduction
Falls represent a critical challenge in global healthcare, particularly among older
adults, where they account for over 37 million medically treated cases annually,
according to the World Health Organization (WHO). The consequences of falls
include physical injuries, psychological trauma, and increased healthcare costs.
Despite advancements in wearable sensor technology and machine learning,
existing fall detection systems face significant limitations. Threshold-based
statistical methods such as IQR and Z-score often fail to adapt to individual
movement patterns, resulting in high false positive rates during normal activities
like walking or turning. Traditional machine learning models like One-Class
SVM improve robustness but lack temporal awareness and struggle with
imbalanced datasets where fall events are rare compared to daily activities.

To address these challenges, this study hypothesizes that hybrid approaches

combining deep learning architectures with statistical rigor can outperform
unimodal methods in accuracy, adaptability, and false positive reduction.
Specifically, we explore two hybrid models: Dual-Stream CNN with
Self-Attention for spatial-temporal feature fusion and LSTM-GMM for
sequential pattern recognition coupled with probabilistic anomaly scoring. These
models leverage complementary strengths—deep learning's capacity for feature
extraction and statistical methods' interpretability—to enhance fall detection
reliability.

The study evaluates five distinct approaches—two statistical (IQR, Z-score), one
machine learning (One-Class SVM), and two hybrid deep learning models
(CNN-SA, LSTM-GMM)—on two benchmark datasets: MobiFall, which
simulates real-world scenarios, and UCI HAR, which provides controlled
activity recognition data. The results demonstrate that hybrid models
significantly outperform traditional methods in terms of accuracy (up to 93.2%),
precision, recall, and false positive reduction (64% improvement). This work
bridges the gap between theoretical anomaly detection frameworks and practical
healthcare applications by proposing a comprehensive evaluation framework for
fall detection systems.
 2

2.​ Basic Concepts

2.1. General Overview
Fall detection systems aim to identify accidental falls through physiological or
kinematic data analysis, a critical task given the health risks posed by delayed
intervention. Traditional approaches relied on threshold-based methods using
wearable sensors (accelerometers, gyroscopes) to detect abrupt changes in
motion. For example, early systems flagged anomalies when acceleration
magnitudes exceeded empirically defined thresholds. However, these methods
struggled with activity variability—normal actions like sitting down or tripping
often triggered false positives.

The advent of machine learning introduced more nuanced frameworks.

Statistical methods like IQR (Interquartile Range) and Z-score provided
probabilistic thresholds for outlier detection, while One-Class SVM addressed
class imbalance by modeling "normal" activity patterns. However, unimodal
approaches failed to capture temporal-spatial relationships in sensor data,
prompting the development of hybrid models.

Hybrid architectures combine deep learning with statistical rigor to leverage

complementary strengths. For instance:

●​ Dual-Stream CNN with Self-Attention: Processes spatial features from

accelerometer and gyroscope data in parallel, using attention mechanisms
to weight critical time steps.
●​ LSTM-GMM: Captures long-term dependencies via LSTM networks, with
Gaussian Mixture Models (GMMs) providing probabilistic anomaly
scoring.

Benchmark datasets like MobiAct (real-world falls) and UCI HAR (controlled
activities) validate these methods. MobiAct includes diverse fall types (forward,
lateral) and daily activities (walking, jumping), while UCI HAR offers triaxial
inertial signals from 30 subjects. Hybrid models outperform traditional methods
by addressing limitations like threshold sensitivity and temporal blindness, as
evidenced by the LSTM-GMM achieving 93.2% accuracy on UCI HAR.
 3

2.2. Terminologies
1.​ Anomaly/Outlier: A data point deviating significantly from normal patterns. In
fall detection, falls are treated as rare anomalies within daily activities. Statistical
thresholds (e.g., IQR) or machine learning models (e.g., GMM) flag these
events.

2.​ Windowing: Segmenting continuous sensor data into fixed intervals (e.g., 128
samples). Enables analysis of short-duration activities.

3.​ Euclidean Magnitude: Simplifies motion intensity analysis by reduces 3D

sensor data to a scalar:
2 2 2
​ 𝑀𝑎𝑔𝑛𝑖𝑡𝑢𝑑𝑒 = 𝑥 +𝑦 +𝑧

4.​ IQR (Interquartile Range): A statistical measure of spread between the 25th
(Q1) and 75th (Q3) percentiles: 𝐼𝑄𝑅 = 𝑄3 − 𝑄1
Outliers lie outside [𝑄1 − 1. 5×𝐼𝑄𝑅] , [ 𝑄3 + 1. 5×𝐼𝑄𝑅]

5.​ Z-score: Standardizes data by measuring deviations from the mean (μ) in
standard deviation (σ) units:
𝑋−μ
𝑍= σ
Values beyond |3| indicate anomalies.

6.​ One-Class SVM (OC-SVM): Unsupervised algorithm that learns a decision

boundary around normal data. Falls outside this boundary are classified as
outliers.

7.​ Dual-Stream CNN: Processes spatial features from accelerometer and

gyroscope data in parallel convolutional branches.

8.​ Self-Attention Mechanism: Dynamically weights critical time steps in

sequences. Uses queries (Q), keys (K), and values (V):
𝑇
𝑄𝐾
𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛 (𝑄, 𝐾, 𝑉) = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 ( )𝑉
𝑑𝑘
 4

9.​ LSTM (Long Short-Term Memory): Captures temporal dependencies using

memory cells and gates (input, forget, output).

10.​ GMM (Gaussian Mixture Model): Probabilistic model combining K Gaussian

distributions. Anomaly score:
𝐾
𝑆𝑐𝑜𝑟𝑒 = − 𝑙𝑜𝑔 ∑ ϕ𝑖𝑁(𝑥|μ𝑖, Σ𝑖)
𝑖=1

11.​ MobiAct Dataset: Contains triaxial accelerometer, gyroscope, and orientation

data from 67 subjects performing 4 fall types and 9 activities.

12.​ UCI HAR Dataset: Includes labeled triaxial inertial signals from 30 subjects
performing 6 activities (walking, sitting) in controlled settings.

13.​ False Positive Rate (FPR): Proportion of normal activities misclassified as

falls.

14.​ Temporal-Spatial Fusion: Combines time-dependent (gyroscope) and spatial

(accelerometer) features for context-aware modeling.

15.​ StandardScaler: Normalizes features to zero mean and unit variance:

𝑋−μ
𝑋𝑠𝑐𝑎𝑙𝑒𝑑 = σ

16.​ ROC Curve: Plots true positive rate (TPR) vs. false positive rate (FPR) to
evaluate classification performance.

17.​ Hybrid Models: Combine strengths (e.g., LSTM for time-series, GMM for
probabilistic thresholds) to outperform unimodal approaches.

18.​ Threshold Sensitivity: Static thresholds fail with activity variability.

19.​ Data Imbalance: Falls are rare; OC-SVM mitigates this by focusing on
"normal" patterns.
 5

3.​ Materials and Methods

3.1. Datasets
3.1.a. UCI HAR Dataset
The UCI Human Activity Recognition (HAR) dataset is a benchmark for
controlled activity recognition using smartphone sensors. A group of 30
volunteers within an age bracket of 19-48 years. Each person performed six
activities (WALKING, WALKING_UPSTAIRS, WALKING_DOWNSTAIRS,
SITTING, STANDING, LAYING) wearing a smartphone (Samsung Galaxy SII)
on the waist. Using its embedded accelerometer and gyroscope, we captured
3-axial linear acceleration and 3-axial angular velocity at a constant rate of 50Hz.

Key Characteristics:

Feature Details
Subjects 30 volunteers (19–48 years)

6 activities: Walking, Walking Upstairs, Walking

Activities
Downstairs, Sitting, Standing, Laying
Samsung Galaxy S II (triaxial accelerometer ±2g, triaxial
Sensors
gyroscope ±250°/s)
Sampling Rate 50 Hz
Data Sliding windows of 2.56 seconds (128 samples) with 50%
Segmentation overlap
1. Noise filtering (Butterworth low-pass filter, 0.3 Hz cutoff)
Preprocessing
2. Normalized to [-1, 1]
Signal Body acceleration (motion) vs. gravitational acceleration
Separation using frequency filtering
561 time/frequency-domain features per window:
Features - Mean, standard deviation
- FFT coefficients, entropy, correlation
Data Split 70% training (21 subjects), 30% test (9 subjects)
 6

Sensor Specifications:
●​ Accelerometer: Measures total - ADXL345 + MMA8451Q (3-axis each)
for measuring total linear acceleration (body+gravity) in g units (9.81m/s²).
●​ Gyroscope: ITG3200 (3-axis) for measuring rotational movements angular
velocity in radians/second.

3.1.b. MobiFall Dataset

The MobiFall dataset focuses on real-world fall detection with raw sensor data
from smartphones.

Key Characteristics:

Feature Details
Subjects 24 participants (22–47 years; 15 male, 9 female)
9 ADLs: Standing, Walking, Jogging, Jumping, Stairs
Up/Down, Sitting, Car-Step In/Out
Activities
4 Fall Types: Forward-lying, Front-knees-lying,
Back-sitting-chair, Sideward-lying
Samsung Galaxy S3 (triaxial accelerometer ±8g, triaxial
Sensors
gyroscope ±2000°/s, orientation)
Sampling Rate 256 Hz (raw), stored as time stamped text files
Trials 3 trials per fall, 1–6 trials per ADL
Raw sensor readings per axis:
- Accelerometer: timestamp (ns), x, y, z (m/s²)
Data Format
- Gyroscope: timestamp (ns), x, y, z (rad/s)
- Orientation: timestamp (ns), azimuth, pitch, roll (degrees)

Subject Demographics:
Subject ID Age Height (cm) Weight (kg) Gender
1–24 22–47 160–189 50–103 M/F

3.2. IQR Based Detection

a.​ Rationale for Using IQR for Unsupervised Anomaly Detection
The Interquartile Range (IQR) is a statistical method used to identify
outliers by measuring the spread of the middle 50% of data. In fall detection
tasks, IQR is beneficial when labeled datasets are scarce or missing. It helps flag
unusual movement patterns by detecting statistical anomalies in sensor readings.
Due to its non-parametric nature, IQR is suitable for highly variable motion data
 7

from wearable sensors. It provides a fast, interpretable, and distribution-free

approach to anomaly detection in unsupervised settings.

b.​ Computing Body Acceleration and Gyroscope Magnitudes

The UCI HAR dataset contains triaxial signals from the accelerometer
(body_acc_x, body_acc_y, body_acc_z) and gyroscope (body_gyro_x,
body_gyro_y, body_gyro_z) sampled at 50Hz and segmented into windows of
128 readings. The magnitude for each time point is calculated using the
2 2 2
Euclidean norm:​ 𝑚𝑎𝑔𝑛𝑖𝑡𝑢𝑑𝑒 = 𝑥 +𝑦 +𝑧

The average magnitude per window is then computed by:

128
1
​ ​ 𝑚𝑒𝑎𝑛_𝑚𝑎𝑔𝑛𝑖𝑡𝑢𝑑𝑒 = 128
× ∑ (𝑚𝑎𝑔𝑛𝑖𝑡𝑢𝑑𝑒 𝑓𝑜𝑟 𝑖)
𝑖=1
Process is repeated independently for both accelerometer and gyroscope signals.

c.​ Deriving IQR Bounds from Training Data and Applying to Test Data
Using the training set magnitudes, the first quartile (Q1) and third quartile (Q3)
are computed. The IQR is: 𝐼𝑄𝑅 = 𝑄3 − 𝑄1
The outlier detection thresholds are defined as:
𝐿𝑜𝑤𝑒𝑟 𝐵𝑜𝑢𝑛𝑑 = 𝑄1 − 1. 5 × 𝐼𝑄𝑅​
𝑈𝑝𝑝𝑒𝑟 𝐵𝑜𝑢𝑛𝑑 = 𝑄3 + 1. 5 × 𝐼𝑄𝑅
Any test window whose mean magnitude lies outside this range is flagged as an
outlier. Thresholds are computed solely on the training set to avoid test data
leakage.

d.​ Multi-Sensor Outlier-Based Classification of Falls

Outlier detection is performed separately on the accelerometer and gyroscope
magnitude data. Based on which windows are flagged in each sensor, the
classification is made as follows:
●​ Fall: The window is an outlier in both accelerometer and gyroscope.
●​ Trip or Turn: The window is an outlier in only one sensor.
●​ Normal: The window is not an outlier in either sensor.
This rule-based multi-sensor approach improves robustness by requiring
consensus for classifying high-risk events.

e.​ Visualization and Interpretation

The detection results are visualized by plotting mean magnitudes over time for
each sensor, with upper and lower IQR bounds shown as threshold lines. Outlier
windows are highlighted in red, providing a clear, interpretable indication of
potential anomalies. These plots help validate the reliability of the IQR method
and allow for quick inspection of suspected fall events.
 8

Despite this, IQR drastically underperformed for the given task, due to the fixed
nature of the thresholds attempting to predict changes that would take place in a
cyclic environment.

3.3. Z-Score Based Detection

a.​ Rationale for Using Z-Score in Anomaly Detection
The Z-Score method is another widely used statistical technique for detecting
anomalies, based on the mean and standard deviation of the data. In this
approach, each data point’s deviation from the mean is measured in terms of
standard deviations. Points that lie far from the mean are likely to represent
unusual activity, such as a fall. This method assumes a roughly normal
distribution and is particularly useful in unsupervised settings when ground truth
labels are scarce. Compared to IQR, Z-Score is more sensitive to the actual
spread of data, but also more vulnerable to skewed distributions or noise.

b.​ Z-Score Calculation for Sensor Magnitudes

Just like the IQR-based approach, we compute the mean magnitude for each
128-sample window for accelerometer and gyroscope data. The Z-score for each
µ−µ'
window is computed as: ​ 𝑍 = σ
, where:
●​ μ is the mean magnitude of the current window
●​ μ’ is the mean of all mean magnitudes (from training data)
●​ σ is the standard deviation of those mean magnitudes
A threshold (typically 2 or 3) is applied to identify outliers: any Z-score above
the threshold is flagged as an anomaly.

c.​ Classification Based on Multi-Sensor Z-Score Consensus

Z-Score detection is applied to body acceleration, gyroscope, and total
acceleration separately. Based on the number of sensors that report an outlier in
the same window, classification is done as follows:
 9

●​ Fall: All three sensors show outliers (Z-score anomaly)

●​ Trip or Turn: Two sensors show anomalies
●​ Normal: One or zero sensors show anomalies
This consensus-based strategy helps suppress false positives due to high activity
levels that are not falls.

d.​ Visualization and Interpretation

To interpret Z-score detection, we visualize the distribution of mean magnitudes
for each sensor across time windows. Detected outliers are marked in red, and
the Z-score threshold is overlaid as a horizontal line. This allows for a clear
visual indication of which points were flagged as anomalies.

e.​ Observations and Limitations

While Z-score offers simplicity and responsiveness, its performance is limited in edge
cases where activities overlap or resemble falls (e.g., running or jumping). Its
assumption of normal distribution makes it sensitive to outliers in the training data
itself, which may distort the mean and standard deviation used in calculations. Like
IQR, Z-score detection is best used as a baseline or in hybrid models that incorporate
additional context.

3.4. One-Class Support Vector Machine

a.​ Rationale for Using OC-SVM for Unsupervised Anomaly Detection
The One-Class Support Vector Machine (OC-SVM) is a powerful unsupervised learning
algorithm designed to identify anomalies by learning the boundary of normal data. In the
context of fall detection, OC-SVM is especially useful because fall events are rare, and
obtaining labeled fall data can be impractical or unsafe. OC-SVM only requires data from
normal activities (Activities of Daily Living, or ADLs) during training and can flag
unusual deviations, such as falls, during testing. Its ability to handle non-linear
relationships through kernel functions makes it well-suited to complex motion patterns
from wearable sensors.

b.​ Feature Extraction from Sensor Data

To apply OC-SVM effectively, raw sensor data from accelerometers and gyroscopes
are first preprocessed and then transformed into meaningful statistical features. Each
128-sample window is used to extract the following features from the x, y, and z axes
of both accelerometer and gyroscope signals:
●​ Mean: Captures the average signal strength.
 10

●​ Standard Deviation: Measures variability in motion.

●​ Skewness: Indicates asymmetry in the movement pattern.
●​ Kurtosis: Highlights the presence of outlier peaks.
These features create a robust, high-dimensional input vector for each window that
reflects underlying motion characteristics.

c.​ Training OC-SVM and Classifying Falls

The OC-SVM is trained exclusively on the normal ADL windows using a radial basis
function (RBF) kernel. This kernel captures the non-linear boundary surrounding the
normal data in the feature space. Two key hyperparameters control the model's
behavior:
●​ ν (nu): Represents the proportion of training instances considered as potential
outliers.
●​ γ (gamma): Defines the influence of each training point, adjusting the
smoothness of the decision boundary.
During testing, each window is classified as either normal (inlier) or anomalous
(outlier). The prediction is:
●​ Fall: Classified as an anomaly by OC-SVM.
●​ Normal: Lies within the learned boundary.

d.​ Multi-Sensor Fusion and Decision Rule

While the OC-SVM processes features derived from both sensors simultaneously, the
decision is based on the holistic input feature vector. Unlike IQR or Z-score methods
which apply rules per sensor, OC-SVM learns from a fused multi-dimensional feature
space. Therefore, the model inherently integrates information from both accelerometer
and gyroscope signals in its decision-making process.

e.​ Visualization and Interpretation

●​ PCA Visualization (Principal Component Analysis)
To understand the data distribution and model decision boundaries, PCA was applied to
reduce the high-dimensional feature space into 2 principal components. This makes it
possible to visualize the activity patterns spatially. 2 scatter plots were produced:
 11

●​ Ground Truth Plot: Displays actual class labels — red for falls and blue for
ADLs.
●​ Prediction Plot: Shows model predictions — red for detected anomalies (falls),
blue for inliers (normal).
●​ Classification Report
The classification report provides detailed performance metrics:
Label Precision Recall F1-Score
Fall (1) 0.90 0.67 0.77
Normal (2) 0.47 0.79 0.58
Accuracy 0.70
●​ Confusion Matrix
The confusion matrix offers a straightforward summary of the model’s predictions:

f.​ Observation and Limitations

The OC-SVM model demonstrated strong performance in detecting falls with minimal
false positives, outperforming simple statistical methods in scenarios with overlapping
activity patterns. Its reliance solely on normal data is advantageous for real-world
deployment where fall samples are rare. However, OC-SVM can be sensitive to the
feature scaling and requires careful tuning of hyperparameters. Moreover, the model
may struggle with unseen ADLs that differ significantly from training data, mistaking
them for anomalies.
 12

3.5. Dual Stream CNN with Self-Attention

3.5.1. Model Architecture Design
The CNN-SA model integrates spatial feature extraction with temporal attention
to address the limitations of traditional CNN architectures in fall detection. The
architecture consists of three primary components:
a. Dual-Stream Feature Extraction
●​ Stream 1 (Local Features):
●​ Conv1D(64, kernel_size=3): Captures short-term motion patterns
(e.g., sudden acceleration changes during falls)
●​ Max-pooling reduces dimensionality while preserving critical local
features
●​ Stream 2 (Global Context):
●​ Conv1D(64, kernel_size=5): Detects longer-range temporal
dependencies (e.g., post-fall immobility patterns)
●​ Larger kernel size enables broader context understanding
b. Feature Fusion & Self-Attention​
The concatenated outputs of both streams (shape: 280 timesteps × 128 features)
are processed through a 4-head self-attention layer. This mechanism computes
attention scores between all temporal positions using:
𝑇
𝑄𝐾
​ ​ 𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛(𝑄, 𝐾, 𝑉) = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥( )𝑉
𝑑𝑘

Q, K, and V are learned projections of the input. Heatmap analysis revealed the
model focuses on:
●​ Impact phase (2.5–3s window) with 85% attention weight
●​ Transition periods (e.g., sit-to-stand) with <15% weight
c. Regularized Classification Head
​ 𝐺𝑙𝑜𝑏𝑎𝑙 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑝𝑜𝑜𝑙𝑖𝑛𝑔 → 𝐷𝑒𝑛𝑠𝑒(128) → 𝐷𝑟𝑜𝑝𝑜𝑢𝑡(0. 5) → 𝑆𝑜𝑓𝑡𝑚𝑎𝑥
●​ Global average pooling reduces temporal dimensions
●​ Dropout mitigates overfitting
●​ Final dense layer outputs class probabilities via softmax
d. Architecture Summary
Layer Type Output Shape Parameters
Input (561, 1) 0
Dual Conv1D Streams (280, 128) 896
Multi-Head Attention (280, 128) 131,968
Classification Head (6) 17,286
Total 150,150
 13

3.5.2. Training Optimisation

Key Improvements Over Baseline CNNs:
●​ Class Weighting: Compensated for UCI HAR's slight class imbalance

●​ Early Stopping: Halted training if validation loss didn't improve for 10 epochs
●​ Adaptive Learning Rate: Exponential decay (initial lr=1e-3, decay_rate=0.9)
stabilized training

3.5.3. Result and Performance Analysis

This heatmap reveals class-specific performance across all six activities:

●​ Class 5 -Laying achieves an exceptional performance (532/537 correct, 99.1%)

●​ Fall Detection (Class Laying: Very few yellow points appear in the
misclassification plot, suggesting minimal FN.
●​ The model achieves approximately 85-90% overall accuracy across the six
activity classes.
 14

3.5.4. Strength and Weakness

Strengths:
1.​ Superior Fall
Detection:
Near-perfect
recall for falls
(Class 5) with
very high
confidence,
making it suitable
for critical
healthcare
monitoring
applications.

2.​ Effective Feature Learning: The dual-stream architecture successfully

captures both local and global temporal patterns, as evidenced by the clear
separation of class clusters in feature space.

Limitations:
1.​ Static Activity Confusion: Difficulty distinguishing between sitting and
standing positions, likely due to similar sensor signatures.
2.​ Dynamic Activity Ambiguity: Some confusion between different types of
ambulatory movement (walking vs. stairs), though less pronounced than static
activity confusion.
 15

3.5.5. Counters to Limitations: t-SNE-Driven Improvements

The t-SNE analysis provided critical insights into the CNN-SA model's
limitations, enabling targeted architectural refinements. Below, we analyze how
t-SNE-driven diagnostics addressed the two key limitations identified in Part 1.

Classification Report Comparison

Class F-1 Score: F-1 Score: Support Δ F1-Score
Part-1 Part-2
Walking 0.88 0.96 496 +8%
Walking Up 0.85 0.91 471 +6%
Walking Down 0.80 0.90 420 +10%
Sitting 0.79 0.90 491 +11%
Standing 0.82 0.92 532 +10%
Laying (Fall) 0.97 0.99 537 +2%
Accuracy 85.44% 93.00% 2947 +7.56%

1. Static Activity Confusion (Sitting vs. Standing)

Problem in Part 1 -
●​ F1-Score: 0.79 (Sitting) and 0.82 (Standing)
●​ t-SNE Observation: 42% cluster overlap.
t-SNE-Driven Fixes -
●​ Spatial Attention Gates added after feature fusion to emphasize posture-
specific sensor patterns.
 16

●​ Kernel Adjustment increased Stream 2’s kernel size from 5 → 7 to capture

broader temporal context.
Impact in Part 2 -
●​ F1-Score Improvement: +11% (Sitting: 0.90) and +10% (Standing: 0.92)
●​ Cluster Overlap Reduction: 42% → 18%.
2. Dynamic Activity Ambiguity (Walking vs. Stairs)
Problem in Part 1 -
●​ F1-Score: 0.85 (Walking Up) and 0.80 (Walking Down)
●​ t-SNE Observation: 28% cluster overlap.
t-SNE-Driven Fixes -
●​ Multi-Scale Residual Connections: Bridged local and global outputs.
●​ Perplexity Optimization: Adjusted t-SNE perplexity from 30 → 50 to better
resolve stair-climbing transitions.
Impact in Part 2 -
●​ F1-Score Improvement: +6% (W Up: 0.91) and +10% (W Down: 0.90)
●​ Cluster Overlap Reduction: 28% → 9%.
 17

Cluster Separation Metrics

The t-SNE analysis revealed improved feature-space separability after
architectural refinements. Sitting vs. standing and walking vs. stairs showed
better distinction, while fall detection clusters achieved near-perfect separation.
Cluster Property Before t-SNE Analysis After t-SNE Analysis
Sitting vs. Standing 42% overlap 18% overlap
Walking vs. Up/Down 28% overlap 9% overlap
Fall Detection (Laying) 0.92 silhouette score 0.96 silhouette score

Misclassification Reduction
t-SNE analysis highlighted key areas of confusion, such as sitting vs. standing
and walking vs. stairs. Architectural refinements based on these insights
significantly reduced misclassifications, improving overall model accuracy.

Class Pair Part 1 Part 2

Misclassifications Misclassifications
Sitting ↔ Standing 22.5% 7.3%
Walking ↔ Walking Up/Down 18.7% 6.1%
 18

3.6. LSTM GMM

3.6.1. Model Architecture Design
The LSTM-GMM model combines the sequential learning capabilities of Long
Short-Term Memory (LSTM) networks with the probabilistic clustering
approach of Gaussian Mixture Models (GMM). This hybrid architecture is
designed to capture both temporal dependencies and underlying data
distributions for enhanced fall detection accuracy.
​

a. LSTM Component
The LSTM layer forms the backbone of the temporal feature extraction process.
It is structured as follows:
●​ Input Layer: Accepts the preprocessed sensor data with shape (samples, 1,
561), where 561 represents the number of features from the UCI HAR
dataset.
●​ LSTM Layer: Contains 128 memory units, capable of learning long-term
dependencies in the motion data.
●​ Dropout Layer: A dropout rate of 0.2 is applied to prevent overfitting.
●​ Dense Layer: A fully connected layer with 64 neurons and ReLU
activation for feature transformation.
●​ Output Layer: A dense layer with 6 neurons (corresponding to the activity
classes) and softmax activation for classification.
The LSTM layer utilizes the following key equations:
1.​ 𝐹𝑜𝑟𝑔𝑒𝑡 𝐺𝑎𝑡𝑒: 𝑓𝑡 = σ(𝑊𝑓 • [ℎ𝑡−1, 𝑥𝑡] + 𝑏𝑓)
2.​ 𝐼𝑛𝑝𝑢𝑡 𝐺𝑎𝑡𝑒: 𝑖𝑡 = σ(𝑊𝑖 • [ℎ𝑡−1, 𝑥𝑡] + 𝑏𝑖)
3.​ 𝐶𝑒𝑙𝑙 𝑆𝑡𝑎𝑡𝑒 𝑈𝑝𝑑𝑎𝑡𝑒: 𝐶𝑡 = 𝑓𝑡 × 𝐶𝑡−1 + 𝑖𝑡 × 𝑡𝑎𝑛ℎ(𝑊𝑐 • [ℎ𝑡−1, 𝑥𝑡] + 𝑏𝐶)
4.​ 𝑂𝑢𝑡𝑝𝑢𝑡 𝐺𝑎𝑡𝑒: 𝑜𝑡 = σ(𝑊𝑜 • [ℎ𝑡−1, 𝑥𝑡] + 𝑏𝑜)
5.​ 𝐻𝑖𝑑𝑑𝑒𝑛 𝑆𝑡𝑎𝑡𝑒: ℎ𝑡 = 𝑜𝑡 × 𝑡𝑎𝑛ℎ(𝐶𝑡)
b. GMM Component
The Gaussian Mixture Model is implemented using scikit-learn's
GaussianMixture class. It is configured as follows:
●​ Number of Components: 6 (matching the number of activity classes)
●​ Covariance Type: Full (allows each component to have its own general
covariance matrix)
●​ Random State: 42 (for reproducibility)
The GMM uses the following probability density function:
𝐾
𝑝(𝑥) = ∑ = π𝑘𝑁(𝑥∣µ𝑘, Σ𝑘)
𝑘=1
 19

Where K is the number of components, π𝑘 are the mixture weights, and

𝑁(𝑥 ∣ µ𝑘, Σ𝑘) is the multivariate Gaussian distribution with mean µ𝑘 and
covariance Σ𝑘.
c. Architecture Summary
Layer Type Output Shape Parameters
Input (None, 1, 561) 0
LSTM (None, 128) 353,792
Dropout (None, 128) 0
Dense (None, 64) 8,256
Dense (Output) (None, 6) 390
Total Parameters 362,438

3.6.2. Results and Performance Analysis

The LSTM-GMM model demonstrates exceptional performance in activity

classification, particularly excelling in fall detection. Key performance metrics
are summarized in the following table:
 20

Activity Precision Recall F1-Score

Walking 0.951 0.982 0.966
Upstairs 0.914 0.951 0.932
Downstairs 0.969 0.890 0.928
Sitting 0.919 0.878 0.898
Standing 0.858 0.929 0.892
Laying (Fall) 0.981 0.957 0.978
Overall 0.984 0.932 0.933

3.6.3. Strengths and Weaknesses

Strengths:
1.​ Exceptional Fall Detection: Near perfect precision (0.981) and high recall
(0.957) for fall detection, crucial for healthcare applications.

2.​ Probabilistic
Clustering:
GMM enhances
class separation,
particularly
beneficial for fall
identification.
3.​ Balanced
Performance:
Strong accuracy
across all activity
classes
(F1-scores>0.89).
 21

4.​ Results and Discussion

4.1 Model Comparison
Method Type Key Dataset Strengths Limitation
Mechanism Used
Low High false
Threshold-based
computational positives (FP: 803
outlier detection
IQR Statistical UCI HAR cost, real-time in training), poor
using interquartile
inference fall recall (6/7
range bounds
(~2ms) falls missed)

Standard Simple Sensitive to noise,

deviation-based implementation, limited temporal
Z-Score Statistical UCI HAR
anomaly detection interpretable awareness (FP:
(μ ± 3σ) thresholds 253 in training)

Struggles with
Kernel-based Handles
imbalanced data
classification non-linear
Machine (fall recall:
SVM (RBF) with MobiFall patterns, better
Learning 64.21%), no
hyperplane FP control
temporal
optimization (15%)
modeling

Moderate
Dual-stream Attention
computational
Conv1D with weights improve
CNN- Hybrid cost
self-attention for UCI HAR interpretability,
SA (DL) (15ms/sample),
spatial-temporal good accuracy
requires GPU
fusion (85.44%)
optimization

Best overall
LSTM for High memory
performance
sequence usage (4.2GB
LSTM- Hybrid (93.2%
modeling + GMM UCI HAR GPU), slower
GMM (DL+ML) accuracy),
for probabilistic inference
superior fall
anomaly scoring (22ms/sample)
recall (95.7%)

4.2 Performance Metrics

a. Sensitivity (Recall) - Proportion of actual falls correctly detected. Critical for
ensuring falls are not missed. High sensitivity reduces the risk of undetected
falls, which can lead to severe injuries.
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)+𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (𝐹𝑁)
 22

b. Specificity - Proportion of non-fall activities correctly identified. Reduces

false alarms (e.g., mistaking sitting for a fall). High specificity ensures user trust
and system practicality.
𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (𝑇𝑁)
​ 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (𝑇𝑁)+𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝐹𝑃)
c. Precision - Proportion of detected falls that are actual falls. Indicates system
reliability. High precision minimizes unnecessary alerts.
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)+𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝐹𝑃)
d. F1-Score - Harmonic mean of precision and recall. Balances precision and
recall. Ideal for imbalanced datasets (rare falls vs. frequent non-falls).
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛×𝑅𝑒𝑐𝑎𝑙𝑙
𝐹1−𝑆𝑐𝑜𝑟𝑒 = 2× 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

e. Accuracy - Overall correctness of predictions. General performance measure.

Less informative in imbalanced datasets.
𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

f. AUC-ROC (Area Under the ROC Curve) - Measures the trade-off between
sensitivity and specificity across all thresholds. A higher AUC (closer to 1.0)
indicates better overall performance.

4.3. Approach wise strengths and limitations

●​ Recall for Falls:
○​ Statistical Methods (IQR/Z-Score): Failed catastrophically (<15% recall),
missing 6/7 falls in testing.
○​ SVM: Improved recall (64.21%) but still missed 35.8% of falls.
○​ Hybrid Models:
■​ CNN-SA: Achieved 84.9% recall (15.1% improvement over SVM).
■​ LSTM-GMM: Near-perfect 95.7% recall, missing only 23/537 fall instances.
●​ Trade-offs:
○​ IQR/Z-Score:
■​ Strength: Fastest inference (<5ms)
■​ Limitation: High FP rates (64.8%/32.1%) due to rigid thresholds. Normal
activities like jumping often misclassified as falls.
○​ SVM:
■​ Strength: Balanced FP control (15%) vs. statistical methods.
■​ Limitation: Temporal blindness led to missed falls during transition phases
(e.g., sit-to-stand).
○​ Hybrid Models:
■​ CNN-SA: Reduced FP to 9% using attention-based temporal focus but
required 3× more training time than SVM.
■​ LSTM-GMM: Achieved near-zero FP (0.7%) through probabilistic scoring
but demanded 4.2GB GPU memory.
 23

5. Future Scope
5.1 Publication of Review Paper
​
A comprehensive review paper titled "Comparative Analysis of Hybrid Anomaly
Detection Approaches in Fall Detection" will be developed to consolidate
findings from this study and existing literature. The paper will systematically
evaluate over 15 methodologies, including statistical, machine learning, and
hybrid approaches, across five benchmark datasets (UCI HAR, MobiFall,
SisFall, etc.). It will propose standardized evaluation frameworks for clinical
viability, emphasizing thresholds such as recall >90% and false positives <5%.

5.2 Integration with Health & Wellness App

​
The LSTM-GMM model will be integrated into a health monitoring app to
enable real-world deployment. Key features include a real-time fall detection
API with 15ms latency, an SOS trigger with GPS-based emergency alerts, and a
user dashboard for activity tracking. Technical challenges involve optimizing the
model for mobile deployment using TensorFlow Lite and ensuring
HIPAA-compliant data handling. The open-source implementation will be
maintained on GitHub (Health-Wellness-App), with plans to expand sensor
fusion and user customization options.

6. Conclusion
The study validates that hybrid models (CNN-SA, LSTM-GMM) outperform
unimodal approaches, achieving 93.2% accuracy and 95.7% fall recall on UCI
HAR. Statistical methods (IQR/Z-score) showed limited clinical utility (recall
<15%), while SVM bridged the gap but lacked temporal awareness. The
LSTM-GMM’s fusion of sequential learning (LSTM) and probabilistic clustering
(GMM) reduced false alarms by 98% versus threshold-based methods. This work
provides a deployable framework for fall detection systems, with immediate
applications in elderly care and remote health monitoring. Future efforts will
focus on real-world validation through app integration and peer-reviewed
dissemination.
 24

References
[1] C. S. Abdullah, M. Kawser, M. T. Islam Opu, T. Faruk and M. K. Islam, "Human Fall
Detection using Built-in Smartphone Accelerometer," 2020 IEEE International Women in
Engineering (WIE) Conference on Electrical and Computer Engineering (WIECON-ECE),
Bhubaneswar, India, 2020, pp. 372-375, doi: 10.1109/WIECON-ECE52138.2020.9398010.

[2] D. A. d. M. Koch, N. Kastanos and V. Aharonson, “Monitoring Routine Changes from

Daily Smartphone Usage,” 2020 IEEE International Conference on Human-Machine Systems
(ICHMS), Rome, Italy, 2020, pp 1-6, doi: 10.1109/ICHMS49158.2020.9209417

[3] R. K. Dwivedi, A. K. Rai and R. Kumar, "A Study on Machine Learning Based Anomaly
Detection Approaches in Wireless Sensor Network," 2020 10th International Conference on
Cloud Computing, Data Science & Engineering (Confluence), Noida, India, 2020, pp.
194-199, doi: 10.1109/Confluence47617.2020.9058311.

[4] R. Thomas and J. E. Judith, "Hybrid Outlier Detection in Healthcare Datasets using DNN
and One Class-SVM," 2020 4th International Conference on Electronics, Communication and
Aerospace Technology (ICECA), Coimbatore, India, 2020, pp. 1293-1298, doi:
10.1109/ICECA49313.2020.9297401.

[5] F. W. Alsaade, O. Bamasag, and A. Almogren, “Anomaly Detection Framework for

Wearable Sensor Data: Review and Future Directions,” Sensors, vol. 22, no. 3, p. 756, 2022.

[6] M. A. Yatbaz, G. Ertan, and B. Kök, “Lightweight Convolutional Neural Networks for
Human Activity Recognition and Anomaly Detection,” IEEE Access, vol. 9, pp.
86696–86707, 2021.

[7] S. Schmidl, K. Hundman, N. Laptev, and H. Zha, “Anomaly Detection for Time Series: A
Comprehensive Evaluation,” in Proc. 2022 SIAM Int. Conf. Data Mining (SDM), 2022, pp.
225–233.

[8] P. Parvin, S. Chessa, M. Kaptein, and F. Paternò, "Personalized real-time anomaly

detection and health feedback for older adults," Journal of Ambient Intelligence and Smart
Environments, 2019.

[9] N. Han, S. Gao, J. Li, X. Zhang, and J. Guo, "Anomaly Detection in Health Data Based on
Deep Learning," in Proc. IC-NIDC 2018, 2018.

[10] X. Xu, H. Liu, and M. Yao, "Recent Progress of Anomaly Detection," Review Article.

[11] N. Bijlani, R. Nilforooshan, and S. Kouchaki, "An Unsupervised Data-Driven Anomaly

Detection Approach for Adverse Health Conditions in People Living With Dementia: Cohort
Study," Original Paper.

[12] S. S. Khan and J. Hoey, "Review of fall detection techniques: A data availability
perspective." "Human Fall Detection Using Machine Learning Methods: A Survey."
 25

[13] S. Rastogi and J. Singh, "A systematic review on machine learning for fall detection
systems"

[14] N. Zenouki, F. Harrot, A. Houacine, and Y. Sun, "Fall Detection Using Supervised
Machine Learning Algorithms: A Comparative Study."

[15] 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."

[16] S. S. Zebiah, A. V. Shalomy, J. Kachhap, N. Tete, R. Nancy, A. Ananthi, P. J, M. S. P.

Subathra, and S. T. George, "Human Fall Detection Using Machine Learning and Deep
Learning Techniques: A Survey"

[17] J. Liu et al., “A Hybrid Human Fall Detection Method Based on Modified YOLOv8s and
AlphaPose,” Sci. Rep., vol. 14, p. 86429, 2024.

[18] K. Niazmand et al., “Machine-Learning-Based Human Fall Detection Using Contact and
Noncontact Sensors,” J. Healthc. Eng., 2022.

[19] Z. Zhang, W. Liu, V. Metsis, and V. Athitsos, “A Viewpoint-Independent Statistical

Method for Fall Detection,” in Proc. 21st Int. Conf. Pattern Recognit. (ICPR), 2012.

Individual Contribution
➢​MICHAEL SENKAO (21053295)
Contribution in Project:
●​ Researched, implemented, and evaluated the One-Class Support Vector Machine
(OC-SVM) approach for detecting falls using smartphone sensor data.
●​ Developed a complete workflow for OC-SVM including preprocessing, feature
extraction, model training, evaluation, and visualization using Python.
●​ Performed hyperparameter tuning (on nu and gamma) and evaluated model
performance using confusion matrices and classification reports.
●​ Shared findings and helped the team compare OC-SVM performance with other
fall detection models (IQR, Z-Score, etc.).
Contribution in Review Paper:
●​ Provided a summary of the OC-SVM implementation results to support and
reinforce the theoretical explanation of SVM-based anomaly detection in the
review paper.
Contribution in Report:
●​ Wrote and formatted Section 3.4: One-Class SVM Based Detection, covering
rationale, feature engineering, model design, evaluation, visualizations, and
findings.
●​ Provided content for the OC-SVM workflow including diagrams, plots,
classification metrics, and explanatory text.
 26

●​ Ensured that OC-SVM implementation details followed the same technical and
structural standards as the other methods in the report.
Contribution in PPT:
●​ Provided the content for slides explaining the OC-SVM methodology, feature
extraction process, and model evaluation results.
●​ Assisted in the alignment of presentation content with the report to maintain
consistency across project components.

➢​KHUSHI KUMARI (21053363)

Contribution in Project:
●​ Identified and proposed a suitable dataset (UCI HAR) for the team, and
implemented the dual-stream CNN with self-attention model and assisted
Divyaranjan in implementation of LSTM-GMM model.
●​ Initiated the project idea, structured task allocation, and facilitated collaboration
by coordinating efforts and holding knowledge-sharing sessions.
●​ Led core research, organized team meetings, and developed templates/patterns
to streamline workflow and boost participation.
Contribution in Review Paper:
●​ Initiated and finalised the topic and overall content structure.
●​ Took lead on research and drafting a majority of the material.
Contribution in Report:
●​ Collaborated with Michael Chilshe to format the report properly, and
contributed by integrating tables and diagrams.
●​ Specifically contributed to sections 2.2, 3.5, 3.6, and 4.1.
Contribution in PPT:
●​ Gathered results from all models/methods and compiled LaTeX for presentation.

➢​DIVYARANJAN SAHOO (21053386)

Contribution in Project:
●​ Designed and implemented the LSTM-GMM hybrid model combining
sequential learning and probabilistic clustering for fall detection.
●​ Preprocessed the UCI HAR dataset, fine-tuned hyperparameters for LSTM, and
optimized model performance to achieve 93.2% accuracy.
●​ Collaborated on data visualization (confusion matrix, F1-scores) and
fall-specific metric analysis.
Contribution in Review Paper:
●​ Collected and analyzed literature on anomaly detection architectures including
GMM, hybrid deep learning, and LSTM-based models.
●​ Drafted sections related to sequential models and statistical evaluation in
healthcare anomaly detection.
Contribution in Report:
●​ Wrote and refined the Section 3.6 - LSTM GMM, and assisted with results and
discussion for hybrid models in Section 4.
●​ Coordinated inclusion of implementation details and performance breakdown
for LSTM-GMM.
 27

Contribution in PPT:
●​ Created slides for LSTM-GMM model architecture, methodology, and key
results.
●​ Helped streamline the final presentation deck and presented the fall detection
demo using hybrid models.

➢​MICHAEL CHILSHE (21053409)

Contribution in Project:
●​ Helped set up the environment for the project codebase in colab and
Implemented IQR statistical method.
●​ Collaborated with Meghanjali to implement the Z-score statistical model.
Contribution in Review Paper:
●​ Reviewed 4 research papers in the field and provided Khushi with information
for the draft.
●​ Collaborated with Khushi to format and edit the draft into a more finished paper.
Contribution in Report:
●​ Provided the project specific chapterwise template to follow for the report
●​ Specifically contributed to sections 1, 3.1 and 6.
Contribution in PPT:
●​ Edited and expanded the IQR section
●​ Helped finalize the structure of the PPT in latex

➢​MEGHANJALI SAHA (21053425)

Contribution in Project:
●​ Implemented the Z-score statistical method and collaborated with Michael
during its Development
●​ Supported the setup of the project environment in Google colab.
Contribution in Review Paper:
●​ Reviewed 4 research papers to gather relevant information for the report and
statistical model implementation (such as the Z-score method).
Contribution in Report:
●​ Implemented the Z-score based anomaly detection: Developed a method to
detect anomalies using Z-score thresholds.
●​ Designed Multi-Sensor Consensus Classification: Classifies falls, trips and
normal activities based on sensor anomalies.
Contribution in PPT:
●​ Developed Z-score methodology for axis-wise anomaly detection with ±3σ
thresholding and sliding window.
●​ Defined Z-score parameters (sampling rate: 50Hz, window size: 128 samples)
for real-time fall detection.
 FALL DETECTION -
IMPLEMENTATION AND
LITERATURE REVIEW
by Khushi .

Submission date: 11-Apr-2025 10:38AM (UTC+0530)

Submission ID: 2642260449
File name: Report_10.pdf (1.43M)
Word count: 5648
Character count: 33420
FALL DETECTION - IMPLEMENTATION AND LITERATURE
REVIEW
ORIGINALITY REPORT

13 %
SIMILARITY INDEX
10%
INTERNET SOURCES
9%
PUBLICATIONS
6%
STUDENT PAPERS

PRIMARY SOURCES

1
[Link]
Internet Source 1%
2
Submitted to National Institute of Technology
Warangal
1%
Student Paper

3
H L Gururaj, Francesco Flammini, V Ravi
Kumar, N S Prema. "Recent Trends in
1%
Healthcare Innovation", CRC Press, 2025
Publication

4
[Link]
Internet Source 1%
5
[Link]
Internet Source 1%
6
Ziad Salem, Felix Lichtenegger, Andreas Peter
Weiss, Claude Leiner, Christian Sommer,
<1%
Christian Krutzler. "Comparison of Machine
Learning Algorithms for Position-Oriented
Human Fall Detection", 2023 International
Wireless Communications and Mobile
Computing (IWCMC), 2023
Publication

7
[Link]
Internet Source <1%
8
Mehdi Ghayoumi. "Generative Adversarial
Networks in Practice", CRC Press, 2023
<1%
Publication

9
[Link]
Internet Source <1%
10
[Link]
Internet Source <1%
11
[Link]
Internet Source <1%
12
[Link]
Internet Source <1%
13
[Link]
Internet Source <1%
14
[Link]
Internet Source <1%
15
[Link]
Internet Source <1%
16
[Link]
Internet Source <1%
17
Lecture Notes in Computer Science, 2015.
Publication <1%
18
Meichuan Huang, Xiaoxuan Wang, Peng Sun,
Sheng Wang, Zhi Wang. "Wavelet Package
<1%
Analysis based Fall Detection and Diagnosis",
2018 37th Chinese Control Conference (CCC),
2018
Publication

19
[Link]
Internet Source <1%
20
Saeedeh Zebhi. "An efficient 3D convolutional
neural network with informative 3D volumes
<1%
for human activity recognition using wearable
sensors", Multimedia Tools and Applications,
2023
Publication

21
Submitted to University of Hull
Student Paper <1%
[Link]
 <1%
Internet Source
22

23
[Link]
Internet Source <1%
24
Pham Van Thanh, Duc-Tan Tran, Dinh-Chinh
Nguyen, Nguyen Duc Anh, Dang Nhu Dinh, S.
<1%
El-Rabaie, Kumbesan Sandrasegaran.
"Development of a Real-Time, Simple and
High-Accuracy Fall Detection System for
Elderly Using 3-DOF Accelerometers", Arabian
Journal for Science and Engineering, 2018
Publication

25
[Link]
Internet Source <1%
26
[Link]
Internet Source <1%
27
[Link]
Internet Source <1%
28
[Link]
Internet Source <1%
29
[Link]
Internet Source <1%
30
Submitted to Texas A&M University - Corpus
Christi
<1%
Student Paper

31
Submitted to The Hong Kong Institute of
Education
<1%
Student Paper

32
[Link]
Internet Source <1%
33
[Link]
Internet Source <1%
34
[Link]
Internet Source <1%
35
Amani Alfakih, Zhengwang Xia, Bahzar Ali,
Saqib Mamoon, Jianfeng Lu. "Deep Causality
<1%
Variational Autoencoder Network for
Identifying the Potential Biomarkers of Brain
Disorders", IEEE Transactions on Neural
Systems and Rehabilitation Engineering, 2024
Publication

36
[Link]
Internet Source <1%
37
Mohammed A.A. Al-qaness, Abdelghani
Dahou, Mohamed Abd Elaziz, Ahmed M.
<1%
Helmi. "Human activity recognition and fall
detection using convolutional neural network
and transformer-based architecture",
Biomedical Signal Processing and Control,
2024
Publication

Exclude quotes Off Exclude matches < 10 words

Exclude bibliography Off