Fall Detection Methods in Signal Processing

Falls can be detected through various signal processing methods including accelerometer analysis, Fourier
analysis, machine learning algorithms, and environmental sensor systems.

Abstract
Signal processing for fall detection builds on data from wearable and environmental sensors. Studies re-
port several techniques that transform raw sensor signals into fall events. For example, threshold‐based
algorithms—applied to waist‐mounted tri-axial accelerometer data—detect rapid changes in velocity, im-
pact, and posture; one study using this method noted 100% sensitivity and specificity among 20 volunteers.
Other work computes a signal's magnitude area to flag falls, while frequency-domain procedures, such as
Fourier descriptor analysis of accelerometer and gyroscope data, discern fall‐like patterns. In addition, ma-
chine learning methods play a prominent role. Two studies implemented Hidden Markov Models and two
applied Support Vector Machines to classify signal patterns; k-Nearest Neighbors was used in two investi-
gations, one of which reported 100% sensitivity. Decision tree classifiers, sometimes combined with feature
selection techniques for reduced complexity or improved power efficiency (with one system saving over 75%
power), also form part of the repertoire. Finally, systems based on environmental sensors—using depth
cameras or radar—and strategies such as non-linear classification with Kalman filtering extend fall detection
beyond wearable devices. In sum, available methods range from simple, threshold-based approaches to more
sophisticated models that exploit time–frequency analysis and machine learning, with techniques selected
according to requirements for computational cost, sensor configuration, and real-world operation.

Paper search
Using your research question ”what are available methods in signal processing to detect falling person?”, we
searched across over 126 million academic papers from the Semantic Scholar corpus. We retrieved the 50
papers most relevant to the query.

Screening
We screened in papers that met these criteria:
• Signal Processing Focus: Does the study focus on fall detection using clearly described signal
processing methods?
• Human Data Collection: Does the study involve human subjects or human motion data collected
through sensors (e.g., accelerometers, gyroscopes, cameras)?
• Algorithm Implementation: Does the study present an original signal processing algorithm with
experimental validation?
• Study Type: Is the study either (a) primary research with practical implementation and validation,
or (b) a systematic review/meta-analysis of fall detection methods?
• Detection Focus: Does the study include fall detection components (not exclusively fall prevention
or medical consequences)?
• Target Application: Does the study focus specifically on human falls (not non-human or inanimate
object detection)?
• Practical Validation: Does the study include practical implementation and validation of the proposed
methods?

1
We considered all screening questions together and made a holistic judgement about whether to screen in
each paper.

Data extraction
We asked a large language model to extract each data column below from each paper. We gave the model
the extraction instructions shown below for each column.
• Signal Processing Method for Fall Detection:
Describe the specific signal processing technique used for fall detection. Include:
• Type of signal processing method (e.g., feature selection, decision trees, signal magnitude area)
• Specific algorithms or computational approaches
• Any unique innovations in the signal processing approach
If multiple methods are described, list all. If method details are incomplete, note ”insufficient details pro-
vided”.
• Sensor Technology Used:
Specify:
• Type of sensor(s) used (e.g., tri-axial accelerometer, Microsoft Kinect)
• Sensor placement on body or in environment
• Number of sensors used
• Technical specifications of sensors (if provided)
If sensor details are not fully described, note ”partial sensor information available”.
• Participant Characteristics:
Extract:
• Age range of participants
• Number of participants
• Demographic details (e.g., elderly, young volunteers)
• Any specific inclusion/exclusion criteria
If participant details are incomplete, note ”insufficient participant information”.
• Study Setting:
Describe:
• Location of study (e.g., homes, laboratory)
• Context of fall detection (e.g., simulated falls, real-world environment)
• Number of study sites
If setting is not clearly described, note ”setting details unclear”.
• Fall Detection Performance Metrics:
Extract quantitative performance indicators:
• Accuracy rate

2
 • Error rate
• Sensitivity
• Specificity
• Power efficiency (if applicable)
Record exact numerical values. If multiple metrics are reported, list all. Use percentages or decimal values
as in original text.
• Validation Approach:
Describe:
• Validation method (e.g., cross-validation)
• Number of falls used in validation
• Types of falls (simulated, real-world)
• Comparison with other fall detection methods (if applicable)
If validation details are incomplete, note ”partial validation information”.

Results
Characteristics of Included Studies
Sensor Detection Full text
Study Study Design Type/Placement Method Sample Size retrieved
Bourke et al., Evaluation of Tri-axial Threshold- 20 volunteers No
2010 fall detection accelerometer / based (10 young, 10
algorithms Waist-mounted algorithm using elderly)
velocity,
impact, and
posture
Chen et al., Fall detection Tri-axial Signal No mention No
2010 system accelerometer / magnitude area found
evaluation Waist-mounted method
Lee, 2019 Fall detection Three-axis Fourier No mention No
system accelerometer descriptor- found
evaluation and three-axis based
gyroscope / frequency
Mobile phone analysis
Stone and Fall detection Microsoft Two-stage fall 13 apartments, No
Skubic, 2015 system Kinect / detection 454 falls (445
evaluation Environment system using simulated, 9
depth image real)
frames and
decision trees

3
 Sensor Detection Full text
Study Study Design Type/Placement Method Sample Size retrieved
Sucerquia et Real-life/real- Triaxial Non-linear 38 participants Yes
al., 2018 time fall accelerometer / classification (SisFall
detection Waist feature with dataset) + 3
evaluation Kalman filter elderly for
validation
Tong et al., Fall detection Tri-axial Hidden Markov No mention No
2013 and prediction accelerometer / Model (HMM) found
method Upper trunk based method
evaluation
Wang et al., Power-efficient Wearable fall Feature No mention No
2017 feature detector selection and found
selection for (sensor type binary
fall detection not specified) classification
decision tree
Wu et al., 2013 Fall detection Narrowband Time-frequency No mention No
using radar pulse-Doppler analysis, found
signals radar / matching
Environment pursuit
decomposition,
HMM, and
Support Vector
Machine
(SVM)
Zhang et al., Fall detection Tri-axial 1-Class SVM 32 volunteers No
”Fall Detection system accelerometer / preprocessing (12 elders, 20
by Embedding evaluation Embedded in with KFD and younger)
an cellphone k-Nearest
Accelerometer” Neighbors
(k-NN)
classification
Özdemir and Machine Tri-axial Six machine 14 volunteers Yes
Barshan, 2014 learning accelerometer, learning
techniques for gyroscope, and techniques
fall detection magnetometer (k-NN, LSM,
/ Head, chest, SVM, BDM,
waist, right DTW, ANN)
wrist, right
thigh, right
ankle

Analysis of the included studies:

• Study Design : System evaluation was the most common design (4/10 studies), followed by various

4
 other approaches including algorithm evaluation, real-life evaluation, and machine learning evaluation.
• Sensor Type :
– Accelerometers were the most frequently used sensors (7/10 studies), often in combination with
other sensors
– Gyroscopes were used in 2 studies
– Kinect, radar, and magnetometer were each used in 1 study
– We didn't find sensor type information for 1 study
• Sensor Placement :
– Waist-mounted sensors were most common (3/10 studies)
– Mobile phone placement and environmental placement were used in 2 studies each
– Other placements included upper trunk, wearable, and multiple body parts
• Detection Method : We found a wide variety of detection methods:
– Decision trees, Hidden Markov Models (HMM), and Support Vector Machines (SVM) were each
used in 2 studies
– Other methods included threshold-based algorithms, signal magnitude area, frequency analysis,
and various machine learning techniques
• Sample Size :
– We found sample size information for 5/10 studies, totaling 87 participants across these studies
– One study reported 454 falls
– We didn't find sample size information in the available abstracts or full texts for the other 5
studies

Effects Analysis
Detection Performance

Study Algorithm Type Sensitivity Specificity Processing Time

Bourke et al., 2010 Threshold-based 100% 100% No mention found
Chen et al., 2010 Signal magnitude No mention found No mention found No mention found
area
Lee, 2019 Fourier No mention found No mention found No mention found
descriptor-based
Stone and Skubic, Two-stage with No mention found No mention found No mention found
2015 decision trees
Sucerquia et al., Non-linear No mention found No mention found No mention found
2018 classification with
Kalman filter
Tong et al., 2013 Hidden Markov No mention found No mention found No mention found
Model
Wang et al., 2017 Feature selection No mention found No mention found No mention found
with decision tree

5
Study Algorithm Type Sensitivity Specificity Processing Time
Wu et al., 2013 Time-frequency No mention found No mention found No mention found
analysis with
HMM/SVM
Zhang et al., ”Fall 1-Class SVM with No mention found No mention found No mention found
Detection by KFD and k-NN
Embedding an
Accelerometer”
Özdemir and k-NN classifier 100% No mention found No mention found
Barshan, 2014

Analysis of detection performance:

• We found 10 studies using various fall detection algorithms. The most common approaches were:
– Decision trees (2 studies)
– Hidden Markov Models (HMM) (2 studies)
– Support Vector Machines (SVM) (2 studies)
– k-Nearest Neighbors (k-NN) (2 studies)
• Other algorithms included threshold-based, signal magnitude area, Fourier descriptor-based, non-linear
classification, Kalman filter, feature selection, time-frequency analysis, and KFD.
• We found sensitivity information for only 2 out of 10 studies, both reporting 100% sensitivity.
• We found specificity information for only 1 out of 10 studies, reporting 100% specificity.
• We didn't find sensitivity information in the available abstracts or full texts for 8 studies and specificity
information for 9 studies.

Implementation Considerations

Computational Real-world
Study Method Power Efficiency Cost Applicability
Bourke et al., 2010 Threshold-based No mention found Low (simple Tested with both
thresholds) scripted and
unscripted
activities
Chen et al., 2010 Signal magnitude No mention found Low to moderate Designed for
area elderly residents
Lee, 2019 Fourier No mention found Moderate Tested with real
descriptor-based (frequency human data set
analysis)
Stone and Skubic, Two-stage with No mention found Moderate to high Tested in actual
2015 decision trees homes of older
adults

6
 Computational Real-world
Study Method Power Efficiency Cost Applicability
Sucerquia et al., Non-linear Device lasts 4-5 Moderate (Kalman Validated with
2018 classification with days on a single filter) full-day real-world
Kalman filter charge testing
Tong et al., 2013 Hidden Markov No mention found High (HMM Tested with human
Model training and fall courses
inference)
Wang et al., 2017 Feature selection 75.3% power Low (reduced Tested with
with decision tree saving feature set) simulated falls and
real free-living
trials
Wu et al., 2013 Time-frequency No mention found High (complex Tested with
analysis with signal processing) different falling
HMM/SVM scenarios
Zhang et al., ”Fall 1-Class SVM with No mention found Moderate to high Tested with
Detection by KFD and k-NN volunteers in
Embedding an experimental
Accelerometer” setting
Özdemir and k-NN classifier No mention found Moderate (k-NN Tested with
Barshan, 2014 classification) simulated falls in
laboratory setting

Analysis of implementation considerations:

• We found 10 different fall detection methods across the 10 studies. The most common methods were
decision trees, Hidden Markov Models (HMM), Support Vector Machines (SVM), and k-Nearest Neigh-
bors (k-NN), each used in 2 studies.
• Computational cost varied across studies:
– 2 studies reported low cost
– 1 study reported low to moderate cost
– 3 studies reported moderate cost
– 2 studies reported moderate to high cost
– 2 studies reported high cost
• Real-world applicability testing varied widely, with each study using a different approach. These ranged
from laboratory simulations to full-day real-world testing in actual homes.
• We found power efficiency information for only 2 out of 10 studies. One reported a device lasting 4-5
days on a single charge, while another reported 75.3% power saving.

Technical Analysis
Signal Processing Algorithms
The studies employ a diverse range of signal processing algorithms for fall detection:

7
 • Threshold-based methods : Bourke et al. (2010) used a simple yet effective approach based on thresh-
olds in velocity, impact, and posture. This method achieved high accuracy while likely maintaining
low computational complexity.
• Frequency domain analysis : Lee (2019) utilized Fourier descriptor-based frequency analysis, demon-
strating the potential of spectral analysis in fall detection.
• Machine learning approaches : Several studies employed advanced machine learning techniques:
– Hidden Markov Models (HMM): Used by Tong et al. (2013) and Wu et al. (2013)
– Support Vector Machines (SVM): Utilized by Wu et al. (2013) and Zhang et al. (2006)
– k-Nearest Neighbors (k-NN): Employed by Zhang et al. (2006) and Özdemir and Barshan (2014)
• Feature extraction and selection : Wang et al. (2017) focused on selecting power-efficient features,
while Wu et al. (2013) used time-frequency representations and matching pursuit decomposition for
feature extraction.
• Kalman filtering : Sucerquia et al. (2018) incorporated a Kalman filter to smooth data and assist in
feature extraction.
Key insights:
• Machine learning approaches trend suggests effective capture of complex fall patterns
• Success of simpler methods (e.g., threshold-based) indicates complex computations not always neces-
sary

Sensor Configuration
The studies showcase various sensor configurations:
• Sensor configuration impacts:
– System capabilities
– Practicality
• Wearable single-sensor systems: Simple and easy to use
• Multi-sensor systems: More detailed data, but increased complexity
• Environmental sensors (e.g., depth cameras, radar):
– No wearable devices needed
– Limited to fixed locations
1. Single-sensor systems : Many studies used a single tri-axial accelerometer, often waist-mounted (Bourke
et al., 2010; Chen et al., 2010).
2. Multi-sensor systems : Lee (2019) combined an accelerometer with a gyroscope in a mobile phone, while
Özdemir and Barshan (2014) used a comprehensive setup with sensors on multiple body locations.
3. Environmental sensors : Stone and Skubic (2015) used the Microsoft Kinect depth camera, while Wu
et al. (2013) employed radar technology.

Performance Optimization
Several studies focused on optimizing fall detection performance:

8
 1. Power efficiency : Wang et al. (2017) achieved a 75.3% power saving through feature selection, ad-
dressing a critical concern for wearable devices.
2. Real-time processing : Sucerquia et al. (2018) developed a system capable of real-time fall detection
with high accuracy.
3. Reducing false positives : Stone and Skubic (2015) used a two-stage system to improve accuracy, while
Sucerquia et al. (2018) incorporated a Kalman filter to reduce false positives.
4. Comprehensive validation : Several studies (Stone and Skubic, 2015; Sucerquia et al., 2018) included
real-world testing to ensure practical applicability.
These optimization efforts highlight the multifaceted challenges in developing practical fall detection systems.
Balancing accuracy, power efficiency, and real-time performance remains a key area for future research and
development.

References
A. Bourke, A. Bourke, P. V. D. Ven, M. Gamble, R. O'Connor, K. Murphy, E. Bogan, et al. “Evalua-
tion of Waist-Mounted Tri-Axial Accelerometer Based Fall-Detection Algorithms During Scripted and
Continuous Unscripted Activities.” Journal of Biomechanics, 2010.
A. Özdemir, and B. Barshan. “Detecting Falls with Wearable Sensors Using Machine Learning Techniques.”
Italian National Conference on Sensors, 2014.
Angela Sucerquia, J. López, and J. Vargas-Bonilla. “Real-Life/Real-Time Elderly Fall Detection with a
Triaxial Accelerometer.” Italian National Conference on Sensors, 2018.
Changhong Wang, S. Redmond, W. Lu, Michael C. Stevens, S. Lord, and N. Lovell. “Selecting Power-
Efficient Signal Features for a Low-Power Fall Detector.” IEEE Transactions on Biomedical Engineering,
2017.
Erik E. Stone, and M. Skubic. “Fall Detection in Homes of Older Adults Using the Microsoft Kinect.” IEEE
Journal of Biomedical and Health Informatics, 2015.
Guan-Chun Chen, Chih-Ning Huang, Chih-Yen Chiang, Chia-Juei Hsieh, and Chia-Tai Chan. “A Reliable
Fall Detection System Based on Wearable Sensor and Signal Magnitude Area for Elderly Residents.”
International Conference on Smart Homes and Health Telematics, 2010.
Jong-Ha Lee. “The Novel Computer Aided Falling Detection System Using Mobile Phone for Elderly Person.”
DEStech Transactions on Computer Science and Engineering, 2019.
Lina Tong, Quanjun Song, Y. Ge, and Ming Liu. “HMM-Based Human Fall Detection and Prediction
Method Using Tri-Axial Accelerometer.” IEEE Sensors Journal, 2013.
Meng Wu, Xiaoxiao Dai, Yimin D. Zhang, Bradley S. Davidson, M. Amin, and J. Zhang. “Fall Detection
Based on Sequential Modeling of Radar Signal Time-Frequency Features.” IEEE International Conference
on Healthcare Informatics, 2013.
Tong Zhang, Jue Wang, Ping Liu, and Jing Hou. “Fall Detection by Embedding an Accelerometer in
Cellphone and Using KFD Algorithm,” 2006.