AI-Driven Multimodal Posture and Action Analysis for Detecting

Workplace Fatigue and Productivity

L. Anand, Ansh Nilesh Doshi, Sanskriti Kumari and Sunny Chaudhary

School of Computing, SRM Institute of Science and Technology Kattankulathur, Tamil Nadu, India

Keywords: Fatigue Detection; Multimodal; Posture Recognition; Deep Learning; Classification; Workplace Accidents,

EAR; Drowsiness; Head Count; Screen Activity Detection.

Abstract: We all know how easy it is to lose focus when we’re tired, in many workplaces that can lead to mistakes or

even accidents. To tackle this, we’ve built a smart, real-time system that helps spot early signs of fatigue

before they become a problem. Using AI and computer vision, the system watches for subtle cues like frequent

blinking, frequent head downs and shifted focus, common signals that someone might be getting drowsy

Additionally, a screen activity detection feature ensures that users remain engaged in their work by monitoring

active applications, mouse interactions, and screen content.It runs discreetly in the background with only a

cheap webcam and open-source software, providing live feedback and cheery nudges when it’s time to take

a break or get back on track. Our system is lightweight, easy to use, and is privacy respectful because it is

always about movement patterns, not personal data. By fusing together physical and behavioural insights, this

project is looking to be able to provide a safer, more productive environment that allows people to be at their

best and healthy at the same time.

1 INTRODUCTION

Fatigue and drowsiness pose a set of common issues

that can quietly affect everything from productivity to

safety in many industries from busy offices and

factory floors, to healthcare settings and human

workstations. Traditional supervision or self-

reporting simply isn’t enough to catch these issues

before they become serious.

So, in this project we have our smart AI system

that monitors fatigue signs in real time. Marrying a

YOLO-based object detection model with Media

Pipe’s facial tracking allows the system to identify

when someone’s excessive blinking, falling asleep, or

gazing at something other than their work for too

long. It monitors simple but telling behaviours such

as eye movement, head position and blink patterns,

and sends up clear alerts when someone might need

to refocus or take a break.

We have also included a screen activity detection

feature to further refine focus tracking. Means you

will incredibly screen and test what the user is doing

Applications, activities and mouse clicks with screen

content analysis, so that you know the person is

actually click whether it is Play or application Active.

If it detects you have been inactive for too long or

spent too much time on something unrelated to work

the system goes, "hey, remember to get back on

track."

What set this solution apart is that it’s lightweight,

inexpensive and compatible with standard webcams

and open-source software such as OpenCV and Py

Torch. It also employs adaptive feedback loops,

which modify reminders according to observed user

behaviour patterns, avoiding superfluous

interruptions. It’s designed to scale and to respect

privacy, concerned with patterns as opposed to

personal data. Our goal is simple: To create a tool that

helps you stay awake & be safe; All while supporting

well-behaved productivity in the workplace.

2 RELATED WORK

Fatigue detection has been explored in various

domains, including transportation (R. Sahayadhas,

2012), (Z. Li and J. Ren, 2022), (R. Yuan and H.

Long, 2024), healthcare (G. Liu et., al. 2023), (L.

Yuzhong et., al. 2020), and occupational safety (M.

Moshawrab et., al. 2022). Prior studies primarily

370

Anand, L., Doshi, A. N., Kumari, S. and Chaudhary, S.

AI-Driven Multimodal Posture and Action Analysis for Detecting Workplace Fatigue and Productivity.

DOI: 10.5220/0013898300004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

370-379

ISBN: 978-989-758-777-1

focus on vision-based methods (M. S. Devi and P. R.

Bajaj, 2008) (e.g., RGB cameras) and wearable

devices (J. Lu and C. Qi,2021), (D. Mistry et., al.

2023), (R. J. Wood et., al. 2012), (K. Madushani et.,

al. 2021) (e.g., accelerometers, gyroscopes). While

wearable sensors provide accurate data, they require

user compliance, making them less practical for

continuous monitoring. Vision-based systems offer a

non-intrusive alternative but pose privacy concerns.

Wearable sensors, such as accelerometers and

EMG, track physiological changes like heart rate

variability (HRV) for fatigue detection (B.-L. Lee et.,

al. 2015). While effective, they require user

compliance and can be uncomfortable for prolonged

monitoring. Computer vision techniques analyze eye-

tracking and facial expressions (e.g., blinking,

yawning) to infer fatigue. Deep learning models like

CNNs and LSTMs (L. Lou and T. Yue, 2023)

enhance accuracy, but privacy concerns limit their

adoption in workplace settings. Recent studies

combine multiple data sources to improve fatigue

detection. Feature-level fusion integrates signals from

different modalities, while decision-level fusion

combines classifier outputs. Attention mechanisms

further enhance interpretability and robustness. Deep

learning models outperform traditional classifiers in

recognizing fatigue patterns. CNNs are widely used

for image-based analysis, while YOLO enables real-

time posture detection. Hybrid models integrating

deep learning with classifiers like Random Forest

further boost accuracy.

Besides just spotting physical fatigue, researchers

are also finding ways to understand mental

exhaustion by looking at how people use their

screens. Simple things like which apps someone is

using, how often they type, or how they move their

mouse can reveal whether they’re focused or getting

distracted.

New AI tools can even recognize what’s on the

screen, making it easier to tell if someone is working

or drifting into non-work activities. By blending

screen activity tracking with traditional fatigue

detection, we get a more complete view of how

people stay engaged. This helps create a healthier,

more productive work environment without being

intrusive.

3 RESEARCH GAP AND

CONTRIBUTION

Existing fatigue detection approaches are either

expression (J. Jiménez-Pinto and M. Torres-Torriti,

2015), (S. Park et., al. 2019), (S. Hussain et., al. 2019)

based or posture (S. Park et., al. 2019), (J. Lu and C.

Qi,2021) based, not both. That can make them less

effective because people express tiredness

differently. Plus, many of the so-called traditional

methods (M. S. Devi and P. R. Bajaj, 2008) depend

on human-based feature selection, which are not

always environment-agnostic. Studies indicate that

masking fatigue through multiple sources of input—

from facial signals to body position or work habits—

can increase accuracy by as much as 20%, compared

to using a single form of detection.

To solve this problem, we built an AI engine that

uses RCNNs to analyse faces, YOLO for posture

tracking and Random Forest for decision making.

Research shows that models like Random Forest are

immensely successful in a fatiguing context as they

help identify patterns on higher-level, by analysing

higher quantities of real-world data. It works in real

time, with lightweight, open source tools and is

inexpensive and easy to implement in any workplace.

We’ve also built in screen activity, to track how

engaged a person is with their work. We flag signs of

mental fatigue or distraction by analysing which apps

people use, recognizing on-screen content and

detecting long periods of inactivity. Research also

suggests that app-switching regularly, or inactivity in

digital spaces, can be a sign of cognitive load. As a

result, the system leverages a blend of both physical

and digital fatigue measures in a comprehensive,

privacy-compliant approach that encourages

improved focus and output.

4 METHODOLOGY AND

SYSTEM ARCHITECTURE

In the world of work, fatigue detection methods play

an important role. Real-time monitoring of fatigue

can play a significant role in preventing sleepiness

related accidents or in increasing productivity during

long hours of work.

This study suggests an AI-enabled fatigue detection

solution that integrates:

• Drowsiness classification

based on deep

learning (YOLOv11)

• Video-based facial landmark tracking (via

Media Pipe)

• Real-time monitoring (OpenCV)

• Environment detection

AI-Driven Multimodal Posture and Action Analysis for Detecting Workplace Fatigue and Productivity

371

Screen Activity Detection for Engagement

Analysis. By analysing eye blinks, head position, and

facial expressions, the system can effectively detect

signs of fatigue and provide immediate alerts.

4.1 System Overview

Our system follows a step-by-step process, from

capturing video data to triggering alerts when fatigue

is detected.

4.1.1 Capturing and Processing Video Data

The system continuously captures video from a

webcam or external camera. Each frame is then

processed in real time to extract facial landmarks and

classify the person’s state (awake or drowsy).

4.1.2 Preparing the Data for Analysis

Before analyzing the video frames, some

preprocessing is done to make sure the system runs

efficiently:

• Converting frames to grayscale and

adding noice: This reduces computational

load while preserving important facial

details and increasing robustness [Figure 1]

attached below.

Figure 1: Adding noise (salt and pepper).

• Detecting the face: The system uses

YOLOv11 to locate the face and extract a

bounding box.

• Environment Detection: The system uses

object detection models to recognize

environmental factors such as backpacks,

books, cups, and the presence of people

around the user [Figure 2]. This helps

analyze the working environment and

possible distractions.

Figure 2: Detection of the environment.

• Tracking facial landmarks:

Figure 3: Face mesh imposed on a person in real time.

MediaPipe FaceMesh detects 468 key facial

points as shown in [Figure 3], focusing on

the eyes, mouth, and head position.

• Tracking Screen Engagement: The system

monitors cursor movement, typing activity,

and screen interaction duration to assess

engagement.

This helps ensure that only the relevant features are

used for further analysis.

4.2 Detecting Fatigue

To determine if a person is fatigued, the system looks

at three key factors:

4.2.1 Eye Blink Rate (EBR) and Eye Aspect

Ratio (EAR)

When people get drowsy, they tend to blink more

slowly or keep their eyes closed longer than usual.

The system calculates Eye Aspect Ratio (EAR) using

six key points around the eye:

𝐸𝐴𝑅 =







|| ||

|| ||

(1)

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

372

• P1 and P4 represent the horizontal eye

corners.

• P2, P3, P5, and P6 are vertical landmarks.

Figure 4: Eye Aspect Ratio.

How It Detects Fatigue:

• If the EAR falls below 0.3 for several

consecutive frames, the system flags

drowsiness.

• It also tracks blink duration and frequency to

check for patterns of fatigue over time.

4.2.2 Head Position Tracking

A common sign of fatigue is nodding off or looking

away from the screen for extended periods.

How It Works:

• The system monitors the Y-coordinate of the

face.

• If the head moves below a certain threshold

or disappears for more than five seconds, an

alert is triggered.

4.2.3 Deep Learning-Based Drowsiness

Detection (YOLOv11)

Instead of relying only on rule-based features like

EAR, the system also uses a trained deep learning

model (YOLOv11) to recognize drowsy facial

expressions.

Steps in Detection:

1. YOLOv11 processes each video frame,

identifying facial features.

2. The model then classifies the frame as

“Awake” or “Drowsy”.

3. If the person is continuously classified as

drowsy, an alert is triggered.

This method improves accuracy by recognizing

subtle signs of fatigue that rule-based approaches

might miss.

4.2.4 Screen Activity Detection

In addition to facial cues, the system monitors screen

engagement patterns to detect fatigue or

disengagement.

How It Works:

• Cursor movement tracking: A reduction in

cursor activity may indicate a drop-in focus.

• Typing frequency analysis: Slower or

erratic typing may signal fatigue.

• Prolonged inactivity: If no interaction is

detected for a set period, the system triggers

a reminder.

4.2.5 Environment Detection

To improve contextual awareness, the system detects

objects in the user’s surroundings. Using YOLO-

based object detection, the YOLO11x summary can

be seen in [Figure 5]:

Figure 5: List of Class (object identified by the model)

along with Mean Precision, Mean Recall, F1 Score, Mean

AP50 and Mean AP also the number if images used.

This additional layer of analysis enhances

accuracy by incorporating both physiological and

behavioral fatigue indicators.

4.3 Training the Model

For the deep learning model to detect drowsiness

accurately, it needs to be trained on a dataset of awake

and drowsy faces.

4.3.1 Dataset and Augmentation

The model is trained using thousands of labelled

images showing people in both alert and drowsy

states. Prerecorded videos were given to train the

multimodal, each divided into different frames, the

data set also includes images from different sources

(Daisee and Roboflow) and the distribution can be

seen in Table 1.

• Screen activity data is collected from real-

world usage scenarios.

AI-Driven Multimodal Posture and Action Analysis for Detecting Workplace Fatigue and Productivity

373

• To improve accuracy, data augmentation is

used to simulate different lighting conditions

and angles as shown in [Figure 6] below.

Table 1: Distribution of Dataset.

Dataset

Daisee

Datase

Roboflow

Datase

Trainin

3800 1400

Validation 1490 200

Tes

1710 400

Total 7000 2000

Figure 6: Few samples of the dataset.

4.3.2 Dataset Analysis for Environment

Object Detection

To make sure our Fatigue Detection System works

well in real life, we need to understand what objects

appear around people. The label distribution chart

[Figure 7a] shows how often things like backpacks,

phones, cups, and laptops show up in the video. This

helps us train the model with different environments,

so it can work well anywhere.

We also used a label correlogram [Figure 7b] to

see which objects are often found together. For

example, people and phones usually appear in the

same scene because people use their phones a lot.

Similarly, laptops and coffee cups often show up

together in work settings. This helps the model learn

not just about faces, but also about the surroundings

where fatigue happens.

Figure 7a: Label distribution chart, Bounding (Top Left

Chart) Box Distribution (Top Right Plot), Object Position

& Size Distribution (Bottom Plots).

Figure 7b: Label correlogram.

4.3.3 Training Process

• The training is done using PyTorch and

YOLOv11, with a focus on optimizing the

model for real-time performance.

4.4 Real-Time Alert System

When the system detects fatigue, it immediately alerts

the user to take a break.

4.4.1 Drowsiness Alerts

• If the system detects eye closure or

drowsiness for more than three seconds, a

popup notification appears.

• In workplace settings, these alerts can be

integrated with a fatigue management

dashboard.

4.4.2 Head-Down Alerts

• If the system cannot detect the user’s face for

more than five seconds, an alert is triggered.

• This helps ensure that users remain engaged,

especially in work or learning environments.

4.4.3 Screen Inactivity Alerts

• If no screen activity (cursor movement,

typing, or scrolling) is detected for a

prolonged period, the system sends an

engagement alert which can be seen in

[Figure 8].

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

374

Figure 8: Inactive alert message.

• This ensures that users remain attentive,

particularly in educational and professional

settings.

4.5 Optimizing the System for Real-

World Use

To ensure the system runs efficiently, several

optimizations are applied:

4.5.1 Performance Optimization

• CUDA acceleration speeds up deep learning

inference on compatible GPUs.

• The model is optimized to run smoothly on

desktop computers, embedded devices, and

cloud systems.

• All specifications and requirements of the

model can be seen in [Figure 9].

4.5.2 Model Compression

• The deep learning model is converted to

ONNX format, reducing its size without

losing accuracy.

4.6 Evaluating the System

To measure how well the system performs, it is tested

under different conditions.

4.6.1 Performance Metrics

• Accuracy – Measures how correctly the

system detects fatigue.

• Latency – Ensures real-time performance

without lag.

• False Positive Rate – Minimizes incorrect

fatigue detections.

The fatigue detection system follows a structured

workflow to ensure real-time and accurate monitoring

as shown in [FIG 10]. It starts by capturing video

frames, shutting down if none are detected to save

resources. When a frame is acquired, it undergoes

preprocessing, including resizing and grayscale

conversion, to enhance efficiency. The system then

extracts key behavioural features: posture analysis to

detect slouching, eye tracking to monitor blinking and

eye closure, and behavioural cues like head nodding

or sudden movements linked to fatigue. These

features are fed into a pre-trained model, which

analyses patterns and predicts fatigue levels. If

fatigue is detected, the system triggers an alert to

prevent potential risks. By combining multiple

behavioural indicators, this system ensures accurate,

real-time fatigue assessment, making it highly useful

in workplace safety.

Figure 9: Specification of the working model.

AI-Driven Multimodal Posture and Action Analysis for Detecting Workplace Fatigue and Productivity

375

Figure 10: Fatigue Detection System Flowchart

5 RESULTS AND DISCUSSION

To understand how well our Real-Time Fatigue

Detection System performs, we evaluated its

accuracy using different metrics like precision, recall,

and F1-score. We tested the model using a dataset

containing videos of individuals in both awake and

drowsy states, under various real-world conditions

like different lighting and head positions.

5.1 Understanding Model Performance

with a Confusion Matrix

A confusion matrix helps visualize how often the

system correctly detects fatigue and where it makes

mistakes. Here’s what it looks like [Table 2].

Table 2: Confusion Matrix in tabular form.

Actual

State

Predicted Awake

Predicted

Drowsy

Awake Correct (TP)

Misclassified

(

)

Drowsy

Misclassified

(FP)

Correct (TN)

• True Positives (TP) → Correctly detected

drowsy instances.

• True Negatives (TN) → Correctly detected

awake instances.

• False Positives (FP) → Mistakenly flagged

awake individuals as drowsy.

• False Negatives (FN) → Failed to detect

drowsiness.

Our model had very few false negatives, meaning it

rarely missed signs of fatigue [Figure 11a] a crucial

aspect for safety applications.

To get a clearer picture of how well the model

performs, we also use a normalized confusion matrix

[Figure 11b]. Instead of just counting correct and

incorrect predictions, this version shows the results as

percentages, making it easier to compare accuracy

across different categories. This is especially useful

when there are more awake cases than drowsy ones in

the dataset. Ideally, we want high numbers along the

diagonal, which means the model is making mostly

correct predictions, and very low numbers elsewhere,

indicating fewer mistakes.

Figure 11a: Confusion matrix of the proposed model.

Figure 11b: Confusion matrix normalized.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

376

5.2 Precision, Recall, and F1-Score

To measure the system’s reliability, we calculated

three important metrics:

• Precision → How many of the detected

drowsy cases were actually drowsy?

Formula used:

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =





(2)

• Recall → How many of the actual drowsy

cases did the model catch?

Formula Used:

𝑅𝑒𝑐𝑎𝑙𝑙 =





(3)

• F1-score → A balance between precision

and recall to ensure overall reliability.

Formula Used:

𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2 ∗

∗ 



(4)

Our model achieved an F1-score of 92.5%,

meaning it performs well in detecting fatigue without

making too many mistakes.

5.3 Confidence Curves and Model

Stability

We also tested how the model behaves when

adjusting its confidence level:

• Precision-Recall Curve → Showed that the

model maintains high accuracy even when

handling uncertain cases.

The Precision-Recall Curve of our model

can be seen in [Figure 12a], It shows 0.853

mAP @ 0.5 when all classes are taken.

Here mPA => Mean Average Precision and

@0.5 refers to the IoU (Intersection over

Union) threshold of 0.5.

• F1 Confidence Curve → Proved the system

remains stable across different confidence

thresholds, which for our model can be seen

in [Figure 12b] with a value 0.69 at 0.426

when all classes are considered.

• Recall Confidence Curve → Confirmed the

model rarely misses signs of drowsiness,

which is essential for real-world

applications. Our model gives out a valur of

0.91 at 0.000 as can be seen in [Figure 12c]

when all classes are considered.

• Precision Confidence Curve: Indicates that

even at higher confidence levels, the model

maintains high precision, ensuring that

detected fatigue cases are truly drowsy

individuals. The value being 1.00 at 1.000

for our model under the case when all cases

are taken [Figure 12d].

These findings indicate that our system is reliable

and practical for real-time fatigue monitoring.

Figure 12: Confidence curves.

5.4 Output

The experimental result when fatigue detection model

is executed [Figure 13] shows various results as alert,

blink count, head down count and if the head is down

then for how many seconds.

Figure 13: result showing different alerts 1. You are awake,

2. Your head is down, please raise it , 3., 4. You are drowsy,

please stay alert.

AI-Driven Multimodal Posture and Action Analysis for Detecting Workplace Fatigue and Productivity

377

When any person is not detected in the cabin or work

place it detects various objects in the surrounding as

seen in [Figure 14] this also includes people who are

not facing the camera or workig on other system as

seen in [Figure 2] .

Figure 14: object detected by AI model (mouse, laptop,

bottle, cup, chair, book, potted plant etc.

6 CONCLUSIONS

In this project, we built a Real-Time Fatigue

Detection System that monitors the persons

performance to increase productivity that uses AI to

track facial expressions, body posture, screen activity,

and objects around a person to detect drowsiness and

all over activity. This makes it more accurate and

reliable than traditional methods. It also helps unlock

if the person is actually working or not.

The training loss and validation loss values in Table

3 show how our model's performance evolved across

different epochs (1, 10, 20, 30, 40, 50).

Table 3: Model Losses on training and validation data.

Epoch Training Loss Validation Loss

Box cls dfl Box cls dfl

1 0.645 1.603 1.036 0.675 1.334 1.008

10 0.474 0.641 0.939 0.374 0.597 0.879

20 0.390 0.520 0.896 0.342 0.567 0.882

30 0.346 0.449 0.878 0.318 0.542 0.868

40 0.298 0.365 0.859 0.300 0.498 0.853

50 0.231 0.272 0.822 0.305 0.476 0.865

Below [figure 15] shows graphs that show

significant decrease in training loss and validation

loss indicating that the model made less mistakes as it

proceeds from epoch 1 to 50 over time.

Figure 15: Graphs generated with dataset is tested (model

losses on training and validation data).

Our model showed good performance, with

training loss going down over time and key accuracy

measures like 77.39% m AP and 73.56% recall

proving that it works well in different situations. These

results confirm that our model successfully learned

from the dataset, reducing errors and achieving strong

detection performance.

This system can be useful in workplaces such as

offices to prevent accidents caused by fatigue and to

calculate the correct work hour the employee is giving

to the company. In the future, we can improve it by

making it work better in different lighting conditions,

adding more training data, and even connecting it with

ceiling camera for better monitoring.

This research is a step toward smarter and safer

fatigue detection, along with proper monitoring of

performance and work time.

REFERENCES

M. S. Devi and P. R. Bajaj, Driver fatigue detection based on

eye tracking, in Proc. 1st Int. Conf. Emerging Trends

Eng. Technol., 2008.

R. Sahayadhas, K. Sundaraj, and M. Murugappan, A review

of computer vision techniques for driver drowsiness

detection, in Proc. IEEE Int. Conf. Syst., Man, Cybern.

(SMC), 2012.

J. Jiménez-Pinto and M. Torres-Torriti, Real-time driver

drowsiness detection based on driver’s face image

behavior using a system of human-computer interaction

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

378

implemented in a smartphone, IEEE Trans. Intell.

Transp. Syst., 2015.

S. Park, J. Kim, and B. Ro, Real-time driver drowsiness

detection for embedded system using model

compression of deep neural networks, IEEE Access,

2019.

S. Hussain, S. Hussain, and S. A. Khan, Driver drowsiness

detection using behavioral measures and machine

learning techniques: A review of state-of-art techniques,

in Proc. 2nd Int. Conf. Commun., Comput. Digit. Syst.

(C-CODE), 2019.

L. Yuzhong, L. Zhiqiang, Y. Zhixin, L. Hualiang, and S.

Yali, Research progress of work fatigue detection

technology, in Proc. 2nd Int. Conf. Civil Aviation Safety

Inf. Technol. (ICCASIT), 2020.

S. Gupta, A. Sharma, and R. Kumar, Driver drowsiness

detection system using YOLO and eye aspect ratio, in

Proc. IEEE Int. Conf. Image Process. (ICIP), 2020.

J. Lu and C. Qi, Fatigue detection technology for online

learning, in Proc. Int. Conf. Networking Netw. Appl.

(NaNA), 2021.

Z. Li and J. Ren, Driver fatigue detection algorithm based on

video image, in Proc. 14th Int. Conf. Measuring

Technol. Mechatronics Autom. (ICMTMA), 2022.

L. Lou and T. Yue, Fatigue driving detection based on facial

features, in Proc. 8th Int. Conf. Intell. Comput. Signal

Process. (ICSP), 2023.

G. Liu, J. Liu, and L. Dai, Research on the fatigue detection

method of operators in digital main control rooms of

nuclear power plants based on multi-feature fusion, in

Proc. IEEE 2nd Int. Conf. Electr. Eng., Big Data

Algorithms (EEBDA), 2023.

R. Yuan and H. Long, Driver fatigue detection based on

multi-feature fusion facial features, in Proc. 5th Int.

Conf. Big Data Artif. Intell. Softw. Eng. (ICBASE),

2024.

X. Qiu, F. Tian, Q. Shi, Q. Zhao, and B. Hu, Designing and

application of wearable fatigue detection system based

on multimodal physiological signals, in Proc. IEEE Int.

Conf. Bioinf. Biomed. (BIBM), 2020.

B.-L. Lee, D.-S. Lee, and B.-G. Lee, Mobile-based

wearable-type of driver fatigue detection by GSR and

EMG, in Proc. IEEE Region 10 Conf. (TENCON), 2015.

Y. Song, H. Yin, X. Wang, C. Sun, Y. Ma, Y. Zhang, X.

Zhong, Z. Zhao, and M. Zhang, Design of a wireless

distributed real-time muscle fatigue detection system, in

Proc. IEEE Biomed. Circuits Syst. Conf. (BioCAS),

2022.

T. Kobayshi, S. Okada, M. Makikawa, N. Shiozawa, and M.

Kosaka, Development of wearable muscle fatigue

detection system using capacitance coupling electrodes,

in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.

(EMBC), 2017.

X.-Y. Gao, Y.-F. Zhang, W.-L. Zheng, and B.-L. Lu,

evaluating driving fatigue detection algorithms using eye

tracking glasses, in Proc. Int. IEEE/EMBS Conf. Neural

Eng. (NER), 2015.

M. Moshawrab, M. Adda, A. Bouzouane, H. Ibrahim, and A.

Raad, Detection of occupational fatigue in digital era;

parameters in use, in Proc. Int. Conf. Human-Centric

Smart Environ. Health Well-being (IHSH),

2022.

D. Mistry, M. F. Mridha, M. Safran, S. Alfarhood, A. K.

Saha, and D. Che, Privacy-preserving on-screen activity

tracking and classification in e-learning using federated

learning, IEEE Access, vol. 11, 2023.

R. J. Wood, C. A. McPherson, and J. M. Irvine, Video track

screening using syntactic activity-based methods, in

Proc. IEEE Appl. Imagery Pattern Recognit. Workshop

(AIPR), 2012.

K. Madushani, N. Edirisinghe, K. Koralage, R. Sahabandu,

P. Samarasinghe, and L. Wickremesinghe, Screening

child’s screen time, in Proc. IEEE Int. Conf. Ind. Inf.

Syst. (ICIIS), 2021, also.

AI-Driven Multimodal Posture and Action Analysis for Detecting Workplace Fatigue and Productivity

379