Department Of Computer Science &

Engineering
ZEROTH REVIEW PROJECT WORK

FocusFlow: Leveraging Machine Learning to Track

Student Engagement Presented by:
Roobiga R - 821121104046
Sathya A - 821121104049
Sharmika R - 821121104051
Guided by:
Ms.T.Abitha Kujalambal

1
Abstract
This project focuses on student engagement detection using machine learning to classify
classroom behaviors. A custom dataset of 2,120 labeled images was created, categorized into
Engaged (Confused: 369, Engaged: 347, Frustrated: 360) and Not Engaged (Looking Away:
423, Bored: 358, Drowsy: 263) states. The workflow involves preprocessing images by resizing
them to 150x150 pixels and using a pre-trained MobileNet model to extract features. These
features are flattened into vectors for training. SVM with One-vs-Rest (OvR) classification is
used for model training. The dataset is divided into training and testing sets, with performance
evaluated through accuracy, precision, recall, and F1-score. This system offers real-time
insights into student engagement, enabling personalized interventions to improve
classroom management and learning outcomes.
Keywords: Student Engagement Detection, Machine Learning, Image Classification, MobileNet,
Support Vector Machine (SVM), Classroom Monitoring, Educational AI, Educational
Technology, Behavioral Classification.

2
Introduction
• Introduction: Student engagement is critical to academic success, but traditional methods of monitoring,
such as manual observation, are often time-consuming, subjective, and inconsistent. This project develops
a mobile application that uses machine learning to detect and classify student engagement from
classroom images in real time. By analyzing visual cues such as facial expressions, body posture, and eye
movement, the app categorizes students into engaged and not engaged states, with further breakdowns
into emotional and behavioral subclasses like Confused, Bored, and Looking Away. This automated
solution provides instant feedback to educators, enabling timely interventions and personalized
learning strategies, ultimately fostering a more engaged and productive classroom environment.
• Problem Statement: Monitoring student engagement in the classroom is often subjective and time-
consuming, relying on manual observation that can be inconsistent. Since engagement directly affects
learning outcomes, there is a need for an automated, real-time solution to assess and track student
attention. Current methods lack timely feedback for teachers to intervene and adapt teaching strategies.
This project aims to develop a machine learning-based mobile app to accurately detect and classify
student engagement, providing educators with immediate insights to improve classroom management
and enhance learning outcomes.

3
Literature Review
S.No Title Year Of Tecchniques Used Disadvantages
Publication

Deep learning method with

Student expression detection in e-
1 2014 convolutional neural networks High computational
learning platforms.
(CNNs) overhead

Poor scalability for

Attention Monitoring Using Emotion Combined facial emotion and
2015 large classrooms with
2 and Gesture Recognition gesture analysis.
diverse environments.

Deep Facial Analysis for Engagement CNN and Multi-classification High memory
3 2017 requirements
Tracking

Facial Analysis-Based Attention Multi-scale pyramid representation Optimal scales and

Detection 2020 for extracting interesting points. parts defined
Human attention guides selection heuristically by
4 of spatial scales. specialists.
Mutual position and
spatial resolution of
parts are arbitrary.

4
S.No Title Year Of Publications Technique Used Diasadvantages

Real-Time Student
Literature Review Hybrid ML models combining deep and
5 Engagement Prediction Using 2022 Slower inference time
traditional classifiers.
Facial Cues

6 Student Engagement 2023 Deep CNN-based Learning Approach. using limited dataset
Detection in Classroom

2023 YOLOv5 combined with Contextual Attention Complex classroom

7 Student Behavior Recognition mechanism environments with
in Classroom Based on Deep OpenPose with improved YOLOv5 feature map occlusion challenges
Learning behavior recognition.

Engagement Detection in
complexity in model
Online Learning Based on Pre-
Pre-trained Vision Transformer (ViT) network training, integration, and
8 trained Vision Transformer 2024
Temporal Convolutional Network (TCN) computational
and Temporal Convolutional
requirements.
Network

Student’s Attention and Computer vision, machine learning, and AI Limited exploration of
9 Alertness Monitoring System 2024 Analyzing webcam footage for real-time variations in
using AI feedback engagement across
academic disciplines

5
DATASET

6
Existing System and Disadvantages
Existing System:
Utilizes deep learning models, particularly Convolutional Neural Networks (CNNs), for detecting
student engagement.
Leverages complex architectures like ResNet, VGG, and Inception for high-accuracy classification of
student behaviors and activities.
Often requires significant computational power, memory, and internet access, limiting usability in
remote or low-resource educational settings.

Disadvantages:
• Costly and inaccessible for many educational institutions due to resource requirements.
• Inefficient in real-time scenarios due to slow inference times and high computational demand.
• Dependency on stable internet connection for cloud-based solutions, making it impractical in
areas with poor connectivity.
• Lacks emotional insight, unable to detect nuanced student engagement states like confusion or
frustration for timely intervention.

7
Proposed System and Advantages
Existing System:
Uses MobileNet, a lightweight AI model, optimized for real-time student
engagement detection on mobile devices.
Low computational requirements, designed to function offline, eliminating the need
for internet connectivity.

Advantages:
• Enhanced Accuracy: Achieves high accuracy in multi-class engagement classification,
detecting engaged and not engaged states with subclasses (Confused, Bored, Drowsy,
etc.).
• Real-Time Feedback: Provides instant insights into student engagement, enabling
timely interventions for educators.
• Scalable & Accessible: Affordable and mobile-based, making it scalable for various
educational settings, even in low-resource or remote classrooms.
• Emotionally Intelligent: Capable of detecting emotional states such as confusion,
frustration, and boredom, providing actionable feedback for personalized teaching.

8
Software Requirements:
System Requirements
• Android/iOS Development Environment: For developing and deploying the mobile application.
• TensorFlow Lite: For deploying the trained machine learning model on mobile devices.
• Python: For model training, data preprocessing, and feature extraction.

Hardware Requirements:

• Mobile devices with moderate processing capacity

Dataset Requirements:

• Custom student engagement dataset with 2,120 labeled images, representing various engagement states
(Engaged: Confused, Engaged, Frustrated; Not Engaged: Looking Away, Bored, Drowsy).

10
Architecture Diagram

11
Modules List

1.Data Preprocessing

2.Model Development

3.Model Training

4.Model Evaluation

12
Modules
Algorithm:
Explanation and Algorithm
• MobileNet for feature extraction and SVM for classifying student engagement based on those features.
The steps in the data preprocessing includes:
Data Collection
• Gather a diverse dataset of classroom images, including various engagement states such as
Engaged and Not Engaged with subclasses like Confused, Bored, Drowsy, etc..
• Here we have used Student-engagement Dataset from Kaggle.
Data Augmentation
• Apply techniques such as rotation, flipping, and zooming to increase dataset size and variability.
Image Resizing
• Standardize all images to a uniform size (e.g., 224 x 224 pixels) for consistent input to the model.
Normalization
• Scale pixel values to a range of 0 to 1 to improve model performance and convergence speed.
Data Splitting
• Divide the dataset into training, validation, and test sets to ensure robust evaluation of model
performance and prevent overfitting.

13
Performance Evaluation
Metrics
• Accuracy: Measures the overall reliability of the model in correctly classifying
student engagement states (e.g., Engaged vs Not Engaged).
• Precision: Indicates the true positive rate for engagement detection, assessing how
accurately the model identifies students in specific engagement states (e.g.,
Confused, Bored).
• Recall: Measures the model's effectiveness in identifying all engagement cases,
ensuring that no critical states are missed.
• F1-Score: Provides a balanced measure of precision and recall, combining both to
evaluate the model’s ability to detect engagement without sacrificing accuracy.
Model Size
• Approximately 2.5 MB, ideal for mobile devices.

Inference Time
• The model is fast and optimized for real-time applications, ensuring quick
responses and efficient engagement detection during live classroom activities.
14
References
[1] Al-Sakib, A., Islam, F., Haque, R., Islam, M. B., Siddiqua, A., & Rahman, M. M.
(2024). "Classroom Activity Classification with Deep Learning." 2024 International
Conference on Integrated Circuits and Communication Systems (ICICACS). IEEE
Xplore.

[2] Saba, T., Rehman, A., Jamail, N. S. M., Marie-Sainte, S. L., Raza, M., & Sharif, M.
(2021). "Categorizing the Students’ Activities for Automated Exam Proctoring Using
Proposed Deep L2-GraftNet CNN Network and ASO Based Feature Selection
Approach." IEEE Access, 9, 47639-47656. IEEE Xplore.

[3] Zhang, L., Sengan, S., & Manivannan, P. (2022). "The Capture and Evaluation
System of Student Actions in Physical Education Classroom Based on Deep
Learning." Sensors, 22(4), 1476. MDPI.

15
Thank you
16