01
Model Development
Integrated detection, alignment, feature extraction, and identity classification into one real-time pipeline.
Data Science & AIReal-Time Facial RecognitionFacial Recognition
BSc. Data Science Research Project · 2024
ICTlab, University of Science and Technology of Hanoi
Hanoi, Vietnam
Real-Time Facial Recognition by Deep Learning
Computer Vision · Deep Learning · Machine Learning · Real-Time Systems
Supervision
Project Overview
A group data science project focused on developing a real-time facial recognition system for contactless attendance. The system combined face detection, alignment, feature extraction, and identity classification to recognise multiple individuals from live camera input.
Role
Project Lead
Data Science Researcher & Developer
Context
Final Group Project · 2024
Collaboration
Mau Minh Phuc Le
Duc Thai Phung
Hai Minh Hua
Minh Hoang Pham
Areas
Computer Vision · Deep Learning · Real-Time Recognition
Tools & Technologies
Python · YuNet · ResNet34 · Support Vector Machine · Face Alignment · Real-Time Video Processing
My Contribution
02
System Testing
Tested recognition accuracy and processing speed across live-camera conditions.
03
Performance Analysis
Analysed evaluation results and identified improvements for recognition reliability.
Dataset
WIDER FACE · GLINT360K · AgeDB Database
Manual collection dataset · 416 face images of 25 people
System Pipeline
Live video frames pass through five processing stages before the system returns an identity prediction.
01
Live Camera Input
Capture continuous video frames from a connected camera.
02
Face Detection
Locate one or more faces in each frame using YuNet.
03
Face Alignment
Normalise detected faces using facial landmark positions.
04
Feature Extraction
Generate numerical face representations using a ResNet34 model.
05
Identity Classification
Match extracted features with registered identities using an SVM.
Project Outputs
A functional real-time facial recognition prototype for contactless attendance, capable of identifying multiple people from live camera input at approximately 30 FPS.
View Report