Challenge:
A startup needed a real-time face recognition system to track customers in a retail environment. The project faced key challenges, including:
- • Masked Face Detection – The system needed to detect and verify faces even if partially covered.
- • Low-Resolution Image Handling – Cameras produced variable quality images, requiring adaptive processing.
- • Appearance Variations – The system needed to recognize repeat visitors despite changes in hairstyles, clothing, and accessories.
Solution:
QuantumHub developed and deployed an ML-driven face recognition system designed for high-traffic retail spaces.
Key Components and Approach:
- Person Detection & Face Recognition:
- Implemented YOLOv5 for real-time person detection.
- Utilized FaceNet for face recognition, ensuring high accuracy even with partial occlusions.
- OSnet deep neural network for re-identification, allowing visitor tracking across multiple cameras.
- Data Preprocessing and Model Training:
- Structured datasets with diverse face angles, resolutions, and lighting conditions.
- Preprocessed image data to enhance clarity and reduce noise.
- Trained models using precision-recall curves, ROC curves, and mean average precision (mAP) for optimal performance.
- Workflow & Logic:
- Designed ML workflows to handle entry/exit detection, visitor tracking, and masked face verification.
- Integrated Redis for real-time data handling and alert management.
Outcome:
- Improved Customer Analytics: Tracked visitor patterns and repeat customers, providing valuable insights for marketing and operations.
- Enhanced Security: Masked face detection reduced security vulnerabilities by identifying individuals even with face coverings.
- Scalable Infrastructure: System capable of handling multiple camera streams simultaneously with minimal latency.
Technology Stack:
- Machine Learning Models: YOLOv5, OSnet, FaceNet
- Backend: NodeJS, Express
- Frontend: ReactJS
- Cloud Services: AWS Lambda, RDS, DynamoDB, Cognito
- Security: JWT, Helmet