Face Recognition Ticketing Application

Project Overview

Developed a native Android-based facial recognition application that served as a gate access mechanism for a football match ticketing system at a stadium in Jakarta. This technology was used by gate operators to quickly and accurately verify spectator identities using mobile device cameras.

Project Duration: August 2024 - November 2024

User Interface Application

View Screenshots

1	2	3	4

Disclaimer

All product names, logos, and brands are anonymized. Screenshots are used solely to demonstrate technical experience.

Project Description

General Overview

General Problem

• Conventional ticket validation processes are vulnerable to abuse
• Long queues at stadium gates on match days
• High network load due to the transmission of large image data

General Solution

• Implementing Face Recognition directly on Android devices
• Reducing network load by extracting facial feature vectors
• Integrating mobile applications with ticketing backend systems

System Architecture

• Client Application for Operator Gate

  • TensorFlow Lite for Face Recognition model inference

• Backend Service

  • User and ticket management API
  • Face vector matching service

• AI Pipeline (Existing)

  • Model Face Recognition (baseline Python, converted to TFLite)

---

Problem Identification (Role-Specific)

Application Usage Context

• The application is used by stadium gate operators
• Runs on a dedicated client-managed Android device

Problem

• Initial Face Recognition implementations send full JPG image to backend
• Large data size causes:
  • High latency
  • Excessive network load
• Matching process becomes less efficient

---

Technical Solutions Implemented

Solution Approach

• Integrating AI model directly into mobile devices using TensorFlow Lite
• Ensuring optimal performance on Android devices

Technical Implementation

• Camera captures real-time face frames
• Frames are processed on device to extract 512-dimensional vector embeddings
• Only vector embeddings are sent to backend for matching

---

Challenges

Cross-Functional Collaboration

• Cross-Functional Collaboration: Working with the AI Team to standardize image preprocessing (such as cropping and normalization) to ensure consistent vector results.

• Adjusting inference logic between:

  • Python model (existing)
  • TensorFlow Lite model on mobile

• Validating the consistency of embedding results

Cross-Functional Collaboration

• Intensive collaboration with:
  • AI Team for preprocessing, normalization, and threshold adjustment
  • Backend Team for API design, user management, and Face Recognition data transmission

---

Impact

• Data transmission becomes very light compared to sending full images
• Decreased latency during the verification process at the gate
• Utilization of the computational power of client devices for facial vector extraction
• The system becomes more scalable and stable for large events

---

My Contribution to the Project

• Integrating Face Recognition Vector Extraction using TensorFlow Lite
• Implementing authentication and user profile management
• Integrating match data and common endpoints
• Distributing the application using MDM (Mobile Device Management) Dashboard provided by the client