Real-time Personnel Detection System with Adaptive Scheduling for Large-scale Retail Deployment

Microservices-based Framework for Real-time Retail Surveillance

Keywords: Microservices Architecture, Retail Security, Computer Vision, AWS

Brief Description

This project extends my camera-feed-processing pipeline (github repo) to implement a scalable intruder detection system for multiple retail locations. The primary challenge addressed was scaling the system across numerous stores, which was achieved through an asynchronous microservices architecture.

The system processes multiple video streams, detecting unauthorized personnel, storing relevant frames, and generating email alerts. Its architecture is designed for horizontal scalability, accommodating an increasing number of retail locations without significant performance degradation.

underlying system design - discussed on github; this post is about a specific processing service
Key Technologies and Architecture
  • Computer Vision: YOLOv8 for real-time personnel detection
  • Scheduler Service: Manages which feeds to capture at a given time, with potential for future expansion into process scheduling and distributed scaling
  • Data Management:
    • Redis for in-memory frame storage
    • RabbitMQ as a message queue for efficient frame-metadata distribution
  • Cloud Storage: AWS S3 for alert frame storage
  • Visualization: Dashboard for cross-location detection monitoring and analytics

note: omitting the illustration of the dashboard and detections for proprietary reasons