How the Problem Was Solved

Technology Stack:

Python (Django) for backend, a mix of OpenCV and machine learning algorithms for video analytics, and a PostgreSQL database.

Key Features:

• Real-time video feed processing
• Intelligent alerts for anomalies (crowd detection, PPE compliance, restricted access)
• Configurable dashboards for different use cases (construction sites vs. schools)

Approach:

• Machine Learning Models: Trained image recognition models to detect PPE, count people, and identify restricted zone breaches.
• RESTful APIs: Django REST framework for feeding real-time alerts to the monitoring dashboard.
• Scalable Architecture: Containerized microservices to process multiple camera feeds concurrently.

The Results

Impact on the Business:

• Reduced security incidents by 40% in pilot locations.
• Minimized the need for large on-site security teams.

Cost:

• ~150K USD
• Investment in GPU-based servers for real-time video processing.

Time:

• 8 months to build and train initial models, with ongoing improvements as new anomaly types were added.

Lessons Learned

• What Went Well:
  • Integration of ML models into Django REST services was straightforward.
  • Containerization (Docker/Kubernetes) allowed for efficient scaling and deployment.
• What We Would Do Differently Next Time
  • Dedicate more effort to data collection and annotation upfront to improve model accuracy.
  • Include built-in analytics to measure false-positive/false-negative rates from the start.