GenAI – Image Annotation - Camera-Based Personnel Detection in Hospitals

SUMMARY

A hospital surveillance AI project aimed to identify and track personnel across 25,000 image frames without using facial recognition. We supported privacy-first data annotation using visual cues like clothing and accessories. The goal was to train AI for personnel recognition in real-world, privacy-sensitive environments.

THE CHALLENGE

• No facial recognition allowed due to privacy—faces were masked.
• Identification relied solely on clothing, gear, hairstyle, and footwear.
• Tracking 18 individuals (60% male, 40% female) in various scenes.
• Needed consistent accuracy across 8 unique real-world hospital scenarios.

SOLUTION

• Human detection and tracking across sequential image frames.
• Used COCO format for output compatibility and AI training.
• Integrated with partner annotation platform.
• Team of 10 annotators and 4 quality inspectors with a layered QA model.
• 100% quality control at Layer 1, and 30% sampled QC at Layer 2 for robustness.

KEY OUTCOMES

• Successfully annotated 25,000 surveillance frames with masked faces.
• Enabled AI to learn pattern-based recognition beyond facial features.
• Maintained high-quality annotations with a two-layered QC approach.
• Delivered within a tight 2-month timeline with zero data privacy concerns.