SiteWatch Edge

Offline-first PPE and restricted-zone screening with OpenCV 5 and privacy-minimised AWS sync.

Comment

Inspiration

Cloud-only safety analytics are a poor fit for sites with unreliable connectivity. Continuous video upload also creates privacy and bandwidth costs. SiteWatch Edge keeps the immediate vision decision local while sending only minimal incident metadata to an authorized cloud receiver when connectivity returns.

What it does

SiteWatch Edge analyzes images, recorded video, webcams, or RTSP feeds locally. OpenCV 5 DNN detects workers, hardhats, safety vests, selected missing-PPE classes, and heavy equipment. A local rule engine associates PPE with workers and checks polygonal restricted zones. Operators receive an immediate advisory overlay, while events queue offline for delayed synchronization.

How we built it

Python and OpenCV 5 provide capture, preprocessing, ONNX inference, non-maximum suppression, geometry, annotation, and video encoding. A local Flask interface supports upload-and-review workflows. The offline queue uses JSONL events and retries safely after connectivity returns.

The AWS component is deployed in ap-south-1 through AWS SAM and CloudFormation. API Gateway invokes a Python Lambda receiver, which authenticates the client and writes privacy-minimized metadata to an encrypted DynamoDB table with a 30-day TTL. An AWS Budget monitors the demonstration account.

Evaluation and results

  • Verified OpenCV 5.0.0 offline ONNX inference.
  • Detected worker, hardhat, and safety-vest classes on licensed evaluation images.
  • Generated restricted-zone incidents without network access.
  • Confirmed that an invalid cloud key returns HTTP 401 and an authorized synthetic incident returns HTTP 202.
  • Verified the resulting DynamoDB record and confirmed that the local evidence path was not stored.
  • Thirteen automated tests and Ruff checks pass.

The seven-image evaluation is a functional check, not a production accuracy benchmark. The submission documents failure cases, including a quantized YOLOX model that produced zero class scores in the tested OpenCV 5 DNN environment; the verified FP32 model is used instead.

Responsible use and limitations

SiteWatch Edge is an advisory screening tool, not a replacement for trained safety personnel or certified protective systems. It does not perform face recognition or identity tracking. Video remains local in the demonstrated architecture, cloud records contain only incident metadata, and the interface makes uncertainty visible. Field-camera performance, occlusion, lighting, temporal tracking, and site-specific calibration still require supervised validation.

What's next

Next steps are supervised field testing, temporal tracking, improved PPE-absence calibration, and a supervisor-reviewed zone editor.

Entrant

Dr. N. Chaitanya Kumar — Professor & Coordinator · KLH University, Bowrampet.

Share this project:

Updates

Submission history