HygieneVision

This repository implements a real-time computer vision monitoring system designed to enforce environment hygiene compliance. Utilizing a custom-trained YOLOv8 model and OpenCV, the system detects critical adherence failures (e.g., missing hair caps, missing gloves) via webcam streams, recorded video, or static evaluation.

It features a modular inference engine, a FastAPI integration for stateless API consumption, and a Streamlit dashboard with persistent PostgreSQL/MongoDB event logging.

Project Architecture

• hygiene_engine.py: The core analytical engine. Encapsulates the YOLOv8 model loading and inference logic. Computes custom Intersection-over-Face (IoF) overlap metrics against secondary Haar Cascade datasets to accurately isolate spatial non-compliance (synthesizing predictions for occluded or heavily bounded classifications).
• dashboard.py: A synchronized, single-page web application built with Streamlit. Manages local hardware interfacing (webcam) without WebRTC overhead, handles concurrent multi-modal uploads, and provides a real-time reactive UI.
• hair-test.py: A headless, high-efficiency Python script optimized for edge-device deployment. It removes frontend rendering overhead to strictly execute localized OpenCV operations and console logging.
• app.py: A FastAPI application providing a REST interface (/predict). Returns bounded image arrays natively encoded in Base64 alongside raw class-confidence dictionaries.

Prerequisites

• Python 3.10+
• MongoDB instance (locally hosted or URI string)
• pip / pyenv

Installation

1. Clone this repository and navigate to the root directory.
2. Install the required dependencies:

pip install -r requirement.txt
pip install pymongo

System Configuration & Execution

The system supports multiple execution vectors depending on the deployment environment.

1. Persistent Data Layer (MongoDB)

The analytics dashboard natively writes compliance violations to a local MongoDB instance. Start a localized mongod background task bound to port 27019 to prevent conflicts with standard operational databases:

mongod --dbpath ./data_27019 --port 27019 --logpath ./data_27019/mongod.log &

2. Interactive Dashboard (Streamlit)

For continuous monitoring via local peripherals with full database synchronization:

streamlit run dashboard.py

Access via: http://localhost:8501

3. Edge-Device Headless Script

For maximum compute efficiency and frame processing throughput (e.g., deployment on SOC infrastructure without a GUI/X-server):

python hair-test.py

4. REST API (FastAPI)

To expose the inference engine to external services, trigger the Uvicorn ASGI server:

uvicorn app:app --host 127.0.0.1 --port 8000

API documentation available at: http://127.0.0.1:8000/docs

Technical Notes

• Debounce Logic: The dashboard.py state machine implements a 10-second structural debounce per staff identity to prevent database fragmentation during continuous exposure to single-event violations.
• Bounding Box Heuristics: The system utilizes coordinate scaling (minSize) and aggressive minNeighbors configurations within the auxiliary face detection cascade to mitigate dense background artifacts.