AI-Powered PPE Compliance System for Construction

Gabriel Adebayo

ML Engineer

Bert

Python

Streamlit

Building & Construction

SafeSight AI Assistant(PPE_VISION_360)

Project tagline: From helmets to boots — AI-powered PPE compliance for safer construction worksites.

Overview (Problem → MVP solution)

Problem: In construction and industrial environments, workers often forget or misuse PPE (helmets, boots, gloves, vests). Current inspection workflows are manual, slow, and error-prone, increasing risk of accidents and compliance issues.

Solution / MVP: SafeSight AI Assistant is an image-based PPE compliance system. Users upload images of workers or site conditions, and the system:

Detects PPE items using YOLOv8.

Classifies worn vs. not-worn PPE.

Generates human-readable compliance reasoning using BERT + NER.

Provides a retrieval-augmented chatbot (FAISS + LLM) for PPE-specific questions.

Note: This MVP currently only works with images, not live camera feeds. Accuracy is limited due to small dataset and GPU constraints. Adding more labeled images will make the system more robust and reliable. Future deployment to embedded systems (Raspberry Pi / edge devices) is planned.

Single-sentence MVP description

An image-based PPE compliance assistant that detects PPE items in photos, classifies presence/absence, and provides explainable compliance reasoning and corrective recommendations.

What we built (current features)

YOLOv8 object detection for helmets, boots, gloves, and vests.

Segmentation & classification for partial or incorrectly worn PPE.

BERT-based OSHA-style question classifier for compliance reasoning.

NER module to extract PPE-related entities from reports.

LLM-powered chatbot (RAG) that answers only PPE compliance questions using FAISS retrieval.

Streamlit dashboard to visualize images, detections, explanations, and compliance reports.

Note: Future improvements include adding more PPE classes, expanding chatbot knowledge, and integrating real-time video monitoring.

🗂 Project Structure (Main Skeleton)

This is the main structure; the production codebase contains additional modules/configs beyond this outline.

project/
├── app/
│   ├── streamlit_app.py                # Home/Dashboard
│   ├── pages/
│   │   ├── 1_PPE_Image_Compliance.py
│   │   ├── 2_PPE_OSHA_QA.py
│   │   ├── 3_PPE_BERT_Classifier.py
│   │   ├── 4_PPE_NER_Tagger.py
│   │   ├── 5_PPE_Chat.py
│   ├── fastapi.py                          
│
├── src/
│   └── PPE_VISION_360/
│       ├── __init__.py
│       ├── image_detection.py
│       ├── osha_qa.py
│       ├── bert_classifier.py
│       ├── ner_tagger.py
│       ├── chat_llm.py                
│       ├── compliance_reasoning.py     
│       ├── hybrid_recommender.py     
│       
│
├── models/
│   ├── saved_distillbert/
│   ├── ppe_ner_model/
│
├── datasets/
│   ├── nlp/
│       ├── osha_qa_cleaned.csv
│       ├── faiss_index.bin
│       ├── qa_embeddings.npy
│
├── requirements.txt
└── README.md

Architecture

System overview (flowchart)

flowchart LR
    A[User uploads Image] --> B[Preprocessing - resize & augment]
    B --> C[YOLOv8 Detection]
    C --> D{Detected PPE Items}
    D -->|Helmet| E[Helmet classifier & segmentation]
    D -->|Boots| F[Boots classifier]
    D -->|Gloves| G[Gloves classifier]
    D -->|Vest| H[Vest classifier]
    E --> I[Compliance rules & thresholds]
    F --> I
    G --> I
    H --> I
    I --> J[Compliance reasoning - NER & BERT]
    J --> K[LLM Assistant / Chatbot]
    K --> L[Dashboard / Exportable Report - CSV/PDF]

Sequence: Image → Decision → Report

sequenceDiagram
    participant U as User
    participant I as Image
    participant M as Model (YOLO + Classifiers)
    participant R as Reasoner (BERT + NER)
    participant L as LLM
    participant D as Dashboard

    U->>I: Upload image
    I->>M: Detect PPE items
    M->>R: Classified & verified PPE
    R->>L: Generate compliance reasoning
    L->>D: Return explanation & chatbot response
    D->>U: Display results + download options

Chatbot RAG Explanation (Hybrid LLM + FAISS)

flowchart LR
    A[User Query / PPE Question] --> B[FAISS Vector Search]
    B --> C{Relevant PPE Context Found?}
    C -->|Yes| D[Retrieve Context & Metadata]
    C -->|No| E[Return Cannot Answer Outside PPE Scope]
    D --> F[LLM Generates Response Using Retrieved Context]
    F --> G[Provide Human-Friendly Answer / Recommendation]
    G --> H[Dashboard / Chat Interface]
    H --> A[User Sees Answer]

Quickstart — How to Run Locally

Prerequisites

Python 3.9+

pip

Virtual environment (venv / conda)

(Optional) GPU + CUDA for faster YOLO training

Install Dependencies

python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt

Prepare Data

Place labeled images in data/ (YOLO format).

Include osha_qna.csv for BERT + FAISS compliance queries.

Run Inference / Dashboard

# Detect PPE in images
python run_inference.py --weights runs/best.pt --source demo_images/

# Run Streamlit dashboard
streamlit run app/streamlit_app.py

Streamlit Cloud / Demo Link Placeholder

Limitations

Dataset size: Current dataset is small → limited model accuracy.

Only images supported: No live camera feed yet.

FAISS + LLM chatbot: Can only answer PPE-related questions.

GPU requirements: Training larger models may fail without GPU.

These limitations are expected for an MVP; scaling the dataset and improving compute resources will increase robustness.

Roadmap & Future Improvements

Expand dataset with more images from diverse construction sites.

Add more PPE classes and edge cases for YOLO.

Improve chatbot coverage with larger FAISS index and dynamic RAG.

Deploy on embedded devices for real-time monitoring.

Add role-based dashboard with audit logs and report exports.

Screenshots / UI Preview

You can add images here to show your dashboard:

How to Contribute

Fork the repository.

Add or improve dataset, annotations, or model code.

Submit a Pull Request (PR) including tests and an example image demonstrating your changes.

👤 Author

Adebayo Gabriel – ML Engineer (**AI × PPE **)

Project lead: You (SafeSight AI Assistant)

Tools & models used: YOLOv8, BERT, FAISS, Streamlit, Python

🔗 Links

📧 iyanuoluwaadebayo04@gmail.com | 🌐 LinkedIn | GitHub

Like this project

Posted Sep 1, 2025

Developed an AI-powered PPE compliance system for construction safety.

Likes

Views