PPE Compliance Detector

Real-time construction-site PPE compliance detection, built by fine-tuning YOLOv11 (Ultralytics) on a labeled construction-safety dataset. Detects people, hardhats, safety vests, and violations (missing hardhat / missing vest) from images, video, a live webcam, or a batch folder — with a Streamlit demo app for all four.

Why this exists

Manual PPE compliance checks on a construction site are slow and inconsistent. This fine-tunes a real-time detector to flag violations automatically from a camera feed or recorded footage — a lightweight, deployable alternative to a manual walk-through audit.

Tech stack

• Model: YOLOv11s (Ultralytics), transfer-learned from COCO-pretrained weights
• Inference: OpenCV
• Deployment: Streamlit
• Training: Google Colab (T4 GPU)

Dataset

Fine-tuned on the Construction Site Safety dataset from Roboflow Universe (CC BY 4.0), ~2,801 images across 10 classes: Hardhat, Mask, NO-Hardhat, NO-Mask, NO-Safety Vest, Person, Safety Cone, Safety Vest, machinery, vehicle.

Setup

git clone https://github.com/SuryadevChippada/Object-Detection.git
cd Object-Detection
pip install -r requirements.txt

Training

1. Export the dataset above from Roboflow in YOLO format into data/processed/ (Roboflow generates its own data.yaml there — don't hand-write one).
2. Run training (Colab recommended — see hyperparameters in configs/train.yaml):

   python src/train.py

3. Copy the resulting weights into the repo:

   cp runs/detect/train/weights/best.pt models/weights/best.pt

Evaluation

python src/evaluate.py --weights models/weights/best.pt --split test

Reports precision, recall, mAP50, and mAP50-95 on the held-out test split. A confusion matrix and PR curves are saved alongside the run.

Results: to be added once training completes.

Inference (CLI)

# Single image
python src/infer.py --source path/to/image.jpg --output out.jpg

# Video
python src/infer.py --source path/to/video.mp4 --output out.mp4

# Webcam
python src/infer.py --source 0

# Batch folder
python src/infer.py --source path/to/folder --output path/to/folder/annotated

Web app

streamlit run app.py

Supports image upload, video upload, live webcam, and batch-folder modes, with an adjustable confidence threshold.

Project structure

├── configs/
│   └── train.yaml       # training hyperparameters
├── models/weights/       # best.pt (produced by training)
├── src/
│   ├── train.py
│   ├── evaluate.py
│   ├── infer.py
│   └── utils.py
├── app.py                # Streamlit entrypoint
└── requirements.txt