YOLO-PPE-Detection

YOLOv11n - based PPE detection model trained for iOS real-time inference via CoreML

A YOLOv11n-based Personal Protective Equipment (PPE) detection model trained for real-time inference on iOS via CoreML.

This repository is part of a workplace safety inspection app that combines PPE detection with pose estimation (MoveNet) in a single camera pipeline.

---

Detection Classes

ID	Class	Description
0	Hardhat	Safety helmet worn
1	NO-Hardhat	Safety helmet not worn
2	NO-Safety Vest	Safety vest not worn
3	Person	Person
4	Safety Vest	Safety vest worn

---

Results

Metric	Value
Precision	0.650
Recall	0.726
mAP50	0.706
mAP50-95	0.463

Training curves, confusion matrix, and per-epoch metrics are available in results/.

---

Repository Structure

yolo-ppe-detection/
├── data.yaml              # Class definitions (set your own dataset path)
├── train.py               # Training script with full hyperparameter config
├── export_coreml.py       # CoreML export script with NMS support
├── results/
│   ├── results.png        # Loss & mAP curves
│   ├── confusion_matrix.png
│   ├── confusion_matrix_normalized.png
│   ├── BoxF1_curve.png
│   ├── BoxPR_curve.png
│   ├── results.csv        # Per-epoch metrics
│   └── args.yaml          # Full training arguments
├── weights/
│   └── README.md          # Download links for best.pt and .mlpackage
└── .gitignore

---

Dataset

• Source: Roboflow PPE public datasets
• Training images and label files are not included in this repository due to dataset license restrictions.
• Refer to data.yaml for the class structure and configure the dataset path for your local environment.

---

Setup

pip install ultralytics coremltools onnx

---

Training

python train.py

Key training configuration:

Parameter	Value	Notes
Model	yolo11n.pt	YOLOv11 nano
Epochs	100
Batch	10
imgsz	832	Larger input for small PPE objects
Optimizer	auto
cos_lr	True	Cosine annealing LR schedule
close_mosaic	10	Mosaic disabled for last 10 epochs
patience	100

---

CoreML Export

python export_coreml.py --weights runs/ppe_yolo11n_832/weights/best.pt --imgsz 640 --nms

The export process requires multiple steps. See export_coreml.py for details on options and known issues.

Option	Value	Notes
imgsz	640	Reduced from 832 for mobile inference speed
nms	True	NMS included inside the model
half	False	FP32 (FP16 optional, verify on device)

---

iOS Integration

The exported .mlpackage is used in an iOS app with the following pipeline:

AVFoundation (camera frames)
    ↓
Vision (CoreML request handling)
    ↓
CoreML (PPE object detection)
    ↓
UIKit (bbox & label overlay)

Key considerations when integrating with Xcode:

• Vision returns bounding boxes with bottom-left origin — Y-axis flip required for UIKit
• Explicit orientation must be passed to VNImageRequestHandler to avoid rotated detections in portrait mode
• iOS 16 and iOS 17+ handle videoOrientation differently

---

Weights

Pre-trained weights are available on the Releases page.

File	Description
best.pt	YOLOv11n trained weights (PyTorch)
DetectionYolov11.mlpackage	CoreML model for iOS

---

Training Environment

Item	Details
Platform	Google Colab Pro
GPU	A100
Framework	Ultralytics 8.3.179
Python	3.13

---