Object Detection with YOLO (Ultralytics)

This project demonstrates an end-to-end object detection pipeline using Ultralytics YOLO and PyTorch.
It is designed as a clean, reproducible portfolio project that covers dataset handling, training, evaluation, inference, and model export.

The project uses a custom PPE Helmet Detection dataset (single class: helmet) that is already included in the repository.

---

Project Overview

Task: Detect safety helmets in images
Model: YOLO (Ultralytics)
Framework: PyTorch
Classes: 1 (helmet)
Dataset: PPE Helmet Detection (custom)

This repository is structured so that it can be cloned and run immediately, without external dataset downloads or API keys.

---

Repository Structure

Object_Detection/
│
├── data/
│   └── PPE/
│       ├── train/
│       │   ├── images/
│       │   └── labels/
│       ├── valid/
│       │   ├── images/
│       │   └── labels/
│       ├── test/
│       │   ├── images/
│       │   └── labels/
│       └── data.yaml
│
├── scripts/
│   ├── train.py
│   ├── eval.py
│   ├── infer.py
│   ├── export.py
│   └── check_dataset.py
│
├── assets/
│   └── results/         # Training curves & prediction images (added manually)
│
├── runs/                # YOLO training outputs (auto-generated, not committed)
├── environment.yml
└── README.md

---

Dataset

• The dataset is included locally under data/PPE
• No external download or API keys are required
• Paths in data.yaml are relative for portability

data.yaml

path: data/PPE
train: train/images
val: valid/images
test: test/images

nc: 1
names: ['helmet']

---

Environment Setup

This project uses conda.

conda env create -f environment.yml
conda activate yolo

---

Dataset Sanity Check

Before training, verify the dataset:

python scripts/check_dataset.py --data data/PPE/data.yaml

This checks:

• Image–label matching
• Missing files
• Dataset size

---

Training

python scripts/train.py --data data/PPE/data.yaml

Outputs:

• Training curves
• Validation metrics
• Saved model weights

All results are saved automatically under runs/.

---

Evaluation

python scripts/eval.py --data data/PPE/data.yaml

Evaluates:

• Precision
• Recall
• mAP

---

Inference

Run inference on sample images:

python scripts/infer.py

Predictions are saved with bounding boxes and confidence scores.

---

Model Export

Export the trained model for deployment:

python scripts/export.py

Supported formats include:

• ONNX
• TorchScript

---

Results

Note: The images below are placeholders.
After training, copy selected result images from runs/ into assets/results/ and update or keep the filenames as shown.

Training Metrics

• Training and validation losses decrease smoothly
• Indicates stable convergence and good generalization
• No significant overfitting observed

Detection Metrics

• mAP@0.5 ≈ 0.8
• mAP@0.5:0.95 ≈ 0.5–0.55
• Precision reaches ~0.85–0.9
• Recall stabilizes around ~0.65–0.7

These values indicate reliable helmet detection with strong precision.

Confusion Matrix

• 48 true helmet detections
• 6 missed helmets
• 24 false positives from background

Sample Prediction

---

Key Features

• Clean, modular script-based workflow
• Portable dataset configuration
• End-to-end ML pipeline
• Suitable for recruiters and technical reviews

---

Notes

• Dataset download scripts were intentionally removed to keep the project fully self-contained
• This reflects real-world ML workflows where datasets are versioned locally

---

Author

Muhammad Qasim