🪖 Helmet Violation Detection using YOLOv8

This project detects violations of helmet rules by motorcyclists using computer vision and deep learning. It leverages the YOLOv8 object detection model to identify motorcyclists and classify whether they are wearing helmets.

---

📌 Project Overview

Motorcycle riders are at high risk in road traffic, especially in developing countries. Wearing helmets is mandatory in many regions, but enforcement is difficult. This project aims to automatically detect riders and passengers who are not wearing helmets using video surveillance data.

---

🔧 Technologies Used

• Python 3.x
• Ultralytics YOLOv8
• OpenCV
• Matplotlib
• Spyder (IDE)
• GPU: Trained and tested on NVIDIA RTX 3050 Ti

---

🗂️ Project Structure

helmet_detection_project/
├── videos/                    # Original video files (not uploaded to GitHub)
├── frames/                    # Extracted frames from videos
├── yolo_dataset/
│   ├── images/train/          # Images in YOLO format
│   └── labels/train/          # Corresponding YOLO labels
├── results/                   # Output prediction examples (.jpg, .mp4)
├── report/                    # Evaluation metrics and training results
├── train_polo.py              # Model training script
├── video_predict.py           # Runs inference on video files
├── convert_to_yolo.py         # Converts gt_cleaned.txt to YOLO format
├── label_visualizer.py        # Draws bounding boxes for visual check
├── fast_test.py               # Runs inference on sample image
├── gt_cleaned.txt             # Cleaned ground truth (1–9 ➜ 0–8)
├── data.yaml                  # YOLOv8 training configuration
└── README.md                  # This file

---

🧠 Dataset

The dataset consists of 100 real-world traffic surveillance videos (20s each).
Due to licensing restrictions, the dataset and extracted frames are not included in this repository.

• Annotation format:

  <video_id>, <frame>, <x>, <y>, <width>, <height>, <class_id>
  

• Class labels (after cleaning):

  0: motorbike
  1: DHelmet
  2: DNoHelmet
  3: P1Helmet
  4: P1NoHelmet
  5: P2Helmet
  6: P2NoHelmet
  7: P0Helmet
  8: P0NoHelmet
  

📦 Dataset can be requested from the official challenge page:
https://www.aicitychallenge.org/2024-challenge-tracks/

---

🏃‍♂️ How to Run

1. Install requirements

pip install ultralytics opencv-python matplotlib

2. Extract frames from videos

python process_videos.py

3. Clean and convert annotations

python clean_gt.py
python convert_to_yolo.py

4. Train the model

python train_polo.py

5. Run inference on images

python fast_test.py

6. Run inference on videos

python video_predict.py

---

🖼️ Example Predictions

Here are some sample outputs from the trained YOLOv8 model:

---

📈 Outputs

Trained model and logs are saved under:

runs/helmet-detection/

Evaluation files are stored under the report/ folder:

• report/results.png: training curves
• report/results.csv: contains mAP, precision, recall scores for each epoch
• report/confusion_matrix.png: class-wise performance overview
• report/P_curve.png, report/R_curve.png, report/F1_curve.png: class-wise precision-recall-f1 charts

---

📊 Evaluation Summary

• mAP@50: ~0.80
• Precision: ~0.75
• Recall: ~0.70

The model was trained using YOLOv8 on a cleaned dataset of helmet usage in real traffic videos.
It successfully detects both motorbike riders and their helmet status with high confidence.

---

🛑 License Notice

The dataset is licensed for academic and challenge use only under the original DATASET LICENSE AGREEMENT.
Therefore, the dataset is not included in this GitHub repository in accordance with the license terms. Interested users may request access through the official competition website.

---

🧑‍💻 Author

• Kağan Kılıç and Mehmet Çokol
  Eskişehir Teknik Üniversitesi
  Computer Vision Challenge Track 5 – Helmet Rule Violation Detection

---