Helmet Detection

A deep learning project that classifies whether a person in an image is wearing a helmet or not, using Convolutional Neural Networks (CNNs) built with PyTorch.

Overview

This project uses a Kaggle helmet detection dataset containing 764 labeled images across two categories:

• With Helmet (label: 0)
• Without Helmet (label: 1)

The primary motivation is road and construction site safety — this classifier can be integrated into automated safety monitoring systems.

Project Structure

helmet-detection/
├── data/                        # Preprocessed cropped images and CSVs
│   ├── objects.csv              # Bounding box info, cropped image names, labels
│   └── images.csv               # Original image names
├── preparation.ipynb            # Data preprocessing and EDA
├── final-project.ipynb          # Model training and evaluation
├── final-report.md              # Detailed project report
├── project_proposal.md          # Initial project proposal
├── progress_report.md           # Mid-project progress update
└── 23f_hw04.pdf                 # Assignment reference

Methodology

Data Preprocessing

The raw dataset consists of 764 images with XML annotation files containing bounding box coordinates and labels. Images were cropped to their bounding boxes, resized to 45×45 pixels, and stored alongside two CSV files (objects.csv, images.csv). An 80/20 train-validation split was applied based on original images to prevent data leakage.

Model Architecture

A custom ObjectClassifier class was used to systematically experiment with CNN architectures, varying:

• Number of convolutional layers
• Kernel sizes and output channels
• Dropout rates
• Loss functions (NLLLoss vs CrossEntropyLoss)

All models shared:

• Activation: ReLU (conv layers), LogSoftmax (output)
• Pooling: Max pooling with kernel size 2
• FC layers: 128 → 2 units
• Optimizer: Adam (lr = 0.001)
• Epochs: 50 | Batch size: 32

Model Comparison

Model	Conv Layers	Loss Function	ROC-AUC	Test Accuracy
Model 1	2 (16, 32 ch)	NLLLoss	0.812	0.884
Model 2	3 (16, 32, 64 ch)	NLLLoss	0.628	0.894
Model 3	3 (16, 32, 64 ch)	CrossEntropyLoss	0.787	0.893
Model 4	2 (16, 32 ch)	CrossEntropyLoss	0.803	0.880

ROC-AUC was prioritized over accuracy to minimize false positives (i.e., classifying someone without a helmet as wearing one).

Results

Model 1 achieved the best ROC-AUC score of 0.812 with a test accuracy of 88.4%. The model performs well overall but struggles with edge cases where a person is wearing some other head covering that resembles a helmet — a limitation tied to the small dataset size.

Requirements

• Python 3.x
• PyTorch
• torchvision
• NumPy
• pandas
• scikit-learn
• matplotlib
• Pillow

Install dependencies:

pip install torch torchvision numpy pandas scikit-learn matplotlib pillow

Usage

1. Data Preparation: Run preparation.ipynb to preprocess raw images and generate the data/ directory with cropped images and CSVs.
2. Training & Evaluation: Run final-project.ipynb to train the CNN models and evaluate performance.

Dataset

The dataset is sourced from Kaggle: Helmet Detection by Andrew MVD. Download and place the raw images in the appropriate directory before running the preprocessing notebook.

Author

Harshvardhan Bhatnagar — GitHub