Deteksi Aktivitas Merokok Menggunakan Mobile Vision Transformer dan Class Activation Mapping

Abstract

Smoking is one of the leading causes of various chronic diseases, including cardiovascular disease, lung cancer, and respiratory disorders, and contributes significantly to the increasing global mortality rate. Monitoring smoking activities in public areas, particularly in smoke-free zones, remains a major challenge due to limited human resources, the extensive coverage area, and the complexity of smoking behavior, which is difficult to detect manually. This study proposes a computer vision-based smoking activity detection system using the Mobile Vision Transformer (MobileViT) architecture, a lightweight and efficient deep learning model capable of extracting local features while capturing global contextual information from images. To improve the interpretability of the classification results, the Gradient-weighted Class Activation Mapping (Grad-CAM) method is employed to visualize the image regions that most influence the model's predictions. The Cigarette Smoker Detection dataset from Kaggle is used for training and testing, with preprocessing steps including image normalization, data augmentation, image resizing, and an 80:20 split for training and validation/testing, respectively. The model performance is evaluated using accuracy, precision, recall, F1-score, confusion matrix, and Grad-CAM visualization. The experimental results demonstrate that MobileViT effectively classifies smoking and non-smoking images with high accuracy, while Grad-CAM enhances the transparency and interpretability of the model's predictions. Therefore, the proposed system has strong potential to support the automated, efficient, and accurate monitoring of smoke-free zones as part of broader public health improvement efforts.