Drunken4VN: A Thermal Infrared Imaging Dataset for Alcohol Intoxication Detection in Vietnamese under Diverse Environmental Conditions

This data set consists of thermal infrared images taken to enable binary classification of alcohol intoxication conditions of sobriety and drunkenness in the context of non-invasive imaging. It has two complementary subsets aimed at encouraging physiological validity as well as machine learning models' capacity to generalize in real-world applications.
Thermal Drunk Processing Dataset (TDPD): The TDPD dataset includes 560 photos of 20 Vietnamese volunteers (17 men, 3 women, aged 19–56 years). The participants drank two cans of Saigon Lager beer (ABV 4.6%) in three administrations, 20 minutes apart. Thermographic images were captured by a Total Mentor HT-03 thermal images (120 × 90 pixels, 8–14 µm wavelength) across multiple regions of the body, such as the whole face, left and right halves of the face, arms, and legs. Images were captured in different daily environmental conditions that ranged from indoor temperatures (23–25°C) to outdoor temperatures (30–37°C) at various times of the day. This dataset contains a variable number of images per class, with 140 sober and 420 drunk images from three drinking sessions.
The thermal diversity dataset (TDD) consists of 1,720 frontal facial thermal images from 60 subjects (45 sober, 15 drunk). The drunk subjects drank at least 660 ml of beer. Frontal facial images were taken from different facial poses (front, left, right) for additional pose diversity and balancing of the dataset. This subset includes facial thermal images only to enhance model generalization towards real-time detection.
By incorporating both methods and testing with an additional 20 people beyond the initial dataset, the total number of images in the dataset now amounts to 2,737. Specifically, there are 1,583 images depicting drunken individuals and 1,154 images showing non-drunken individuals. Both methods are extended to give further pairs of samples that are then combined to form an equilibrium binary classification problem. All images are saved in JPEG format with an average size of 30-50 KB, a fair compromise between storage capacity and quality. The dataset records important physiological temperature variations related to alcohol intake. Assembled from various environmental settings and time frames, this dataset fully represents the real climate in Vietnam and population diversity, and therefore is a precious source for constructing robust, population-driven drunkenness detection models in Southeast Asia.