Drunken4VN: A Thermal Infrared Imaging Dataset for Alcohol Intoxication Detection in Vietnamese under Diverse Environmental Conditions
Description
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.
Files
Steps to reproduce
The Thermal Drunk Processing Dataset has been created to remedy the lack of a culturally and environmentally sound thermal imaging dataset of the Vietnamese population. The data were recorded with the Total Meter HT-03 thermal camera, which has a thermal image resolution of 120 x 90 pixels within the 8–14 µm wavelength range. The data is all about binary classification of alcohol intoxication states into sober and drunk. 20 volunteers (17 males and 3 females) between the ages of 19-56 completed the study. All subjects were verified sober at baseline with breathalyzer testing or self-report of abstinence from alcohol for 24 hours. The data collection procedure included thermal scans of several body areas, including the entire face, left facial side, right facial side, the arms, and the legs. 220 ml of two bottles of Saigon Lager beer (alcohol by volume, 4.6%; one bottle, 330 ml) were taken in three doses at intervals of 20 minutes within the experiment. BAC 0.25 to 0.32 mg/L was obtained after consumption to measure levels of intoxication. Interestingly, observations were taken under diversified real-life conditions of the environment, primarily in the evenings indoors at 23–25°C ambient temperature and outdoor areas during mornings at 35–37°C temperature. This is a contrast to other data that were recorded in laboratory-controlled situations, and therefore, there is improved ecological validity. Thermal measurement confirmed large physiological responses to the consumption of alcohol, especially in facial areas. In particular, skin temperature rose 2–3°C during drinking and fell 3–4°C after the last dose, reflecting thermal changes that are directly associated with measures of intoxication. Such observations are consonant with the promise of the data for the support of efficient, real-world applications in the detection of alcohol intoxication by thermal imaging The Thermal Diversity Dataset was created. The Thermal Diversity Dataset enhances the generalization and robustness of intoxication detection models, especially for real-world, time-varying scenarios. The Thermal Diversity Dataset is larger in sample size and image diversity. It consists of two primary classes: sober and drunk. For the sober class, thermal frames were captured for 45 subjects, more than 800 images from significant body regions (the entire face and side facial angles), similar to the regions scanned for the original dataset. For the drunk class, 15 subjects were instructed to consume at least 2 beers (approximately 660ml). A collection of a range of different thermal face images was then gathered, with a range of distinct facial orientations, to further improve dataset balance and coverage. The dataset comprises 1,720 high-resolution thermal face images, providing greater pose, face shape, and physiological state variability.
Institutions
- FPT University