Infrared Deep Learning System for Non-Invasive Drunk Detection and Alcohol Level Classification
Date
2024-12
Authors
P Acharya, Lipsa
Major Professor
Aleksandar, Dorgandzic
Advisor
Committee Member
Dorgandzic
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This project proposes a novel approach to alcohol detection using an infrared (IR) camera integrated with deep learning technology. Designed to address delays and contamination issues associated with conventional methods such as breath analyzers, the system utilizes an FLIR ONE thermal camera to capture both infrared and visual images, which are processed via a deep learning-based classification system. The system operates on an iPhone, offering portability and convenience. The dataset comprises images from 50 volunteers, each contributing data across four levels of alcohol consumption: sober, one glass, two glasses, and three glasses. Data augmentation techniques expanded the dataset to over 1,200 images, improving the robustness and generalization of the classification models. Two classification tasks were performed: binary classification (sober vs. drunk) and four-level classification (sober, 1 glass, 2 glasses, 3 glasses). The NasNetMobile model achieved the highest accuracy for the four-level classification task at 85.10%, while MobileNet performed best in the binary classification task, achieving an accuracy of 74.07%. The use of Grad-CAM interpretability methods demonstrated that the models focused effectively on relevant features, such as facial regions affected by alcohol consumption.
Despite these promising results, challenges such as distinguishing minor differences between consumption levels and accounting for external factors like body temperature variations remain. This research highlights the potential for non-invasive, real-time alcohol detection systems and their broader implications for public safety and health monitoring. Future work will focus on addressing limitations and improving system scalability for real-world applications.
Series Number
Journal Issue
Is Version Of
Versions
Series
Academic or Administrative Unit
Type
creative component
Comments
Rights Statement
CC0 1.0 Universal
Copyright
2024