DrunkSelfie: Intoxication Detection from Smartphone Facial Images
Colin Eric Willoughby, Ian Banatoski, Paul Roberts, Emmanuel Agu · 2019
Drunk Driving killed over 10,000 million people in 2015, accounting for nearly a third of all traffic-related deaths in the US. In many cases, drivers do not know they are over the limit. Passive methods to detect intoxication so that drinkers can be warned proactively, are desirable. Many young people take selfportraits (selfies) while drinking. We explore whether user intoxication levels can be inferred by image analysis and classification of selfies. We analyzed a corpus of the facial images of 53 subjects after drinking 0-3 glasses of wine, extracted features from the photographs and used machine learning to classify subjects as either sober or drunk. We found that facial lines changed significantly after consuming alcohol and that facial landmark vectors were the most predictive features. We achieved a classification accuracy of 81% using Gradient Boosted Machines for classifying subjects as either "sober" (0 or 1 glasses of wine) or "non-sober" (2 or 3 glasses of wine). Augmenting the original dataset of studio images by blurring, rotating, and altering lighting in order to capture more realistic party/bar scenarios, also improved classification accuracy. We used our intoxication classifiers to build DrunkSelfie, an Android application that estimates the subject's drunkenness from a selfie.