Alcohol consumption detection through behavioural analysis using intelligent systems

• We instrumented a realistic driving simulator and monitored users behaviour.
• We generated features and associated those to various BAC values.
• We were able to classify drunk and sober example from 1 subject.
• With multiple subjects we could estimate the BAC.
• The system is generic and should adapt well to other problems.

We describe in this paper a new methodology for blood alcohol content (BAC) estimation of a subject. Rather than using external devices to determine the BAC value of a subject, we perform a behaviour analysis of this subject using intelligent systems. We monitor the user’s actions in an ordinary task and label those data to various measured BAC values. The obtained data-set is then used to train learning systems to detect alcoholic consumption and perform BAC estimation. We obtain good results on a mono-user base, and lower results with multiple users. We improve the results by combining multiple classifiers and regression algorithms.