A Nonintrusive System for Detecting Drunk Drivers in Modern Vehicles

Abstract

In this work, a nonintrusive system has been developed using features from inertial sensors, car telemetry, and road lane data, enabling to recognize the driving style of a drunk driver. Drunk drivers caused 10,497 deaths on USA roads in 2016 according to NHTSA. The Naturalistic Driver Behavior Dataset (NDBD) was created specifically for this work and it was used to test the proposed system. The proposed system was designed to study drunk driving situations, but it can also be used to detect any other psychoactive drugs consumption that causes abnormal driver behaviors during driving. The classifier system's output is "no risk" (normal driving) or "risk" (drunk/abnormal driving). If the system is connected to an autonomous or semi-autonomous car control system, it can be enabled to step in and act in order to avoid dangerous situations, or it can activate an alarm, or also ask for external help (e.g. contact authorities). The best results achieved in the experiments obtained 98% of accuracy in NDBD frames and only 1.5% of frames labeled in NDBD as "no risk" had a wrong prediction. The proposed system is composed by an MLP neural classifier using sigmoidal activation function and with 14 neurons in input layer, 18 neurons in hidden layer, and 1 neuron in output layer of the network. It uses periods of 220 frames (22 seconds) for the predictions and a buffer of the last 3 predictions was used for reducing the number of false predictions for "risk" output. Thus, it could avoid wrong predictions (false positives), avoiding to incorrectly enable the alarms and semi-autonomous car control system.