ABSTRACT
Driving is a dynamic activity, which requires quick reflexes and decision making in order to respond to sudden changes in traffic conditions. Alcohol consumption impairs motor and cognitive skills, and causes many driving-related accidents annually. Passive methods of proactively detecting drivers who are too drunk to drive in order to notify them and prevent accidents, have recently been proposed. The effects of alcohol on a drinker's gait (walk) is a reliable indicator of their intoxication level. In this paper, we investigate detecting drinkers' intoxication levels from their gait by using neural networks to analyze sensor data gathered from their smartphone. Using data gathered from a large controlled alcohol study, we perform regression analysis using a Bi-directional Long Short Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN) architectures to predict a person's Blood Alcohol Concentration (BAC) from their smartphone's accelerometer and gyroscope data. We innovatively proposed a comprehensive suite of pre-processing techniques and model-specific extensions to vanilla CNN and bi-LSTM models, which are well thought out and adapted specifically for BAC estimation. Our Bi-LSTM architecture achieves an RMSE of 0.0167 and the CNN architecture achieves an RMSE of 0.0168, outperforming state-of-the-art intoxication detection models using Bayesian Regularized Multilayer Perceptrons (MLP) (RMSE of 0.017) and the Random Forest (RF), with hand-crafted features. Moreover, our models learn features from raw sensor data, obviating the need for hand-crafted features, which is time consuming. Moreover, they achieve lower variance across folds and are hence more generalizable.
ABSTRACT
Intoxicated driving causes 10,000 deaths annually. Smartphone sensing of user gait (walk) to identify intoxicated users in order to prevent drunk driving, have recently emerged. Such systems gather motion sensor (accelerometer and gyroscope) data from the users' smartphone as they walk and classify them using machine or deep learning. Standard Field Sobriety Tests (SFSTs) involve various types of walks designed to cause an intoxicated person to lose their balance. However, SFSTs were designed to make intoxication apparent to a trained law enforcement officer who manually proctors them. No prior work has explored which types of walk yields the most accurate results when assessed autonomously by a smartphone intoxicated gait assessment system. In this paper, we compare how accurately Long Short Term Memory (LSTM), Convolution Neural Network (CNN), Random Forest, Gradient Boosted Machines (GBM) and neural network classifiers are able to detect intoxication levels of drunk subjects who performed normal, walk-and-turn and standing on one foot SFST walks. We also compared the accuracy of intoxication detection on the ascending (increasing intoxication) vs descending (decreasing intoxication) limbs of drinking sessions (bi-phasic). We found smartphone intoxication sensing more accurate on the descending limb of the drinking episode and that intoxication detection on the normal walks of subjects were just as accurate as the SFSTs.
