Identifying Distracted and Drowsy Drivers Using Naturalistic Driving Data

Driver fatigue and distraction remain significant safety issues for drivers. Despite substantial developments in driver state detection technology, a reliable system has yet to emerge. Existing systems tend to suffer from reliance on a single metric such as PERCLOS estimated from single expensive in-vehicle cameras and/or a poorly designed and tuned algorithm resulting in lack of effectiveness (high false positive rates). It is not likely that any single, real-time measure of driver drowsiness will be obtainable all of the time from the entire driver population. Therefore, a multi-variable algorithm based on sensors/variables that can be reliably obtained in real time on modern vehicles is essential. In this work, several algorithms for multivariate time-series analysis are tested on the Second Strategic Highway Research Program (SHRP2) Naturalistic Driving Study (NDS) dataset, including a statistical feature extraction method, deep learning-based long short-term memory, and video classification using convolutional neural networks. Given the amount of training and test data currently available, traditional statistical feature extraction methods outperformed the deep learning methods tested.