Using deep learning to detect driver distraction in the Australian Naturalistic Driving Study (ANDS) video data – preliminary results
Keywords: Crash Reconstruction – including computer simulation, Statistical, Epidemiology and Other Road Safety Research Methods, Distraction & Inattention
Submission Date: 2019
This paper reports preliminary results of investigating the use of machine learning techniques to label distraction related events from video data collected from the Australian Naturalistic Driving Study (ANDS). This offline automatic labeling is designed to replace manual coding and accelerate the data reduction process with the view to save effort and money. We adopted the well-known pre-trained deep learning network Alex to label ANDS video data. The pre-trained network was used as a starting point after modifying the fully connected and classification layers. Then the modified model was retrained using ANDS data. The re-trained network achieved promising results despite low video quality.