Driver Distraction Recognition using 3D Convolutional Neural Networks

The number of road accidents and deaths due to distracted driving is continuously increasing in recent years. Usage of mobile phones, talking to passengers and reaching behind to take something while driving are some of the reasons due to which drivers may get distracted. Distractions are of numerous types, out of which we focus on manual distractions which are based on the posture of the driver. We have considered two datasets that incorporate drivers engaging in several different distracting behaviors using left and/or right hands seen from a camera. In this research, we have derived a benefit from temporal information by using a 3D convolutional neural network and optical flow to improve the driver distraction monitoring task. Results have shown that fine-tuning a pertained network on Kinetics dataset for learning driver action achieves a detection accuracy about 90% on State Farm dataset, which outperforms other methods on the same dataset.