An Embedded Deep Learning Computer Vision Method for Driver Distraction Detection

Driver distraction is a modern issue when operating automotive vehicles. It can lead to impaired driving and potential accidents. Detecting driver distraction most often relies on analyzing a photo or video of the driver being distracted. This involves complex deep learning models which often can only be ran on computers too powerful and expensive to implement into automobiles. This paper presents a method of detecting driver distraction using computer vision methods within an embedded environment. By taking the deep learning architecture SqueezeNet, which is optimized for embedded deployment, and benchmarking it on a Jetson Nano embedded computer, this paper demonstrates a viable method of detecting driver distraction in real time. The method shown here involves making slight modifications to SqueezeNet to be trained on the AUC Distracted Driver Dataset, yielding accuracies as high as 93% when detecting distracted driving.