Driver distraction detection via multi‐scale domain adaptation network

Distracted driving is the leading cause of road traffic accidents. It essential to monitor driver's status avoid accidents caused by distracted driving. Current research on detecting distracting behaviours focuses analysing image features using convolutional neural networks (CNNs). However, generalisation ability current models limited. This paper aims improve that are affected factors such as driver himself, background, monitoring angle, and so on. A new distraction detection method, which referred multi-scale domain adaptation network (MSDAN), was proposed model adaptability. The method consists three stages: first, convolution introduced build a backbone accommodate better valuable feature target different scales. Secondly, authors designed model's adaptability difference in data sources through adversarial training. Finally, dropout added fully connected layer increase ability. comparison results large-scale dataset show authors’ can accurately detect has good performance, with an accuracy improvement cross-driver cross-dataset experiments.