Distracted driving detection based on the improved CenterNet with attention mechanism

Distracted driving detection has many significant application scenarios in intelligent transportation, driver assistance, and other fields. However, these distracted behaviors are difficult to be recognized due to the variable background and different scale targets. To solve these problems, a distracted driving detection scheme is proposed based on the improved CenterNet with attention mechanism in this paper. Given the complexity of driving environments, an image classification method was first designed to divide the images into person and unmanned areas, which can reduce the interference in unmanned situations. And then a novel attention mechanism module was introduced into CenterNet to improve its detection ability for small targets. Numerous experiments were conducted with a public dataset and newly built targeted dataset that included three categories of distracted driving behaviors with 6481 pictures. The results demonstrated that, the proposed scheme can detect distracted behaviors in real time while driving with a mean average precision (mAP) of 97.0%, which outperforms some representative detection methods, such as CornerNet, YOLO v3 and YOLO v4.