Incorporating bidirectional feature pyramid network and lightweight network: a YOLOv5-GBC distracted driving behavior detection model

Distracted driving is one of the leading causes of traffic accidents and has become a bottleneck for improving driver assistance technologies. It is still a challenge to detect distracted driving behavior in real-life scenarios, which have the features of complex backgrounds, different target scales, and resolutions. In this context, a lightweight YOLOv5-GBC model is proposed for real-time distracted driving detection in this work. Firstly, the lightweight network GhostConv is used to perform lightweight operations on the convolutional layers, aiming to reduce a large number of parameters and computations. Secondly, the path aggregation network structure is improved to enhance the model fusion ability for different scale features, and coordinated attention is introduced to enhance the model extraction ability for effective information. The proposed YOLOv5-GBC model can predict different types of distracted driving. Finally, this work conducts extensive experiments; the results show that the proposed model has a mean accuracy (mAP) of 91.8%, which is 3.9% better than the baseline model, with a reduction of 6.5% and 9.1% in the weight file and Floating-point Operations Per Second, respectively. It outperforms the models of Faster-RCNN, SSD, YOLOv3-tiny, and YOLOv4-tiny, which indicates that the proposed model can identify distracted driving behaviors efficiently and rapidly.