Distracted Driving Detection by Combining ViT and CNN

The risk of road accidents is rising rapidly. Distracted driving remains one of the leading causes of traffic accidents. Therefore, the identifying of the distracted driving become significant. Extensive methods based on the convolutional neural network (CNN) have been applied to the detection of the distracted driving. Within Convolutional Neural Network (CNN), the convolution operations are good at extracting local features but experience difficulty to capture global representations. Within visual transformer (ViT), the cascaded self-attention modules perform surpassingly in capturing content-based global interactions but unfortunately deteriorate local feature details. In order to address those challenges mentioned before, we propose a new distracted driving detection method that utilizes the driver and related object cues as guidance and combines CNN with ViT as a backbone to capture the local and global features. Besides, the simulation module is introduced to obtain the result of classification during a certain time period in the stage of inference. Under the widely used StateFarm benchmark, our proposed method presents the best performance.