Real-Time Driver Behavior Detection Based on Deep Deformable Inverted Residual Network With an Attention Mechanism for Human-Vehicle Co-Driving System

Driver Distraction Behavior Detection Framework Based on the DWPose Model, Kalman Filtering, and Multi-Transformer

“…Considering that distraction behavior recognition is a finegrained image classification task, to improve the ability of the model to extract subtle features from images with small differences, Li, et al [16] guided the model to learn robust features based on the loss function of the contrast learning and stop-gradient strategies. In addition, the improvement of the classification performance of deep learning models for distraction behaviors by applying attention mechanisms and prior knowledge is also an important research idea [17][18][19][20]. Lu, et al [18] applied the attention channel to convolutional weights, and fused global and keypoint features from driving images of different scales.…”

“…(15) (16) The process of extracting the category token can be represented as (17) where CT represents the separation of the category token from tokens and TE represents the Transformer encoder.…”