基于改进视觉自注意力模型的分心驾驶行为识别研究

Study on recognizing distracted driving behavior based on an improved visual self-attention model

针对分心驾驶行为识别问题,文章提出一种基于改进视觉自注意力模型的方法,构建了模型ViT_CR,用于估计驾驶员头部姿势,通过多任务学习提高角度预测精度,在数据集AFLW上预测误差MAE为4.61;运用ViT_CR处理连续视频帧,并基于分心驾驶识别原则设定安全阈值与辅助参数判断驾驶员是否处于分心状态。实验表明,在真实驾驶数据集Dimags上,该方法能有效利用头部姿势的时序信息进行识别,为分心驾驶监测及预警提供了一种新的思路。

To address the issue of recognizing distracted driving behavior,this study proposes a method that utilizes an improved visual self-attention model.The ViT_CR model is first constructed to estimate the driver’s head pose.Multi-task learning is employed to improve the accuracy of angle prediction,resulting in a prediction error MAE of 4.61 on the dataset AFLW.Subsequently,ViT_CR is used to process continuous video frames.Safety thresholds and auxiliary parameters are set based on the distracted driving recognition principle to determine whether the driver is in a distracted state or not.The experiments demonstrate that the method can effectively utilize temporal information of head pose for recognition on the real driving dataset Dimags.This provides a new idea for monitoring and warning against distracted driving.