基于深度卷积-Tokens降维优化视觉Transformer的分心驾驶行为实时检测

Real-Time Detection of Distracted Driving Behavior Based on Deep Convolution-Tokens Dimensionality Reduction Optimized Visual Transformer Model

针对基于端到端深度卷积神经网络的驾驶行为检测模型缺乏全局特征提取能力以及视觉Transformer(vision transformer,ViT)模型不擅长捕捉底层特征和模型参数量较大的问题,本文提出一种基于深度卷积和Tokens降维的ViT模型用于驾驶人分心驾驶行为实时检测,并通过开展与其他模型的对比试验、所提模型的消融试验和模型注意力区域的可视化试验充分验证了所提模型的优越性。本文所提模型的平均分类准确率和精确率分别为96.93%和96.95%,模型参数量为21.22 M,基于真实车辆平台在线推理速度为23.32 fps,表明所提模型能够实现实时分心驾驶行为检测。研究结果有利于人机共驾系统的控制策略制定和分心预警。

To address the problems that the end-to-end Deep Convolutional Neural Network(DCNN)based driving behavior detection model lacks global feature extraction ability,and the Vision Transformer(ViT)model is not good at capturing underlying features with a large number of model parameters,this paper proposes a ViT model that combines deep convolution and Tokens downscaled optimization for real-time detection of driver dis⁃traction behavior.Comparison experiments with other models,ablation experiments and visualization experiments of the models’attention region are carried out to fully validate the superiority of the proposed model.The mean accura⁃cy and precision of the proposed model are 96.93%and 96.95%,respectively.The number of the model parameters is 21.22 M;and the online inference speed based on the real vehicle platform is 23.32 fps,indicating that the pro⁃posed model can achieve real-time distracted behavior detection.The result of the study is beneficial to the control strategy development and distraction warning of human-machine co-driving system.