基于BiViTNet的轻量级驾驶员分心行为检测方法

A Lightweight Driver Distraction Behavior Detection Method Based on BiViTNet

针对基于卷积神经网络的驾驶员分心行为检测,模型比较复杂、检测效率低下且缺少全局视觉表征的问题,提出了一种双分支并行双向交互神经网络BiViTNet(bidirectional interaction neural network based on vision transformer)对驾驶员行为进行识别,将ViT(vision transformer)引入到网络中对全局信息进行编码,在一定程度上提高检测精度。该网络由两个并行分支组成,第1个分支基于轻量级的CNN结构,第2个分支基于ViT结构。通过双向特征交互模块BiFIM(bidirectional feature interaction module)解决CNN Branch和ViT Branch之间特征不对称的问题,最后将两个分支的特征融合并对驾驶员行为进行检测。实验在自建的多视角驾驶员数据集上展开,验证集准确率达到97.18%,参数量为38.22 MB,计算量为271.20×10^(6)。研究表明:轻量级BiViTNet提高了驾驶员分心行为识别的准确率,可以在一定程度上辅助驾驶员的行车安全。

To address the issues of complex models,low detection efficiency,and lack of global visual representation in driver distraction behavior detection based on convolutional neural networks,a bidirectional interaction neural network based on vision transformer(BiViTNet)was proposed to identify driver behavior.ViT(vision transformer)was introduced into the network to encode global information,which could improve the detection accuracy to a certain extent.The proposed network consisted of two parallel branches,and the first one was based on the lightweight CNN structure and the second one was based on the ViT structure.The bidirectional feature interaction module(BiFIM)was used to solve the problem of feature asymmetry between CNN branch and ViT branch.Finally,the features of the two branches were fused and driver behaviors were detected.The experiment was carried out on the self-built multi-view driver dataset.The accuracy of the verification set reached 97.18%,the parameter quantity was 38.22 MB,and the MAdds was 271.20×10^(6).The research shows that lightweight BiViTNet improves the accuracy of drivers distracted behavior identification and can assist drivers driving safety to a certain extent.