驾驶员在开车时玩手机是一种危险驾驶行为，针对目前现有的驾驶员玩手机行为检测方法存在检测精度低以及检测速度慢等一系列的问题，提出了一种基于改进YOLOv5s(You Only Look Once v5s)的驾驶员玩手机行为检测方法。该方法将YOLOv5s作为基础网络，并在此基础上进行改进。首先，在网络中引入混合空洞卷积(Hybrid Dilated Convolution，HDC)，增大感受野，减小图像信息的丢失，并保证图像信息的连续性;加入ECA(Efficient Channel Attention)通道注意力模块，使得整个网络更加专注于提取显著特征。经过实验测试，改进后的方法与原始的YOLOv5s网络相比，平均精度均值(mean Average Precision，mAP)提高了3。7%，检测帧率提高了5帧/s，检测的精度与速度得到了有效的提升;证明了该方法可行有效，具有良好的工程应用前景。
Mobile Phone Behavior Detection for Drivers Based on Improved YOLOv5
It is a dangerous driving behavior for a driver to play with a mobile phone while driving.Aiming at a series of prob-lems such as low detection accuracy and slow detection speed in the existing detection methods of driver's mobile phone behavior,an improved YOLOv5s(You Only Look Once v5s)driver's mobile phone behavior detection method is proposed.This method uses YOLOv5s as the basic network and improves on it.First,Hybrid Dilated Convolution(HDC)is introduced into the network to in-crease the receptive field,reduce the loss of image information,and ensure the continuity of image information.ECA(Efficient Channel Attention)channel attention module is added to make the entire network more focused on extracting salient features.After experimental testing,compared with the original YOLOv5s network,the improved method has a mean Average Precision(mAP)in-crease of 3.7%,a detection frame rate of 5 frames/s,and an effective detection accuracy and speed.The results show that the im-proved method is feasible and effective,and has good engineering application prospects.
mobile phone behaviorYOLOv5sHybrid Dilated Convolutionchannel attentionfeature extraction
万方数据
维普
