摘要
为实现对常见分心驾驶行为(打电话、抽烟、喝水)进行高精度的实时检测,提出基于改进YOLOv4-tiny的分心驾驶行为检测算法。首先针对香烟这类小目标检测精度差的问题,将主干特征提取网络中最后一个跨阶段局部连接(Cross Stage Partial connections,CSP)层的输出特征进行卷积和上采样之后与第二个CSP层的输出特征进行特征融合,增加了一个52×52的预测尺度,提升对小目标的检测能力;其次在特征金字塔中添加高效通道注意力机制(Efficient Channel Attention,ECA)模块,提升模型的检测精度;最后使用k-means聚类算法在自制分心驾驶行为数据集上重新确定先验框,使用迁移学习、余弦退火学习率和标签平滑进行模型训练。结果表明,本文方法的所有类别平均精度(mAP)为98.88%,相较于原YOLOv4-tiny算法提高了2.83%,其中对于检测香烟的平均精度值提高了8.46%,提升了对香烟的检测能力。改进YOLOv4-tiny的分心驾驶行为检测算法具有较好的综合性能,有利于车载系统对驾驶员的分心驾驶行为进行实时检测并提醒,对减少交通安全事故具有一定的现实意义。
To achieve high-precision real-time detection of common distracted driving behaviors(such as calling,smoking,and drinking water),a distracted driving behavior detection algorithm based on improved YOLOv4-tiny has been proposed.First of all,to solve the problem of poor detection accuracy of small targets such as cigarettes,the output features of the last Cross Stage Partial connections(CSP)layer in the backbone feature extraction network are convolved and upsampled.After which the feature fusion is performed with the output features of the second CSP layer,and a prediction scale of 52×52 is added to improve the detection ability of small targets.Secondly,the Efficient Channel Attention(ECA)attention mechanism module is added to the feature pyramid to improve the detection accuracy of the model.Finally,the k-means clustering algorithm is used to redetermine bounding box priors on the self-made distracted driving behavior data set,and transfer learning,cosine annealing learning rate,and label smoothing are used for model training.The results show that the mean average precision(mAP)of the method proposed in the present study is 98.88%,which is 2.83%higher than the original YOLOv4-tiny algorithm,and the average precision for cigarette detection is increased by 8.46%,which improves the detection ability of cigarettes.Therefore,the distracted driving behavior detection algorithm based on improved YOLOv4-tiny show good comprehensive performance,which is conducive to the on-board system to detect and remind the driver’s distracted driving behavior in real-time and has certain practical significance for reducing the occurrence of traffic safety accidents.
作者
魏启康
朱文忠
江嘉文
谢鑫煌
WEI Qikang;ZHU Wenzhong;JIANG Jiawen;XIE Xinhuang(School of Computer Science and Engineering,Sichuan University of Science&Engineering,Yibin 644000,China)
出处
《四川轻化工大学学报（自然科学版）》
CAS
2023年第2期67-76,共10页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
基金
四川省科技创新(苗子工程)培育项目(2022049)
企业信息化与物联网测控技术四川省高校重点实验室开放基金(2020WZJ02)。
