基于改进Mobile Net-SSD网络的驾驶员分心行为检测  

Detecting Driver’s Distracted Behavior Based on Improved Mobile Net-SSD Network

作　　者：杜虓龙 余华平[1] DU Xiao-long;YU Hua-ping(School of Computer Science,Yangtze University,Jingzhou Hubei 434000,China)

机构地区：[1]长江大学计算机科学学院,湖北荆州434000

出　　处：《公路交通科技》2022年第3期160-166,共7页Journal of Highway and Transportation Research and Development

基　　金：国家自然科学基金项目(61440023)。

摘　　要：驾驶员在驾驶时出现分心行为容易诱发交通事故。为了能对驾驶员在驾驶时的分心行为进行快速、准确识别,从而降低交通事故的发生率,保障人民生命财产安全。首先,将数据集进行分析评估,筛选出7种驾驶员分心状态,并制作成Voc2017格式数据集;其次,通过深度可分离卷积替换SSD(VGG16)网络中特征提取层的方法减少网络参数量,形成Mobile Net-SSD网络模型,使模型使用场景更适合车内检测,并在浅层网络中加入HDC处理模块改进网络的特征提取层,在网络中利用该模块提高特征提取能力,有效应对浅层网络特征提取时的特征丢失现象。然后对改进后的特征提取层进行网络叠加处理,使其可以进行多尺度融合提取特征,使网络鲁棒性提升,增强网络对驾驶员行为检测的性能,构成新的MH-SSD检测网络模型,随后,使用迁移学习方法对改进后的网络进行训练。最后,使用测试集和自制的短视频对改进后的网络进行测试评估,再通过对照组进一步说明改进后网络优势。结果表明,改进后的网络mAP值达到94.01%,较Mobile Net-SSD网络模型高2%,网络参数量为SSD(VGG16)的1/2,网络检测实时帧数保持在25 fps以上,改进后的网络可以实时,准确地识别7种分心行为。MH-SSD网络可实现驾驶员分心行为实时检测,为下一步研究打下了良好基础。The driver’s distracted behavior while driving can easily induce traffic accidents. In order to be able to quickly and accurately identify the driver’s distracted behavior while driving, thereby reduce the incidence of traffic accidents and ensuring the safety of people’s lives and property, first, the data set is analyzed and evaluated, and 7 driver distraction states are screened out and made into a Voc2017 format data set. Second, the feature extraction layer in the SSD(VGG16) network is replaced with deep separable convolution to reduce network parameters, and the Mobile Net-SSD network model is formed to make the model usage scene more suitable for in-vehicle detection, the HDC processing module is added to the shallow network to improve the feature extraction layer of the network, and the feature extraction ability is improved by using the module in the network to effectively cope with the feature loss phenomenon during shallow network feature extraction. Then, the network overlay processing on the improved feature extraction layer is performed, so that it can perform multi-scale fusion extraction of features, improve the robustness of the network, and enhance the performance of the network to detect driver’s behavior and form a new MH-SSD detection network model. Afterwards, the improved network is trained by using the transfer learning method. Finally, the improved network is tested and evaluated by using the test set and self-made short video, and the advantages of the improved network is illustrated by using the control group. The result shows that(1) the mAP value of the improved network reaches 94.01%, which is 2% higher than that of the Mobile Net-SSD network model;(2) the amount of network parameters is 1/2 of that of SSD(VGG16), the real-time frame rate of network detection remains above 25 fps, and the improved network can accurately identify 7 kinds of distraction in real time. The MH-SSD network can realize real-time detection of driver’s distracted behavior, laying a good foundation f

关 键 词：交通安全 分心行为检测 计算机视觉 卷积神经网络 驾驶员行为 目标检测 

分 类 号：U492.84[交通运输工程—交通运输规划与管理]