| 基于人体姿态估计的分心驾驶行为检测 |
| |
| 引用本文： | 尹智帅,钟恕,聂琳真,马晨.基于人体姿态估计的分心驾驶行为检测[J].中国公路学报,2022,35(6):312-323. |
| |
| 作者姓名： | 尹智帅 钟恕 聂琳真 马晨 |
| |
| 作者单位： | 武汉理工大学 汽车工程学院, 湖北 武汉 430000 |
| |
| 基金项目： | 国家重点研发计划项目(2018YFB0105203)；国家自然科学基金项目(51805388) |
| |
| 摘 要： | 为弥补现有驾驶特征提取方法的不足,提高分心驾驶行为检测的准确性和鲁棒性,将2D/3D人体姿态估计应用于驾驶人行为检测,提出一种适用于驾驶舱环境下的驾驶特征提取方法。首先通过将2D姿态估计网络Simple Baseline和分类网络ResNet进行融合,构建基于2D姿态估计的分心驾驶行为检测模型,并在分心驾驶数据集State Farm上分析不同数据增强方法、不同超参数、不同分类网络对模型性能的影响。其次,融合3D密集姿态估计网络DensePose与分类网络ResNet,构建基于3D姿态估计的分心驾驶行为检测模型。接着,在State Farm数据集上,针对模型的实时性和泛化能力,对比分析基于原始图像和基于2D/3D姿态的分心驾驶行为检测模型。最后,针对效果更优的基于2D姿态估计的分心驾驶行为检测模型,在分心驾驶数据集State Farm上,对使用不同姿态估计算法和分类网络的分心驾驶行为检测模型做了交叉试验,对比分析4个不同检测模型的优缺点。进一步地,将基于2D姿态估计的分心驾驶行为检测模型应用于实际采集的驾驶图片,对模型的泛化能力和有效性进行了测试验证。研究结果表明:与基于原始图像的检测模型相比,基于2D和3D姿态的检测模型都能显著提高分心驾驶行为的检测准确率;基于3D姿态的检测模型在检测精度方面略优,但基于2D姿态的检测实时性更好,检测效率是基于3D姿态检测的4倍;在驾驶舱单一环境下,基于2D姿态估计的分心驾驶行为检测模型能够满足分心驾驶行为检测的需求,在分心驾驶行为检测方面具有重要应用价值。
|
| 关 键 词： | 汽车工程 驾驶分心 分心驾驶行为检测 人体姿态估计 人机共驾 |
| 收稿时间： | 2020-06-03 |
Distracted Driving Behavior Detection Based on Human Pose Estimation |
| |
| YIN Zhi-shuai,ZHONG Shu,NIE Lin-zhen,MA Chen.Distracted Driving Behavior Detection Based on Human Pose Estimation[J].China Journal of Highway and Transport,2022,35(6):312-323. |
| |
| Authors: | YIN Zhi-shuai ZHONG Shu NIE Lin-zhen MA Chen |
| |
| Affiliation: | School of Automotive Engineering, Wuhan University of Technology, Wuhan 430000, Hubei, China |
| |
| Abstract: | In this study, we applied the 2D/3D human pose estimation for detecting the driver distraction behavior to compensate for the shortcomings of the existing driving feature extraction methods and improve the detection accuracy and robustness. In particular, a driving feature extraction method suitable for cockpit environment was proposed. First, by integrating 2D pose estimation network Simple Baseline into the classification network ResNet, a driver's distraction behavior detection model based on 2D pose estimation was constructed. Moreover, the influence of different data enhancement methods, hyper parameters, and classification networks on the performance was analyzed on the driving distraction dataset State Farm to select the optimal network structure. Subsequently, the 3D dense pose estimation network DensePose and the classification network ResNet were combined to construct a driving distraction behavior detection model based on 3D pose estimation. Then, aiming at the real-time and generalization ability of the model, the driving distraction behavior detection model based on the original image and that based on 2D/3D pose were compared and analyzed on the State Farm dataset. Finally, aiming to obtain a superior distraction behavior detection model based on 2D pose estimation, a cross-over study using different pose estimation algorithms and classification networks was performed. The advantages and disadvantages of the four combinations thus obtained were analyzed. Furthermore, to examine the generalization ability and effectiveness of our proposed model, the detector was tested on real-life driving scenes collected by us. The research results show that compared to the driving distraction behavior detection model based on the original image, the one based on 2D/3D pose can provide significantly higher detection accuracy. The detection accuracies of the two models are similar. The 3D pose-based model performs slightly better than the 2D pose-based model. In contrast, the 2D pose-based model is relatively superior in terms of detection rate, and runs at a frame rate four times that of the 3D pose-based model. In a single cockpit environment, the 2D pose-based detection model can satisfactorily detect driving distraction behavior and has important application value in driving distraction behavior detection. |
| |
| Keywords: | automotive engineering driving distraction distracted driving behavior distraction human pose estimation co-piloting |
|
| 点击此处可从《中国公路学报》浏览原始摘要信息 |
|
点击此处可从《中国公路学报》下载全文 |
|
