基于双流全卷积网络的驾驶员姿态估计方法  

Driver pose estimation based on dual-stream fully convolutional network

作　　者：王彬[1,2] 赵作鹏 WANG Bin;ZHAO Zuopeng(School of Computer Science and Technology,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China;Department of Information Technology,Jiangsu Union Technical Institute,Xuzhou,Jiangsu 221008,China)

机构地区：[1]中国矿业大学计算机科学与技术学院,江苏徐州221116 [2]江苏联合职业技术学院信息技术系,江苏徐州221008

出　　处：《江苏大学学报（自然科学版）》2022年第2期161-168,共8页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61976217);徐州市重点研发计划项目(KC18082)。

摘　　要：针对现有姿态估计方法在驾驶室复杂环境条件下发生的非目标误检测和检测精度低的问题,提出了一种基于双流全卷积网络的驾驶员姿态估计方法.该方法通过建立2条独立的FCN(fully convolutional network)分支,分别对关键点坐标及关键点间的连接信息进行预测,同时在2个分支中构建沙漏状的网络结构,增强了网络提取关键信息的能力.为了进一步提高模型的特征提取能力,将浅层与深层网络得到的特征图进行融合.为了验证所提方法的检测效果,采用COCO(common objects in context)数据集和DDS(driver′s driving situation)数据集进行验证.试验结果表明:该方法在COCO数据集和DDS数据集上的检测平均精度分别达到64.5%和78.4%,优于其他3种对比算法;该方法可以提高驾驶员人体姿态的检测精度,具有较好的鲁棒性.To solve the problems of non-target misdetection and low accuracy under the complex environmental conditions in the cab by the existing pose estimation method,a driver pose estimation method was proposed based on dual-stream fully convolutional network(FCN).Two independent FCN branches were established to predict the coordinates of keypoints and the connection information between keypoints,and the hourglass network structure was set in the two branches to enhance the ability of network for extracting key information.In order to further improve the feature extraction capability of the network,the feature maps obtained from the shallow and deep networks were fused.Common objects in context(COCO)data set and driver′s driving situation(DDS)data set were used to verify the detection effect of the proposed method.The experimental results show that the detection accuracies of the proposed method in COCO data set and DDS data set are respective 64.5%and 78.4%,which illuminates that the proposed method is superior to other three comparison algorithms.The proposed method can improve the detection accuracy of the driver posture with good robustness.

关 键 词：驾驶员 姿态估计 特征融合 全卷积网络 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]