结合单目视觉成像和深度测量的行为检测研究  

Study on action detection by combination of monocular vision imaging and depth measure

作　　者：包林霞[1,2] 王云良 BAO Linxia;WANG Yunliang(School of Information Engineeing,ChangZhou Vocational Institute of Mechatronic Technology,Changzhou 213164,China;Jiangsu Internet of Things and Manufacturing Information Engineering Technology Research and Development Center,Changzhou 213000,China)

机构地区：[1]常州机电职业技术学院信息工程学院,江苏常州213164 [2]江苏省物联网与制造业信息化工程技术研究中心,江苏常州213000

出　　处：《光学技术》2021年第2期196-202,共7页Optical Technique

基　　金：常州市工业互联网数据智能技术重点实验室资助(CM20183002)。

摘　　要：人体行为自动检测技术易受成像角度、环境因素的影响,造成检测准确性下滑,结合单目视觉和深度传感器提出了一种新的行为检测技术。首先筛选出人体行为相关的关键帧,以解决行为速度不统一的问题;结合卷积神经网络特征和幅度直方图作为行为描述符;利用对抗网络实现人体行为的域适应迁移学习,解决单目视觉成像角度偏差的问题。在公开的行为检测数据集上完成了验证实验,结果显示该技术提高了行为检测的性能,并能有效改善单目视觉成像角度差异所导致的检测性能衰减问题。Body actions automatic detection techniques are usually influenced by imaging angle and imaging environment,these factors lead to detection accuracy reduction,in view of this,a new action detection technique combined monocular vision and depth sensor is proposed.Firstly,the key frames corresponding to body actions are filtered,this process helps to fix the problem of action speed difference;then,the convolutional neural network features and magnitude histogram are combined as action descriptors;finally,adversarial networks are adopted to realize domain adaption transfer learning for body action,in order to solve the problem of imaging anger difference of monocular vision.Validation experiments are carried on a public action detection dataset,and the results of which show that the proposed technique improves action detection performance,and the technique overcomes the detection performance reduction caused by monocular vision imaging anger difference,effectively.

关 键 词：自动行为检测 单目视觉成像 人工智能技术 深度神经网络 迁移学习 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]