基于CNN的3D姿势估计在建筑工人行为分析中的应用  

Application of convolutional neural network-based 3D posture estimation in behavioral analysis of construction workers

作　　者：熊若鑫 宋元斌[1] 王宇轩 段彦娟[3] XIONG Ruoxin;SONG Yuanbin;WANG Yuxuan;DUAN Yanjuan(School of Naval Architecture,Ocean&Civil Engineering,Shanghai Jiaotong University,Shanghai 200240,China;School of Transportation,Southeast University,Nanjing Jiangsu 211189,China;Department of Dermatology,Seventh People's Hospital of Shanghai University of Traditional Chinese Medicine,Shanghai 200137,China)

机构地区：[1]上海交通大学船舶海洋与建筑工程学院,上海200240 [2]东南大学交通学院,江苏南京211189 [3]上海中医药大学附属第七人民医院皮肤科,上海200137

出　　处：《中国安全科学学报》2019年第7期64-69,共6页China Safety Science Journal

基　　金：国家自然科学基金资助(71271137);浦东新区中医药研发与创新专项(PDZYYFCX-201804)

摘　　要：为实现建筑工人现场行为的自动化分析,采用卷积神经网络(CNN)检测3D人体姿势并根据现场条件对连续图像进行姿态估计;考虑到动态和杂乱的施工现场环境(部分遮挡等)及多变的工人行为,开发建筑工人姿势图像数据集,从定性和定量2方面综合测试算法性能;将所提出的方法用于施工作业姿势风险评估,利用视频中工人的3D姿势驱动人体生物力学模型,快速、定量计算工人作业时易损伤的部位。结果表明:该人体姿势估计方法具有较好的鲁棒性和较高的准确性,结合生物力学模型可实现更精细的工人行为分析与评估。In order to facilitate an automated behavioral analysis of construction workers,CNN was applied for 3D human pose estimation on sequential images.Considering the complicated site environment and dynamic operator behaviors,the data set of construction workers'postures was developed and the performance of the proposed algorithm was analyzed from both qualitative and quantitative aspects.Furthermore,the derived 3D postures in the video were used to drive the biomechanical model for more detailed and quantitative analysis of worker behavior.The results show that the proposed 3D pose estimation method is accurate and robust,and that combined with biomechanical model,more detailed analysis and evaluation of workers'behavior can be achieved.

关 键 词：3D姿势估计 卷积神经网络(CNN) 行为分析 现场测试 施工安全 

分 类 号：X912.9[环境科学与工程—安全科学]