摘要
针对目前多人图像三维人体重建姿态失准、模型精确度低的问题,提出一种基于单张多人图像的三维人体重建算法,算法首先对图片进行特征提取生成多个人体框架得到人体区域图,接着对所得到人体图像的体型姿态参数进行进一步提取,将得到的人体参数与SMPL模型相结合后生成三维人体模型,再将三维投影至二维与图像的二维关节点进行比对利用卷积神经网络回归,最后利用生成对抗网络对模型进行判别与损失回归得到合适的生成器模型,完成整体设计。设计使用Human3.6M与MPII数据集进行测试,与cmr、smplify等算法进行对比实验,能够更准确地恢复三维人体模型,在提高了模型重建精确度的同时,模型的重建速度也得到提升。
To address the current problems of inaccurate pose and low model accuracy in 3D human reconstruction of multi-person images,a 3D human reconstruction algorithm based on a single multi-person image is proposed.The algorithm first generates multiple human frames by parameter extraction of image features,gets its 2D joint point pa-rameters after human positioning,and further extracts the body shape pose parameters of the obtained human images,combines the parameters with SMPL to generate The 3D human model is then projected to 2D and the 2D joints of the image are regressed using convolutional neural network,and finally the model is discriminated and loss regressed us-ing generative adversarial network to obtain a suitable generator model to complete the overall design.The design is tested using Human3.6Mwith MPII dataset and compared with cmr and smplify algorithms for experiments,which can recover the 3D human model more accurately and improve the model reconstruction accuracy while the model recon-struction speed is also improved.
作者
何恩德
张永爱
严群
林志贤
姚剑敏
HE En-de;ZHANG Yong-ai;YAN Qun;LIN Zhi-xian;YAO Jian-min(College of Physics and Information Engineering,Fuzhou University)
出处
《中国集成电路》
2022年第5期51-57,共7页
China lntegrated Circuit
基金
国家重点研发计划课题(2016YFB0401503)
广东省科技重大专项(2016B090906001)
福建省科技重大专项(2014HZ0003-1)
广东省光信息材料与技术重点实验室开放基金资助项目(2017B030301007)。
关键词
三维人体重建
smpl
生成对抗网络
多人人体检测
3D human reconstruction
smpl
generative adversarial network
multi-person human detection
