摘要
现有基于轮廓图的步态识别方法易受服装等外部条件干扰,而基于3D模型的识别方法虽然一定程度上抵抗了外部干扰,但对摄像设备有额外的要求,且模型计算复杂。针对上述问题,利用3D姿态估计技术,建立了行人运动的"轻"模型,利用神经网络框架,提取行人3D空间运动的时空信息,并且与轮廓图的信息相融合,进一步丰富了步态特征。在CASIA-B的数据集上的实验结果表明:融合了3D时空运动信息增强了步态特征的鲁棒性,进一步提升了识别率。
The existing gait recognition methods based on silhouettes are easy to be interfered by clothing and other external conditions. Although the 3 D model-based recognition method resists external interference to a certain extent,it has additional requirements of camera equipment and complicated model calculations. In order to solve the above problems, " light" model for pedestrian motion is established by using 3 D pose estimation technology.The spatiotemporal information of pedestrian 3 D spatial motion is extracted by using neural network framework,and the information is fused with the information of skeleton map to further enrich the gait features. The results on CASIA-B dataset show that the fusion of 3 D spatiotemporal motion information enhances the robustness of gait features and further improves recognition rate.
作者
赵黎明
张荣
张超越
ZHAO Liming;ZHANG Rong;ZHANG Chaoyue(College of Information Science and Engineering,Ningbo University,Ningbo 315211,China)
出处
《传感器与微系统》
CSCD
北大核心
2021年第2期23-25,29,共4页
Transducer and Microsystem Technologies
基金
浙江省公益性技术研究项目(LGF18F020007,LGF21F020008)
宁波市自然科学基金资助项目(2018A610057,2018A610163)。
关键词
深度学习
步态识别
3D姿态
时空特征融合
deep learning
gait recognition
3D pose
spatiotemporal feature fusion
