基于运动特征的骨骼行为识别方法
Human skeleton activity recognition based on motion feature
孙浩 1何宏 1汪焰兵 1朱子豪1
作者信息
- 1. 上海理工大学健康科学与工程学院,上海 200093
- 折叠
摘要
针对现有的骨骼行为识别方法对人体行为的运动信息利用不足的问题,提出一种基于运动特征的时空注意力图卷积(STA-GCN)行为识别模型.对动作捕捉设备采集到的关节点运动轨迹和速度信息进行建模,在时间和空间构建注意力权重矩阵,结合图卷积网络进行特征提取,能够关注到具有判别力的关节点和时间帧.通过在自建动作捕捉数据集和NTU-RGB+D数据集的CS和CV标准上进行实验,其结果表明,该模型增强了对人体骨骼行为信息的理解能力,验证了模型对行为识别的有效性.
Abstract
Aiming at the problem that the existing skeleton-based activity recognition algorithms can not make full advantage of motion information of human activity,a spatial-temporal attention graph convolutional neural network(STA-GCN)based on motion feature was proposed to model the trajectory and velocity collected by motion capture system.The spatial and temporal weight matrix was constructed and it was combined with graph convolutional network for feature extraction to focus the discrimi-native joints and frames.Experimental results on dataset established using activity information collected by motion capture sys-tem and NTU-RGB+D dataset's both CS and CV evaluation criterion show that the model enhances the ability to understand the information of human skeleton activity,and its effectiveness is verified for activity recognition.
关键词
行为识别/深度学习/动作捕捉/骨骼信息/特征提取/图卷积/时空注意力Key words
activity recognition/deep learning/motion cap/skeleton data/feature extraction/graph convolution/spatial-tempo-ral attention引用本文复制引用
基金项目
科技部项目(G2021013008)
科技部项目(2020YFC2005802)
上海市科委项目(18070503000)
上海理工大学医工交叉重点基金(1020308405)
上海理工大学医工交叉重点基金(1022308502)
出版年
2024
NETL
NSTL
万方数据