摘要
近年来基于骨架数据集的行为识别已成为行为识别领域研究热点.以往基于骨架信息的行为识别研究,是从原始骨架数据中直接提取一系列行为特征信息,作为行为识别网络模型的输入,这种方法忽略了行为中骨架间存在的时空依赖关系,不利于提高行为识别准确度.本文直接利用骨架数据作为网络模型的输入,提出一种基于骨架数据的多流卷积神经网络的行为识别方法.该方法首先进行骨架动作建模;然后利用多流的卷积网络框架提取骨架行为时空特征信息,最后进行行为特征融合和识别.在NTU_RGB+D数据集上进行行为识别,实验结果表明此方法能够提高识别准确率.
In recent years,action recognition based on skeleton dataset has become a research hotspot in the field of action recognition.In the past,the action recognition research based on skeleton information directly extracted a series of behavioral feature information from the original skeleton data as the input of the behavior recognition network model.This method ignores the spatio-temporal dependence between the skeletons in the behavior,and is not conducive to improving the accuracy of behavior recognition.Using skeleton data as the input of network model,this paper proposes a multi-stream convolutional neural network action recognition method based on skeleton data.firstly,the method performs dynamic skeleton modeling.Then,the multi-stream convolutional network framework is used to extract the skeleton action spatio-temporal feature information,and finally the action feature fusion and recognition.The action recognition is carried out on the NTU_RGB+D dataset.The experimental results show that this method can improve the recognition accuracy.
作者
华钢
曹青峰
朱艾春
张赛
唐士宇
崔冉
HUA Gang;CAO Qing-feng;ZHU Ai-chun;ZHANG Sai;TANG Shi-yu;CUI Ran(School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221008,China;College of Computer Science and Technology,Nanjing University of Technology,Nanjing 211816,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第6期1286-1290,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(51574232)资助.
关键词
卷积神经网络
骨架行为识别
骨架动作建模
时空特征信息
特征融合
convolutional neural network
skeleton action recognition
skeleton action modeling
spatio-temporal information
feature fusion
