摘要
为解决人体行为识别中关节特征不显著、噪声干扰严重等问题,提出一种基于自适应融合权重的人体行为识别方法。为充分利用长时时序信息,需要对长视频进行时域分割,而后设计基于BN-Inception的双流网络结构,分别以RGB图像和骨骼关节点构成的运动特征向量作为空间流和时间流网络的输入,基于熵值法对各片段内单支流网络融合后,根据关节点显著程度赋予各支流以惩罚值,使双流网络自适应调节权重进行融合,输出行为类别。实验结果表明,该方法在人体行为识别中具有可行性和有效性。
To solve the problems of insignificant joint features and severe noise interference in human action recognition,a human action recognition method based on adaptive fusion weights was proposed.To make full use of the long-term temporal information,it was necessary to segment the long video in temporal domain.A two-stream network based on BN-Inception was built,and RGB images and the motion feature vectors composed of skeleton joint points were used as the input of the spatial stream and the temporal stream network.After fusing the single network in each segment based on the entropy method,the penalty value was assigned to the single network according to the significance of the joint points,so that the spatio-temporal network adaptively adjusted the weight for fusion to get the final action category.Experimental results show that the network model is feasible and effective in human action recognition.
作者
乔迤
曲毅
QIAO Yi;QU Yi(College of Information Engineering,Engineering University of PAP,Xi’an 710086,China)
出处
《计算机工程与设计》
北大核心
2023年第3期845-851,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(61801516)
陕西省自然科学基础研究计划基金项目(2019JQ-238)。
关键词
行为识别
双流卷积神经网络
骨骼关节点
运动特征向量
自适应融合
熵值法
时域分割
action recognition
two-stream convolutional neural network
skeleton joint points
motion feature matrix
adaptive fusion
entropy method
temporal domain segmentation
