基于光流约束自编码器的动作识别
发布时间：2019-01-01 16:18
【摘要】：为了改进特征学习在提取目标运动方向及运动幅度等方面的能力,提高动作识别精度,提出一种基于光流约束自编码器的动作特征学习算法.该算法是一种基于单层正则化自编码器的无监督特征学习算法,使用神经网络重构视频像素并将对应的运动光流作为正则化项.该神经网络在学习动作外观信息的同时能够编码物体的运动信息,生成联合编码动作特征.在多个标准动作数据集上的实验结果表明,光流约束自编码器能有效提取目标的运动部分,增加动作特征的判别能力,在相同的动作识别框架下该算法超越了经典的单层动作特征学习算法.
[Abstract]:In order to improve the ability of feature learning to extract the moving direction and amplitude of the target, and to improve the accuracy of motion recognition, an action feature learning algorithm based on optical flow constraint self-encoder is proposed. The algorithm is an unsupervised feature learning algorithm based on single-layer regularized self-encoder. Neural network is used to reconstruct the video pixels and the corresponding moving optical flow is used as the regularization term. The neural network can encode the motion information of objects while learning action appearance information and generate joint coded action features. The experimental results on several standard action data sets show that the optical flow constrained self-encoder can effectively extract the moving parts of the target and increase the ability to distinguish the motion features. Under the same framework of motion recognition, this algorithm surpasses the classical single-layer action feature learning algorithm.
【作者单位】： 东南大学自动化学院;中国电科集团28所;
【基金】：国家自然科学基金资助项目(61402426)
【分类号】：TP181;TP391.41
[Abstract]:In order to improve the ability of feature learning to extract the moving direction and amplitude of the target, and to improve the accuracy of motion recognition, an action feature learning algorithm based on optical flow constraint self-encoder is proposed. The algorithm is an unsupervised feature learning algorithm based on single-layer regularized self-encoder. Neural network is used to reconstruct the video pixels and the corresponding moving optical flow is used as the regularization term. The neural network can encode the motion information of objects while learning action appearance information and generate joint coded action features. The experimental results on several standard action data sets show that the optical flow constrained self-encoder can effectively extract the moving parts of the target and increase the ability to distinguish the motion features. Under the same framework of motion recognition, this algorithm surpasses the classical single-layer action feature learning algorithm.
【作者单位】： 东南大学自动化学院;中国电科集团28所;
【基金】：国家自然科学基金资助项目(61402426)
【分类号】：TP181;TP391.41
【相似文献】
相关期刊论文 前10条
1 韩雷;王洪庆;林隐静;;光流法在强对流天气临近预报中的应用[J];北京大学学报(自然科学版);2008年05期
2 龚大墉;汪治华;杨数强;刘岩;;基于目标的局部约束光流分析[J];重庆工学院学报(自然科学版);2009年07期
3 严强;黄增喜;曹丽萍;黄蓉刚;;光流法在机车安全行驶中的应用[J];计算机应用研究;2013年04期
4 马鹏飞;杨金孝;;基于光流法的粒子图像测速[J];科学技术与工程;2012年32期
5 张永亮;卢焕章;贺兴华;谢耀华;;基于光流预测的直线对应算法[J];信号处理;2010年05期
6 孙承志;熊田忠;吉顺平;张家海;;基于差分的光流法在目标检测跟踪中的应用[J];机床与液压;2010年14期
7 王，
本文编号：2397804
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2397804.html
