201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

论文标题:Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

来源/作者机构情况:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

卧龙岗大学(世界排名230~),第一次听说这个学校。竟然是在澳大利亚的一个学校。好吧,华人果然全球了

李老师是本硕都是浙大的,李老师个人链接如下:

https://www.uow.edu.au/~wanqing/#UOWActionDatasets

解决问题/主要思想贡献:

使用一个加权协方差因子,来积累前几帧的信息,使用增强学习来实现online learning,可以不用使用分好段的视频来预测动作

成果/优点:

1.延时,错误率和丢失率都有很多提升

缺点:

反思改进/灵感:

#############################################################

论文主要内容与关键点:

1.Introduction

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

前人研究的主要分类方法,缺点是没有办法实时检测

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

视频动作的特征表现,主要依靠这两种

2.Related Work

主要介绍了一下,上面两种分类方法,主要的几个研究方法,讲了一下这些的缺点。

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

特别强调了一个苏联人的一个方法,并讲解自己的文字解决了他的两个问题:没有权重的对待不同帧

3.The proposed method


权重方差因子:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

时间权重的变化:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

帧权重的变化:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

这里,能量是结点的动量和势能。

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

增量学习:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

后面还给予了证明。

4.Experimental Results

一些度量性能。还有结果展示

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

后面使用了一些去噪和归一化

动作特征向量的选取:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

使用KNN和SVM进行特征分类

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

展示了在三个数据集上面的结果:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

最后一种数据集,是卧龙岗大学自己创造的,适合增量学习。

5The depth and skeleton data will be made available after the paper being accepted for publication at:

http://www.uow.edu.au/˜wanqing/#UOWActionDatasets

https://www.uow.edu.au/~wanqing/#UOWActionDatasets

最后讲解了使用的硬件情况:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

以及对比,KNN和SVM的情况:

201904Online Human Action Recognition Based on Incremental Learning of Weighted Covariance Descriptors

5.Conclusion

6.附录,

一些公式的证明和推导

代码实现:

上一篇:[CF Round #294 div2] D. A and B and Interesting Substrings 【Map】


下一篇:泛函编程(38)-泛函Stream IO:IO Process in action