视频动作识别研究

发布时间：2018-03-08 05:12

本文选题：动作识别　切入点：视频分析　出处：《江西理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：视频中的人体动作识别是一个非常活跃的研究领域,随着相机、手机等电子产品行业的快速发展,对基于视频中人体动作识别的应用提出越来越高的要求。针对人体动作在视频中的定位问题,如何对视频中提取的多种特征进行有效融合的问题以及如何利用动作标签信息提高分类效果等问题,提出了利用流形度量学习的人体动作识别方法。首先根据人体区域利用基于人物肢体伸展程度分析的方法,获取人体区域的面积变化函数。由于面积变化函数随时间不断变化的过程中会产生相应的噪点,为了使得面积函数体现出本质的波动特征,在获取面积变化函数之后对面积函数使用稳健的局部加权平滑方法对面积函数进行平滑。取面积函数的极小值作为动作的切分点对动作进行切分,将后续的动作识别对象具体化。其次从每一段动作片段中分别提取人体区域的时域全局特征、空域特征、帧间光流特征以及帧内局部旋度特征和散度特征,将这些特征构造成为一种7×7的协方差矩阵描述子将多种特征进行融合,在黎曼流形中对动作进行描述。最后在训练阶段结合流形度量学习方法,根据训练样本的标签信息有监督地寻找一种在流形空间中更有效地度量方法,提高同类间的聚合度,加大不同类别之间的差异,从而达到提高动作分类的效果。在实验阶段,对weizmann公共视频库的切分实验统计结果表明本文提出的视频切分方法具有很好的切分能力,能够做好动作识别前的预处理;在weizmann公共视频数据集上进行了流形度量学习前后的识别效果对比,结果表明利用流形度量学习方法对动作识别效果提升2.8%;在weizmann和KTH两个公共视频数据集上的平均识别率分别为95.6%和92.3%,与现有方法的比较结果表明本文提出的动作识别方法有更好的识别效果。多次实验结果表明本文算法在预处理过程中动作切分效果理想,描述动作所构造协方差矩阵对动作的表达有良好的多特征融合能力,流形度量学习方法对动作识别的准确性有明显提高。
[Abstract]:Human motion recognition in video is a very active research field. With the rapid development of camera, mobile phone and other electronic products, Put forward higher and higher requirements for the application of human motion recognition based on video. How to effectively fuse the features extracted from video and how to use the action tag information to improve the classification effect, etc. In this paper, a method of human motion recognition based on manifold metric learning is proposed. Firstly, according to the human body region, the method based on the extension degree of human body is used. Get the area change function of human body. Because the area change function will produce corresponding noise in the process of changing with time, in order to make the area function reflect the essential fluctuation characteristic, After obtaining the area change function, the area function is smoothed by a robust locally weighted smoothing method. The minimum value of the area function is taken as the segmentation point of the action to segment the action. Secondly, the global feature, spatial feature, inter-frame optical flow feature, local curl feature and divergence feature of human body are extracted from each action segment. These features are constructed into a 7 脳 7 covariance matrix descriptor, which combines multiple features to describe actions in Riemannian manifolds. According to the label information of the training sample, we can find a more effective measure method in manifold space, improve the aggregation degree of the same kind, increase the difference between different categories, so as to achieve the effect of improving the classification of action. The experimental results of weizmann common video library show that the proposed video segmentation method has a good segmentation capability and can do a good job of pre-processing before motion recognition. The recognition results of manifold metric before and after learning are compared on weizmann common video data set. The results show that the performance of motion recognition is improved by using manifold metric learning method, and the average recognition rate on two common video data sets of weizmann and KTH is 95.6% and 92.3 respectively. The results of comparison with existing methods show that the motion recognition proposed in this paper is based on this method. The results of many experiments show that the proposed algorithm is effective in action segmentation in the process of preprocessing. The covariance matrix constructed by describing the action has a good multi-feature fusion ability for the expression of the action, and the accuracy of the manifold metric learning method for motion recognition is obviously improved.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】