基于场景-部件的人体行为识别方法

Human Action Recognition Based on Scene-Part

中文关键词:  卷积神经网络  行为识别  场景  部件

英文关键词:convolutional neural network  action recognition  scene  part

中文摘要:

      基于部件的行为识别方法给图像行为识别领域提供了一种新的思路，即将人体行为识别看成是一种人体各个部件行为的组合。但是这种方法完全忽视了除人以外的任何东西，导致了某些姿态过于相似的行为无法区分。针对这一不足，在基于部件(Part-based)的行为识别方法基础上，提出了基于场景-部件(Scene-Part based)的行为识别方法。实验过程中利用卷积神经网络将部件和场景的外观特征转换为行为特征，并通过全连接层将所有特征连接，进行人体行为类别的最终判定。在Standford40和PASCAL VOC2012两种行为识别数据集上的实验结果表明，相对于基于部件的行为识别方法而言，基于场景-部件的行为识别方法能更好地区分相似行为，从而进一步提高行为识别的准确率，提升精度约为1%。

英文摘要:

      The part-based action recognition method provides a new idea for the field of image-based action recognition,which means that human action recognition can be regarded as a combination of the actions of various human body parts.While this method completely ignores other things except humans,and sometimes it’s too weak to distinguish extremely similar behaviors.In order to distinguish these similar behaviors better,scene-part based action network for action recognition is proposed on the basis of part-based.During the experiment,the convolutional neural network was used to transform the appearance features of the parts and scenes into behavior features,and the human behavior was classified by connecting all the features via the fully connected layer.The experimental results of the Standford40 and PASCAL VOC2012 datasets show that the proposed method has better performance,so the accuracy of behavior recognition is improved further by about 1%.