【摘要】 人类行为识别作为视频分类中的重要问题,正成为计算机视觉中的热门话题。由于视频信息较多,有的视频冗余信息过量,判别性帧较少,因此如何无监督地提取关键帧对于行为识别至关重要。为此,本文提出了一种新的基于顺序验证的关键帧提取方法,并将其应用到行为识别中。首先,本文定义了一种顺序验证的模块,验证局部区间中帧的顺序,学习局部区间中帧的关键性描述,接着将其整合得到整段视频中每一帧的关键性描述;其次,根据学习到的视频帧关键性描述提取关键帧;最后通过实验讨论分析提取多少关键帧对行为识别最有利。实验结果表明,本文的方法在UCF-101上可以达到95.40%,在HMDB51上可以达到68.80%,均优于当前的一些先进的方法。更多还原

【Abstract】 As an important issue in video classification,human action recognition is becoming a hot topic in computer vision.Since there are many video information,some videos have redundant information and few discriminative frames,so how to extract key frames unsupervised is very important for action recognition.To this end,the paper proposes a new key frame extraction method based on order verification and apply it to action recognition.First,this paper defines an order verification module that verifies the order of frames in a local interval,learns the key description of the frames in the local interval,and then integrates them to obtain the key description of each frame in the entire video; Second,key frames are extracted based on the learned key descriptions of the video frames; Finally,the paper discusses experimentally how many key frames are extracted to be most beneficial for action recognition.Experimental results show that the proposed method can reach 95.40% on UCF-101 and 68.80% on HMDB51,which are all better than some current advanced methods.更多还原