基于压缩域新时空特征的视频动作识别

Action recognition for new spatiotemporal feature based on compressed video

针对基于压缩域视频动作识别中精度偏低等问题,提出了一种压缩域视频动作识别新方法.主要思路是在提取压缩码流信息阶段,利用压缩视频的运动矢量和残差构造新时空特征.新时空特征不仅具有运动矢量和残差的时空关系,更兼备物体边缘清晰的特点.通过在动作识别主流数据集(HMDB-51、UCF-101)的验证,此方法计算开销相比基于传统像素域的动作识别更小,识别精度相比基于视频压缩域的动作识别更高.实验表明:基于压缩域的新时空特征具有了强时空关系和高信息密度等优点,能使视频动作识别的结果更加准确.

Aiming at the problem of low accuracy in video action recognition based on compressed domain,a new method of video action recognition based on compressed domain is proposed.The main idea is to use the motion vectors and residuals of compressed video to construct new spatiotemporal features in the stage of extracting compressed bitstream information.The new spatiotemporal features not only have the spatiotemporal relationship of motion vectors and residuals,but also have the characteristics of clear edges.Through the verification in the mainstream action recognition data sets(HMDB-51,UCF-101),the computational overhead of this method is smaller than that based on the traditional pixel domain,and the recognition accuracy is higher than that based on the video compression domain.Experiments show that the new spatiotemporal features based on compressed domain have the advantages of strong spatiotemporal relationship and high information density,which can make the results of video action recognition more accurate.