>> 自然科学版期刊 >> 2016年06期 >> 正文
基于先验的动作视频关键帧提取
供稿: 庞亚俊 时间: 2018-11-12 次数:

作者:庞亚俊

作者单位:洛阳理工学院

摘要:针对运动视频关键帧提取结果运动表达能力差的问题,以健美操运动视频关键帧提取为例,将先验语义引入到视频片段分割和关键帧提取特征提取等过程中,提出基于先验的运动视频关键帧提取算法。该算法采用韵律特征和动作节拍连续性等先验知识,将健美操动作视频分解成不同长度的动作视频片段,并利用Hog人体分类器从每一帧图像中识别出人体边界框;通过人体模板将人体边界框分割为16个运动块,并采用光流法计算每个运动块的基本运动方向;通过比较运动块基本运动方向的差异实现了动作视频关键帧提取。实验证明,该方法在保证关键帧视频压缩的情况下,具有更好地动作概括力。

基金:河南省教育厅科学技术研究项目(14B890004); 河南省省教育厅自然科学研究项目(2011A520031); 河南省软科学研究计划项目(142400410459); 河南省社科规划项目(2015BTY001);

关键词:运动视频; 关键帧提取; 光流图; 节拍探测;

DOI:10.16186/j.cnki.1673-9787.2016.06.019

分类号:TP391.41

Abstract:To enhance the motion express ability of key frame,the key frame extraction of calisthenics video as an example,the prior is applied to the video split and key frame feature extraction and an algorithm of key frames extraction of motion video binding prior knowledge is proposed. Based on music beat detection algorithms and rhythm constraints,the calisthenics video is broken down into continuous motion video clips. Then the bounding box of human is identified from the calisthenics video picture with Ho G human classifier. And then,each bounding box of human is divided into 16 motion block by body template and the optical flow is adopted to set the basic direction of motion of each block. Finally,key frames of calisthenics video are gotten by comparing the differences of the basic direction of motion of each block. Experiments show that the proposed algorithm has better motion generalization ability while keeping motion video compression efficiency.

最近更新