姿态估计和跟踪结合的运动视频关键帧提取

提出姿态估计和特定部位跟踪相结合的动作视频关键帧提取算法.首先利用非确定人体部位的时间连续性保持提高基于柔性部件铰接人体模型的单帧图像人体姿态估计准确率,通过实施数据降维得到局部拓扑结构表达能力强的判别性运动特征向量,采用极值判定原理确定候选关键帧集合.然后利用ISODATA动态聚类算法,通过初始聚类中心优化、基于语义的关键帧集合增强等策略确定关键帧.实验表明文中算法具有较高的关键帧提取准确率和召回率,支持基于语义的关键帧提取.提取的视频关键帧可以用于运动视频压缩和批注审阅.

Joint Pose Estimation and Tracking for Keyframe Extraction of Motion Video

A key frame extraction framework for motion video is proposed by using pose estimation and dynamic clustering techniques. The pose configuration of body part is computed with a mixture-of-parts articulated human model in single frame, whose accuracy is promoted by using temporal feature preserving of special body-part. The body-part motion is modeled with a set of discriminative motion vectors that has a good power of locality preserving and is received by performing a feature selection based on Laplacian Scoring on the relative distances and motion directions of each body part. A candidate keyframe set is filter out in the extremum principle. The extracted keyframes are attained by using an ISODATA-based clustering with the initial cluster centers optimization and semantic-based enhancement of the candidate keyframe set. Experimental data suggest that the proposed algorithm can experience a high accuracy and recall rate, and can facilitate the semantic-based keyframe extraction. Key frames extracted can be used to sports video compress and annotation.