%0 Journal Article %T 基于层叠的部件轨迹片段模型的视频人体姿态估计<br>Cascaded tracklet-based spatio-temporal model for video pose estimation %A 史青宣 %A 王谦 %A 田学东< %A br> %A SHI Qingxuan %A WANG Qian %A TIAN Xuedong %J 山东大学学报（工学版） %D 2018 %R 10.6040/j.issn.1672-3961.0.2017.431 %X 摘要： 为解决单目视频中的人体姿态估计问题,从人体的部件模型出发,以人体部件轨迹片段为实体构建时空概率图模型,通过逐步缩减轨迹片段在时域上的覆盖度,形成多级层叠模型,采用迭代的时域和空域交替解析的策略,从完整轨迹的推理开始,逐级过滤状态空间,直至获取人体各部件在每帧图像中的最优状态。为提供高质量的状态候选,引入全局运动信息,将单帧图像中人体姿态检测结果传播到整个视频形成轨迹,构成原始状态空间。在3个数据集上的对比试验表明,该方法较其他视频人体姿态估计方法达到了更高的估计精度。<br>Abstract: To address the problem of full body human pose estimation in video, a coarse-to-fine cascade of spatio-temporal models was developed in which the tracklet of body part was considered as basic unit. The notion of “tracklet” ranges from trajectory covering the whole video to body part in one frame. In this cascade, coarse models filtered the state space for the next level via their max-marginals. Loops in the graphical models made the inference intractable, the models were decomposed into Markov random fields and hidden Markov models. Through iterative spatial and temporal parsing, optimal solution was achieved in polynomial time. To generate reliable state hypotheses, the pose detections were propagated to whole video sequence through global motion cues. Our model was applied on three publicly available datasets and showed remarkable quantitative and qualitative improvements over the state-of-the-art approaches %K 轨迹片段 %K 姿态估计 %K 马尔科夫随机场 %K 隐马尔科夫模型 %K < %K br> %K Markov random field %K tracklet %K pose estimation %K hidden Markov model %U http://gxbwk.njournal.sdu.edu.cn/CN/10.6040/j.issn.1672-3961.0.2017.431