基于深度运动轨迹信息的人体动作描述算法

发布时间：2018-07-13 14:28

【摘要】：受到dense轨迹特征的启发,本文提出了基于深度运动轨迹信息的动作描述算法,首先,利用稠密光流场对L帧深度视频提取稠密(dense)兴趣点并形成稠密轨迹,其次,利用轨迹前后兴趣点的深度信息计算深度变化值,并将它加入到稠密轨迹和HOG描述算子的计算中;再次,在整个数据集上,计算所有动作的平均深度变化值并利用它判断每类动作的深度信息变化情况;最后,根据深度信息变化剧烈程度选择不同的码书,对视频样本进行投影并分类。在两个公开深度动作数据集DHA-17和UTkinect上进行了实验,实验结果表明基于深度运动轨迹信息的动作描述算法具有较好的区分性和鲁棒性,其性能与一些先进的且具有代表性的算法具有可比性。
[Abstract]:Inspired by the characteristics of dense trajectory, an action description algorithm based on depth trajectory information is proposed in this paper. Firstly, dense (dense) interest points are extracted by dense optical flow field to L-frame depth video, and then dense trajectories are formed. The depth variation is calculated by using the depth information of the point of interest before and after the trajectory, and it is added to the calculation of dense trajectory and HOG description operator. Thirdly, on the whole data set, The average depth variation value of all actions is calculated and used to judge the variation of depth information of each action. Finally, different codebooks are selected according to the intensity of the depth information changes, and the video samples are projected and classified. Experiments on two open depth action datasets, DHA-17 and UTkinect, show that the motion description algorithm based on depth motion trajectory information has good distinction and robustness. Its performance is comparable with some advanced and representative algorithms.
【作者单位】：天津理工大学计算机视觉与系统教育部重点实验室;天津理工大学天津市智能计算及软件新技术重点实验室;
【基金】：国家自然科学基金(61572357,61202168) 天津市应用基础与前沿技术研究计划(14JCZDJC31700) 天津市自然科学基金(13JCQNJC0040)资助项目
【分类号】：TP391.41

【参考文献】