结合位姿约束与轨迹寻优的人体姿态估计

发布时间：2018-06-07 12:45

本文选题：人体姿态估计 + 混合部件模型　；参考：《光学精密工程》2017年04期

【摘要】：基于混合部件模型的人体姿态估计方法忽视了人体结构的对称位姿约束关系,从而导致对称部件容易被重复检测、人体姿态估计准确率较低,为此,提出一种基于位姿约束与轨迹寻优的姿态估计新方法。首先估计人体单部件和对称部件在单帧图像中的多个合理位置,利用对称部件之间的位姿约束关系构建标识部件。然后根据单部件和标识部件各自的目标优化函数,通过动态规划算法反复迭代获得初始轨迹候选集,再结合轨迹的全局特征剔除检测得分较低的运动轨迹。最后引入树形合约模型,联系时空上下文信息,准确求解出视频序列光滑且兼容的最优轨迹。在N-best、Outdoor Pose和Scene数据集中的实验结果表明,对于存在背景复杂、运动模糊、部件遮挡等问题的视频序列中,该方法平均姿态估计准确率达87%以上,有效减少了对称部件的误判,提高了视频中人体姿态估计的准确率。
[Abstract]:The method of human body attitude estimation based on hybrid component model neglects the symmetrical pose constraint relation of human body structure, which leads to the symmetry component being easily detected repeatedly, and the accuracy of human body attitude estimation is low. A new attitude estimation method based on pose constraint and trajectory optimization is proposed. Firstly, the reasonable position of human body single component and symmetrical component in a single frame image is estimated, and the identification components are constructed by using the pose constraint relationship between symmetric components. Then according to the objective optimization function of the single component and the marking component, the initial trajectory candidate set is iterated through dynamic programming algorithm, and then the motion track with low score is eliminated by combining the global feature of the trajectory. Finally, a tree contract model is introduced to accurately solve the smooth and compatible optimal trajectory of video sequences with temporal and spatial context information. The experimental results in N-best outdoor Pose and Scene data sets show that for video sequences with complex background, motion blur and component occlusion, the average attitude estimation accuracy of this method is over 87%, which effectively reduces the misjudgment of symmetric components. The accuracy of human pose estimation in video is improved.
【作者单位】：河海大学物联网工程学院;常州市传感网与环境感知重点实验室;
【基金】：国家自然科学基金(41301448) 江苏省重点研发计划(BE2016071)
【分类号】：TP391.41

【相似文献】