基于无监督学习的动作捕捉躯干弯曲度估计方法

Unsupervised Learning-based Method for Motion Capture Torso Curvature Estimation

人体姿态估计是行为识别的研究热点,基于深度学习的人体运动捕捉技术是人体姿态估计的重要方法。然而,基于骨骼模型的研究,通常使用二维的人体姿态估计,在人体中间部位缺少胸部、骨盆、脊柱等关键点,大部分方法只包含人体中间部分有限的关键点。由于人体整体结构的复杂性,跟踪方法只估计人体表面,估计躯干内部的弯曲度较困难。通过在基于骨架的模型中添加新的关键点来优化现有的深度学习模型,并提出一种基于无标记动作骨架的曲线弯曲算法来估计躯干的弯曲度。借助惯性测量智能套装,用惯性测量法对该方法进行验证,该方法能够较好地估计出人体躯干弯曲度。实验表明,无标记的躯干弯曲估计模型,为进一步提高人体估计姿态的躯干弯曲精度提供新的研究思路。

Studies based on skeletal models usually use two-dimensional human pose estimation,which lacks crucial points such as the chest,pelvis,and spine in the middle part of the human body,and most methods only include a limited number of vital points in the central part of the human body.Due to the complexity of the overall body structure,tracking methods only estimate the body's surface and have more difficulty estimating the curvature within the torso.Therefore,this paper aims to optimize an existing deep-learning model by adding new key points to the skeleton-based model and to propose a curve-bending algorithm based on the unmarked action skeleton to estimate the curvature of the torso.The method is validated using inertial measurements with an inertial measurement smart suit.This way can evaluate the human torso curvature well.