融合神经网络与数值计算的人体逆向运动学求解

MNN-IK:Mixing neural network and numerical IK for human posing

人体逆向运动学问题是人体运动合成、人体运动捕获和理解的基本问题.由于人体关节链式系统的复杂性,人体逆向运动学方程往往存在多解或无解的情形.传统的方法通常采用解析或数值迭代方法求解逆向运动学问题,在给定足够多约束的情形下能够得到比较好的解,但无法处理少量约束下生成自然的人体姿态问题.近年来,从大规模数据集中学习统计模型参数的思想被广泛运用,求解人体逆向运动学的机器学习方法中经典工作|混合Gauss逆向运动求解模型(Gaussian mixture model-inverse kinematics,GMM-IK)就提出利用混合Gauss模型建模人体姿态数据分布,并采用期望最大化方法求解参数.随着深度学习技术的发展,本文提出一种自编码神经网络与数值迭代融合的方法,在给定少量约束的情形下依然能够得到自然的人体姿态,相较于GMM-IK方法,本文所提出的方法通过神经网络自动学习姿态分布,省去了模型的假设和特征的设计,且量化实验显示本文方法的关节坐标和角度重建误差相较于GMM-IK模型平均减少了25%和39%.在应用方面,本文方法可处理光学运动捕获数据,也可用于图像视频的人体姿态估计等领域.

The inverse kinematics of human figure is the basic problem of human motion synthesis,human motion capture and understanding.The system of kinematic constraint equations has many or no solutions to the inverse kinematics for human posing because the human system is very complex and its articulated representation has many degrees of freedom.Traditional methods usually use analytical or numerical iterative methods to solve the inverse kinematics problem.They can obtain good results with sufficient constraints.However,it is difficult to find natural human pose only given a small number of constraints.In recent years,the idea of learning statistical models from large-scale data sets has been widely used.Among the machine learning methods for solving inverse kinematics of human body,the classical work GMM-IK proposes to use Gaussian mixture model to construct the human posture data distribution,and uses the expectation maximization method to solve the parameters.This paper presents a method combining neural networks and numerical inverse kinematics for human posing.It can synthesize natural human pose with a small number of constraints.Extensive quantitative experiments show that the joint coordinates and angle reconstruction errors of our method are reduced by an average of 25%and 39%,respectively,compared with the state of the art,i.e.,Gaussian mixture model.Our method can be used to deal with optical motion capture data,and estimate human pose in RGB image or video.