Fast identification of a human skeleton-marker model for motion capture system using stochastic gradient descent method

Abstract

A method to identify a human model as a rigid kinematic chain and marker arrangement, which is referred to as a human skeleton-marker model in this paper, from the data collected by an optical motion capture system is proposed. The model is utilized for convenient and accurate motion analyses based on robotics computations. A chicken-and-egg problem to find the model to estimate the whole-body posture and the whole-body posture to fit the model simultaneously is resolved by a dual-phase nonlinear least square error minimization. The computation cost mainly due to the numerical computation of the gradient of the cost function with the inverse kinematics is significantly reduced by applying the stochastic gradient descent (SGD) method. Despite only using partial samplings for the estimation of the global gradient, it does not sacrifice the accuracy since it stochastically reflects information of the overall motion sequence through the iterations.