Toward on-line fitting of a human skeleton-marker model for accurate motion tracking

Two approaches to on-line fitting of the human kinematics model and marker arrangement are proposed for accurate pose estimation in motion tracking. A ‘flexible’ skeleton-marker model that can adjust the lengths of the body segments and relative locations of markers in addition to the whole joint angles is employed. In order to avoid the ill-posedness due to the huge degrees of freedom and the overfitting, the model and the whole joint angles are updated frame-wise. The particle filter and a frame-wise gradient descent method were examined for the model update. The former was applied based on an expectation that a stochastic technique can help to avoid the overfitting of the model. The latter had a success in the authors’ another work to estimate the global gradient to reduce the estimation error only from one sample. While the both techniques worked from the viewpoint of accuracy, it was found that the particle filter had a drawback in the aspect of computation cost, so that the frame-wise gradient descent method is more preferable although it still has a problem in a high-rate implementation.