An empirical study into the robustness of split covariance addition (SCA) for human motion tracking

Accurate human body tracking is extremely important for many virtual and augmented reality systems. However, tracking human motion is extremely difficult. Some of the difficulties arise from the fact that accurate process models of human motion are difficult to derive. Approximate models can have substantial time-correlated process noise terms. We examine the effectiveness of using the split covariance addition (SCA) algorithm as part of a human head orientation estimation system. We perform a series of empirical experiments to compare the performance of several implementations of SCA with an extended Kalman filter (EKF). The results suggest that the benefits of SCA are mixed. It leads to filters, which are slightly more robust and have slightly more accurate angular velocity estimates than the EKF. However, the absolute orientation estimate is slightly worse than the EKF.