Estimating missing marker positions using low dimensional Kalman smoothing

Motion capture is frequently used for studies in biomechanics, and has proved particularly useful in understanding human motion. Unfortunately, motion capture approaches often fail when markers are occluded or missing and a mechanism by which the position of missing markers can be estimated is highly desirable. Of particular interest is the problem of estimating missing marker positions when no prior knowledge of marker placement is known. Existing approaches to marker completion in this scenario can be broadly divided into tracking approaches using dynamical modelling, and low rank matrix completion. This paper shows that these approaches can be combined to provide a marker completion algorithm that not only outperforms its respective components, but also solves the problem of incremental position error typically associated with tracking approaches.