Autolabeling 3D tracks using neural networks

Motion capturing systems based on monochrome video have problems assigning measured 3D marker positions to the anatomically defined positions or labels of the markers applied to the test subject. This task is usually called “labelling” and is paramount to the reconstruction of 3D trajectories from a set of video frames from multiple cameras––the tracking procedure. Labelling means sorting a set of 3D vectors by their spatial positions. Neural networks can be made to “learn” from examples of marker positions in a given marker set, i.e. previously manually tracked video sequences. Trained neural networks are able to calculate a set of sorted approximate marker positions from an unsorted set of exact marker positions. The set of sorted exact positions can be found by pairing up both sets of marker positions via a minimum distance function. The neural network is trained only once and can then be applied to any number of individuals. The algorithm is designed for cyclic motions like for locomotion analysis.