Quaternion Watershed Transform in Segmentation of Motion Capture Data

The novel approach for segmentation of motion capture data is proposed. It utilizes hierarchical watershed transform for time series containing unit quaternions which describe rotations of human skeleton joints as well as their angular velocities. To approximate gradient magnitudes of subsequent time instants, aggregated geodesic distance on hypersphere (Formula presented) for preceding and following quaternions is computed. The introduced segmentation method was applied to gait analysis. The highly precise motion capture data, registered in a human motion lab (HML), were used. A fully automatic watershed transform with detection of catchment basins as well as a marker controlled one were investigated. The obtained results are promising. By selecting proper hierarchy of a segmentation or by specifying adequate markers, it is even feasible to divide gait cycle into consecutive steps. The segmentation can be improved in respect to a considered problem, if only selected joints are taken into account by watershed transform.