Quaternion lifting scheme applied to the classification of motion data

Abstract In this study, a new method of classification of skeleton-based motion data has been introduced. In the first stage, we performed multiscale feature extraction of rotational data. It is based on the proposed linear quaternion lifting scheme, with respect to the rotations coded by unit quaternions, which computes each scale based on the spherical linear interpolation (SLERP) prediction function and preserves the average signal value on each scale. Consequently, motion descriptors are extracted as quaternion attributes on different scales. The final recognition is performed by the nearest neighbor and minimum distance classifiers, adapted to support nonscalar features. Because of dimensionality of obtained descriptors, an attribute selection with respect to the multiresolution data has been proposed. It takes into consideration a specified number of resolutions, which is similar to low-pass filtering of the frequency domain. This method is utilized to solve the gait-based human identification problem. To validate such an application, a database containing data from 30 subjects was collected at the Human Motion Laboratory of the Polish-Japanese Academy of Information Technology (PJAIT). The obtained results were found to be satisfactory. In the best case, over 96% precision with only seven misclassified gaits of 178 samples was achieved.