Person identity recognition on motion capture data using label propagation

Most activity-based person identity recognition methods operate on video data. Moreover, the vast majority of these methods focus on gait recognition. Obviously, recognition of a subject's identity using only gait imposes limitations to the applicability of the corresponding methods whereas a method capable of recognizing the subject's identity from various activities would be much more widely applicable. In this paper, a new method for activity-based identity recognition operating on motion capture data, that can recognize the subject's identity from a variety of activities is proposed. The method combines an existing approach for feature extraction from motion capture sequences with a label propagation algorithm for classification. The method and its variants (including a novel one, that takes advantage of the fact that, in certain cases, both activity and person identity labels might exist for the labeled sequences) have been tested in two different datasets. Experimental analysis proves that the proposed approach provides very good person identity recognition results, surpassing those obtained by two other methods.