A Self-updating Template Approach for Human Action Recognition Using Improved DTW ?

The existing recognition algorithms always need large number of samples to train models or templates. Nevertheless, they aren’t view-invariant and afiected by performing speed and scale variations of difierent actors. In this paper, we capture the skeleton information by Kinect and present a novel approach based on self-updating template for human action recognition using 4D Quaternions as a compact representation of postures. We adopt the new template generation method and update the template if it satisfles speciflc conditions. This method solves the above problems. On the other hand, we improve the decision function of DTW that the number of alignment paths may have a great in∞uence on recognition. We tested our approach on our dataset containing 960 action samples and the public MSR Action3D dataset. Compared with Li et al.’s [10], our approach produced better results on the most of cases, especially with the increased number of recognition.