Sequence Alignment for RGB-D and Motion Capture Skeletons

RGB-D skeletons are nowadays commonly used e.g. for gesture recognition, and so their accuracy and stability have significant influence on further processing. Skeletons obtained with motion capture are considerably more accurate and can be used to assess the quality of RGB-D skeleton extraction algorithms. In this paper, we record motion sequences with both a Kinect RGB-D sensor and a full motion capture system and align the generated skeletons by subsequence dynamic time warping with a varied step size. To evaluate the alignment, we propose two measures: the minimum overall distance between feature vectors and the distance of transformed skeletons. Experimental results show that our proposed method provides a better alignment between skeletons than the comparison methods. The proposed technique can also be used for content-based retrieval from large motion capture databases.