Primitive actions extraction for a human hand by using SVD

The use of singular values for classification of elementary actions performed by human hands is known in the literature. However, usually only the first singular right-vector is used. This approach is well-suited when a single elementary action is performed, thus it is used for classification, as it reduces a large set of data to a single vector with as many entries as the number of features. However, when considering complex actions, the second singular value may increase its importance with respect to the first one. This suggests a strategy for segmentation of elementary actions based on the analysis of the second-to-first singular value ratio. The idea is first discussed on a simple 2D example and then tested on an experimental set-up employing a data-glove with 18 markers and 5 infra-red cameras, resulting into samples of 54 data (3D-data for each marker). Different maneuvers are executed, a flow of 600 samples (on the average) is stored and processed, resulting in segmentation of the complete action into a suitable number of elementary actions. What is worth noticing is that the procedure is fully data-driven, no prior human knowledge is required to produce segmentation, classification and motion interpretation. The results of experiments show the effectiveness of the proposed procedure, both in segmenting and in classifying and recognizing complex maneuvers.