PUB - Publikationen an der Universität Bielefeld

The recognition of manual actions, i.e., hand movements, hand postures and gestures, plays an important role in human-computer interaction, while belonging to a category of particularly difficult tasks. Using a Vicon system to capture 3D spatial data, we investigate the recognition of manual actions in tasks such as pouring a cup of milk and writing into a book. We propose recognizing sequences in multidimensional time-series by first learning a smooth quantization of the data, and then using a variant of dynamic time warping to recognize short sequences of prototypical motions in a long unknown sequence. An experimental analysis validates our approach. Short manual actions are successfully recognized and the approach is shown to be spatially invariant. We also show that the approach speeds up processing while not decreasing recognition performance.