We propose a new approach to gesture recognition using the properties of Spherical Self-Organizing Map (SSOM). Unbounded mapping of data onto a SSOM creates not only a powerful tool for visualization but also for modeling spatiotemporal information of gesture data. Once mapped onto a SSOM the gesture data is treated as a series of postures. A set of postures describing a specific path on the SSOM for a gesture is used as a trajectory. Although some trajectories may share the same postures, the path consisting of posture transitions will always be unique. Different variations of posture transitions occurring within a gesture trajectory are used to classify new unknown gestures. Experimental results on datasets involving full body and hand gestures show the effectiveness of our proposed method.
展开▼
