Dynamic hand gesture recognition for human-computer interactions

Abstract

In this paper we address the problem of dynamic trajectory segmentation for human-computer interfaces. We are concerned with locally linear trajectories. Trajectory points are obtained from a hand feature detector. First, a tensor voting technique is used to filter the trajectory and to construct a smooth trajectory from the sparse collection of detected points. The tensor voting scheme is also in accordance with perceptual principles. Local linearity of the trajectory permits us to have a decision based on an analysis of the corresponding modes in the Radon space. A mode detector in this space allows us to find the orientation of each trajectory segment. The entire trajectory is encoded by a sequence of directions, thus, allowing a large number of possible meaningful gestures to be defined in the HCI.