A new technique for description and recognition of human gestures is presented. First, a method is proposed to represent gestures as an ordered sequence of simple basic motions. It is based on the rate of change of gesture motion direction (referred to as "curvature") that is estimated using optical flow from monocular motion images. Then, proceeding from the method, a real-time gesture recognition system is presented that can be easily given the gesture model and achieve high recognition rates for unspecific gesture performers. The main difference between the proposed method and previous ones is the fact that the authors' method does not require any preprocessing related to generation of the gesture model such as observation and/or analysis of specific performers' gestures. Finally, the real-time implementation of the proposed algorithm for manipulating a vehicle by gesture motion is described to demonstrate the utility of the proposed method.
展开▼
