(A) study on the gesture spotting from continuous hand motion with a threshold model

Abstract

This paper proposes a new method of spotting meaningful gestures from continuous hand motion in real-time. As the gesturer switches from one gesture to another, his hand makes an intermediate move linking the two gestures. A gesture recognizer may attempt to recognize this inevitable intermediate motion as a meaningful one (segmentation problem). Furthermore, the same gesture varies dynamically in shape and duration; instance by instance even of the same gesturer (spatio-temporal variability problem). The proposed method can solve the problems of segmentation and spatio-temporal variability of gestures using Hidden Markov Model (HMM). To handle non-gesture patterns between gestures, we make use of the internal segmentation property of the HMM and introduce a threshold model that consists of the state copies of all trained gesture models. The internal segmentation property implies that the states and the transitions of a trained HMM represent sub-patterns of a gesture and their sequential order. The threshold model calculates the likelihood threshold of the input pattern and is used to qualify an input pattern as a gesture. However, the threshold model, a weak model of all gestures, has a large number of states, so we reduce the number of states by the use of relative entropy. Through a set of experiments, it has been shown that the proposed method can successfully extract meaningful gestures from continuous hand motion with 93.14% reliability. The proposed method has been incorporated into PowerGesture to provide gestural commands for browsing the slide of $PowerPoint^TM$.