A Real Time Temporal Segmentation Method for Continuous Gestures Analysis

The forthcoming demand of the natural use of the human hand as human–computer interaction motivates research on continuous hand gesture recognition. Gesture recognition relies on gesture segmentation to find the boundary of ges tures with semantic meanings while ignoring unintentional movements. However, gesture segmentation on a stream of continuous gestures poses a challenge due to the movement ambiguity of successive gestures and unconstrained spatiotemporal variation. To address this challenge, our approach entails three major steps: the first step applies Maximum Mean Discrepancy criterion to detect the changepoints over continuous gestures as the initial estimated cuts of the gesture transitions; the second step uses kinematic constraints to revise an initial estimated cut to an accurate gesture transition position; and finally a probability density estimation is used to estimate the hand motion between two cuts to eliminate unintentional movements and non-gesture segments. The proposed method is evaluated by using Chalearn benchmark datasets with 150 sequences of continuous sign gestures. The proposed method achieves the accuracy rate of 82.7 %, which outperforms the state-of-theart approach.