Adaptive segmentation for air gestures identification

Having a user-friendly Human-CAD interaction with high speed and accuracy plays a key role in development of future intelligent modeling environments. A major part of this puzzle is sketch identification using either 2D gestures -which is commonly recorded from mouse, light pen and touchpad-or air gestures captured from some newly emerged devices such as Leap Motion and Soft-Kinect. To this end, we present a leaning based technique for segmentation of air gestures. The proposed technique can detect the separation points of any single-stroke air gesture using specific motion features such as speed, curvature and center of curvature. Two types of separation points are considered: 1) rough separation points or simply corner points and 2) soft separation points such as inflection points. The segmentation is performed in two steps: Support Vector Machine (SVM) is used to adaptively differentiate the corner points from regular points. A soft segmentation method is then implemented to further break the rough segments into a set of smaller arcs and lines based on sudden change in the center of curvature. The experimental validation shows robust performance of the proposed method and low computation expenses.