Feature Extraction and Classification of Dynamic and Static Gestures Based on RealSense

The process of feature selection and the extraction algorithm of dynamic and static gestures based on 3D skeleton data were mainly introduced, and ten customized gestures were classified according to the features above. At first, ten gestures were customized according to teaching demands and with Intel’s 3D camera SR300, a small gesture database based on classroom teaching was established; then, the characteristics of these gestures were analyzed so that features like the trajectory and the local area of some joints point and the extremum of adaptive Euclidean distance can be selected, and a packaged feature selection method (LVW-RF) which combines Las Vegas Wrapper (LVW) and Relief-F was proposed to extract the feature set. With these features input into support vector machine (SVM), an average accuracy of 98.67% was acquired for the ten gestures. The experimental results prove that the features extracted are reasonable and effective.