Machine Learning for Real Time Poses Classification Using Kinect Skeleton Data

Poses recognition is an important research topic because some situations require silent communication (sign language, surgeon poses to the nurse for assistance etc.). Traditionally, poses recognition requires high quality expensive cameras and complicated computer vision algorithms. This is not the case thanks to the Microsoft Kinect sensor which provides an inexpensive and easy way for real time user interaction. In this paper, we proposed a real time human poses classification technique, by using skeleton data provided by the Kinect sensor. Different users performed a set of tasks from a vocabulary of eighteen poses. From skeleton data of each pose, twenty features are extracted so that they are invariant with respect to the user's size and its position in the scene. We then compared the generalization performances of four machine learning algorithms, support vectors machines (SVM), artificial neural networks (ANN), k-nearest neighbors (KNN) and Bayes classifier (BC). The method used in this work shows that SVM outperforms the other algorithms.