View independent human posture identification using Kinect
Yuanning Liu, Zequn Zhang, Ao Li, Minghui Wang · 2012
As an important part of human computer interaction (HCI) system, posture identification has been extensively studied over last years. Recently, Microsoft Kinect Sensor has become a hot spot for posture identification because it is efficient in acquiring body joint location information. In this study, based on Kinect, we proposed a framework for view independent human posture identification. In this framework, a viewpoint rotation transformation was performed on original skeleton location data and then total 9 features were extracted for building a SVM classifier. About 4200 samples including five postures taken from different body orientations were collected to construct a dataset for performance evaluation. The results of PCA analysis showed that the transformation was efficient in distinguishing different postures. Further analysis demonstrated that this method achieved a superior performance of 98.0% when the orientation angle was between -60° and 60°. These results show that this view independent framework is powerful and efficient in viewpoint invariant posture identification.