Adaptive kernels for skeleton based action recognition using geometric feature score fusion

Abstract

This paper presents a novel adaptive kernel based method for classifying the human actions from the skeletal data. Three types of geometric features: joint positions, joint relative distances and joint relative angles were used for action recognition using the information sensed by Microsoft Kinect sensors. Recently, kernel-based methods mostly focused on the action recognition on spatio-temporal data. In this work, the obtained scores from the individual adaptive kernels which are inputted three different geometric feature namely joint positions, joint relative distances and joint relative angles features were fused to detect the action by using simple kernel matching technique which evaluates the similarity between the features. This method is used to rank the actions from the database according to the highest ranked query action. The experimental results conducted over three publicly available datasets, i.e., NTU RGB-D, G3D and UTD-MHAD. The proposed technique has been tested and compared with other state-of-the-art methods on above datasets. The proposed method performed well on all datasets and good recognition rates were recorded.