RBF models with shallow and deep feature for skeleton-based human gesture recognition

摘要

Recognition of human actions is an intelligent way for human-machine communication and Radial basis function (RBF) models are among the most powerful machines on this task. One prerequisite of using this traditional model is that the movement data must be translated into a vector space via the feature extraction process. Recent development of the convolutional neural networks (CNNs) has been shown that this deep architecture could achieve state-of-the-art performance in different recognition tasks. However, it is not possible to apply the CNN directly to many applications like 3D coordinates of human skeletal joints. In this paper, we present an effective way of reorganizing the 3D skeletal joints into feature maps, which enables applying the modern CNNs to recognize human gestures. Experimental results on the Microsoft's Kinect dataset MSRC-12 show that the proposed data re-organization scheme combined with the deep feature learning of CNN could achieve very competitive predictive accuracy.