Efficient Human Action Recognition with Dual-Action Neural Networks for Virtual Sports Training

In this paper, we propose an efficient human action recognition method for virtual sports training. In the proposed method, 3D skeleton data of human actions for four kinds of sports, which are composed of ready actions and main ones, are accurately captured by using RGB-D cameras. Then, skeleton features, which are sophisticatedly designed for each sport training, are extracted from captured 3D skeleton data. Finally, a dual-action neural network, which is developed to simultaneously recognize ready actions and main ones for sports training, is trained with extracted skeleton features. The effectiveness of the proposed method based on the dual-action neural network is verified with several real experiments for various sports.

1. Introduction

Human action recognition has been studied for a decade in the related literature, including computer vision, robotics, surveillance, virtual reality, and so on [1]–[2]. Although a lot of methods for human motion recognition have been proposed with various sensors and algorithms different to each other and some of methods have been successfully applied to real-life applications, human motion recognition is still one of challenging issues due to the very large diversity of general human actions. Therefore, it is technically very difficult to generalize one method based on one approach for recognizing all kinds of human actions.