Machine learning-based skeleton extraction for sports pose estimation

Abstract

With the increasing interests in Augmented Reality (AR), the physical interactive entertainments are becoming popular. Among the entertainments related with AR, especially indoor screen sports are attracting people’s interests since they are not limited to the variations of weather, time, and region. In this paper, we propose a method to estimate human joint position in sports simulator using depth information. While most of the human pose estimation algorithms require the use of expensive sensors or high-performance GPUs, this study proposes a method that allows the achievement of accurate analysis even with lower priced sensor, a single depth camera. In this paper, an SVM (Support Vector Machine)-based human pose estimation algorithm was used with a single depth camera, which performs well even with a CPU without additional calibration. The proposed SVM-based joint estimation algorithm using depth information has been experimentally verified in a screen horse riding simulator environment.