Video Augmentation for Enhanced Skill Learning in Karate Using Homography Transformation and Stacked Hourglass Networks

Abstract

This paper describes an image processing method that can facilitate skill learning in karate using recorded karate competition videos. The proposed method superimposes a partially filmed karate competition court in the input video image onto an overall model of a karate court via a homography transform. This method utilizes the Stacked Hourglass Network, a deep neural network proposed for estimating human poses, to estimate the corresponding points needed for the homography transform. To evaluate our method, a player-focused video was augmented with complete competition field information. The augmented video would be useful for observing both players’ actions as well as the player positioning within the entire competition court. The evaluation of the proposed method by a university karate club showed that it was useful for skill learning.