This paper describes an image processing approach capable for estimating the pose of athletes exercising on indoor rowing machines in video sequences. The proposed algorithm finds and tracks the hand, elbow, shoulder, ankle, knee, hip and head, and the line of the back also. Our contribution is twofold. The first is a new background subtraction method, which can reliable separate the silhouette of athletes under some assumptions related to the videos. Furthermore the paper introduces - as the second contribution - a skeleton fitting method to find the joints of the athletes based on the results of the background subtraction. This algorithm is based on anthropometric data and special movement patterns. The overall solution works on a real time setting in the test environment. Comparing the results it is presented that our method surpasses the most accurate state-of-the-art general pose estimation solution for indoor rowing specific videos based on two common used metrics as well.
© Copyright 2018 Journal of Computing and Information Technology. University of Zagreb. Alle Rechte vorbehalten.