Tracking athletes in diving videos based on pose learning

A fast algorithm is presented in this paper to track athletes in diving videos based on pose learning. In this algorithm, each diving motion is considered as a particular sequence of poses and each pose is represented by a set of landmarks. Firstly, the transformations between poses of successive frames are learned from sample videos. Then, in the video to be tracked, an initial pose is designated to the first frame and fits itself to the body contour by a curve-fitting algorithm. Later, in every following frame, the current pose transforms according to the parameters learned before, adjusts itself from the temporal information and fits itself to the body contour. In this way, the athlete in a diving video is effectively tracked. Experiments are carried out and the tracking results show that the proposed algorithm guarantees both efficiency and robustness.