A convolutional neural network (CNN) has been designed to interpret player actions in ice hockey video. The hourglass network is employed as the base to generate player pose estimation and layers are added to this network to produce action recognition. As such, the unified architecture is referred to as action recognition hourglass network, or ARHN. ARHN has three components. The first component is the latent pose estimator, the second transforms latent features to a common frame of reference, and the third performs action recognition. Since no benchmark dataset for pose estimation or action recognition is available for hockey players, we generate such an annotated dataset. Experimental results show action recognition accuracy of 65% for four types of actions in hockey. When similar poses are merged to three and two classes, the accuracy rate increases to 71% and 78%, proving the efficacy of the methodology for automated action recognition in hockey.
© Copyright 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE. Veröffentlicht von IEEE. Alle Rechte vorbehalten.