Action recognition in sports videos using stacked auto encoder and hog3d features

Abstract

Sports analytics is an emerging area of research with applications to personalized training and entertainment. In this work, an approach for sports action recognition using stacked autoencoder and HOG3D features is presented. We demonstrate that actions in sports videos can be recognized by 2D interpretation of HOG3D features, extracted from the bounding box of the player as input to a deep learning model. The ability of a stacked autoencoder to learn the underlying global patterns associated with each action is used to recognize human actions. We demonstrate the efficacy of the proposed classification system for action recognition on ACASVA dataset.