Improving Human Action Recognition in Australian Rules Football through Video Data Augmentation

For video-based computer vision tasks and deep learning more generally, it is essential to have large amounts of correctly annotated data for model training and evaluation. This is especially the case in team-based sports, where the rapid and concurrent movement of many players is recorded by multiple cameras from different perspectives. Australian Rules Football (AFL) is one such sport where capturing specific human actions is important. Previous work [1] demonstrated how a range of human AFL actions could be automatically classified using a relatively small annotated data set comprising 439 video clippings of AFL games. These included pass, mark, contested mark, kick and other non-action events. In this work, we explore different video data augmentation approaches and assess their impact on the accuracy of the classifications. We achieve an improvement of up to $6.66 \%$ and overall accuracy of $84.76 \%$ for the AFL-related human actions of interest.