Action Recognition in Australian Rules Football Through Deep Learning

Understanding player's actions and activities in sports is crucial to analyze player and team performance. Within Australian Rules football, such data is typically captured manually by multiple (paid) spectators working for sports data analytics companies. This data is augmented with data from GPS tracking devices in player clothing. This paper focuses on exploring the feasibility of action recognition in Australian rules football through deep learning and use of 3-dimensional Convolutional Neural Networks (3D CNNs). We identify several key actions that players perform: kick, pass, mark and contested mark, as well as non-action events such as images of the crowd or players running with the ball. We explore various state-of-the-art deep learning architectures and developed a custom data set containing over 500 video clips targeted specifically to Australian rules football. We fine-tune a variety of models and achieve a top-1 accuracy of 77.45% using R2+1D ResNet-152. We also consider team and player identification and tracking using You Only Look Once (YOLO) and Simple Online and Realtime Tracking with a deep association metric (DeepSORT) algorithms. To the best of our knowledge, this is the first paper to address the topic of action recognition in Australian rules football.