Machine Learning for automatic summarization of sport videos
Automatically extracting key events from sports videos is extremely useful for teams, coaches, fans, and anybody who only wants to watch the highlights of a match. In our work, we propose a solution to automate the detection of key events in GAA hurling matches. In particular, our solution focuses on the detection of the four important events in the sport: puckouts, kickouts, scores, and wides. Due to both the wide variety of detectable events and the scarce availability of training data, we did not use common techniques for video classification & detection tasks (i.e. LSTM, 3D-CNNs). Instead, we designed handcrafted features capable of capturing discriminative information using alternative techniques (i.e. optical flow, people detection & tracking). We then fused these features together in order to predict the probability that a specific frame relates to a specific target event.