A deep learning based framework for badminton rally outcome prediction

Badminton is a fast-paced net-based sport in which players' actions and strategies in-game determine their chances of winning. With sports analytics gaining popularity due to its capability in providing valuable information for players and coaches to counter opponents with tactics, some recent research works attempted to perform stroke recognition. However, there has been little research into using stroke sequences for sports analytics. In this paper, we propose a player-independent framework to investigate the relationship between strokes and rally outcome in badminton games. To classify the rally outcome, strokes are represented by deep features extracted using CNN and fitted into LSTM. Experiments with various variants of GRU and LSTM models demonstrate that Bidirectional LSTM gives the best prediction performance, with ResNet-18 as the feature extractor. Additional experiments were performed to study different features that represent the stroke as plain text and player's pose, as well as methods to augment a small sequential dataset.