Deep-Learning based positional estimation of agents, Automated Real-time Spatiotemporal tracking, & planar visualisation
Football is one of the most played and watched sport in the world with over a billion fans spanning across the seven continents. With the advent of Artificial Intelligence, the way the sport is enjoyed is changing - the most prominent examples being the addition of Video Assistant Referee (VAR) system and Semi-Automated Offside Technology (SAOT), which have improved fairness in the sport, and automated decision making process in detecting offsides. There is still a great scope and potential for many implications of artificial intelligence in this sport.
In this research, I explored and built a novel way of visualising and analysing the games using deep learning and computer vision. Firstly, players, referees and ball are tracked in real-time with broadcast-feeds, and visualised on a 2-dimensional plane with a top-down view using state-of-the-art techniques. Subsequently, the tracked data is synthesised into a spatiotemporal dataset used to make positional estimates of players with respect to time, even when he/she is not visible within the camera’s field of vision. This novelty is achieved using advanced recurrent neural networks such as LSTM and gradient boosting models such as XGBoost. In addition, major events such as game pauses, throw-ins, passes, and goals are automatically detected and recorded as logs for future data analysis.
While the scope of this research is to incorporate artificial intelligence to enhance audience’s viewing experience, the spatiotemporal data and deep-learning prediction and tracking models can be used by football clubs, game developers and analysts to acquire deeper insights into gameplays, strategies, tactics and much more.