Trajectory Prediction and Hypothetical Outcome Visualization in Basketball Using AI Models

Abstract

In modern basketball, trajectory prediction and real-time feedback systems have become useful tools for game analysis, tactical planning, and decision-making. Our paper is a contribution on short-term prediction of player and ball movement in professional basketball, using historical movement and contextual game information as input. We use data from 500 games from a professional basketball league, provided by a sports analytics company. The raw video feeds are processed to extract 3D positions of all ten players and the ball, sampled at 60Hz. We propose a dual transformer architecture, using masked autoencoders for pretraining, and report very good results. We additionally conducted a perceptual study to verify that the predicted trajectories look realistic.