Autoregressive Generation for Basketball Defensive Trajectory

Abstract

Tactics learning in VR has been proved to be effective for basketball training In VR training system the coach can input offensive trajectories by drawing via an electronic tactic board but defensive trajectories should be generated automatically to improve the efficiency and usability To provide a better virtual training process we aim to simulate more realistic defensive trajectories based on an autoregressive method In the proposed method the previous defensive trajectories the current offender positions and the current ball position are taken as the input Then a generative model based on a differential position sampling algorithm and a causal convolution mechanism is designed to learn the relation between player positions The similarity between the generated defensive trajectory and the real defensive trajectory is evaluated in both objective and subjective manners For objective evaluation we compare the defensive position movement speed and acceleration difference between the generated trajectories and the real ones In addition we calculate the empty space for the offender and the defensive pressure based on the Voronoi algorithm to compare defensive trajectories For subjective evaluation we recruited 70 experimenters to conduct questionnaires for judging whether the defensive trajectories shown in the video is realistic According to the questionnaire result the experimenters are difficult to distinguish the real basketball defensive trajectories from the generated ones This implies that the proposed autoregressive model can generate realistic defensive trajectories

Date of Award	2020
Original language	English
Supervisor	Wei-Ta Chu (Supervisor)