Authors: Tobias Björk & Marcus Salzer

Abstract: 

Multimodal trajectory prediction is a task commonly found in fields like autonomous driving in order to model interactions between pedestrians and vehicles, but is also relevant in sports. In the context of football, predicting the future trajectories of the players and ball can be used to control a pan-tilt-zoom (PTZ) camera, allowing for automatic video recording of games. Recently, transformer architectures have gained popularity due to their ability to effectively capture both social and temporal relationships. Previous works based on transformers typically encode past trajectories to predict one or multiple future modes, but usually without quantifying the likelihood for each mode. In this work we propose statistical MART (sMART) which builds on a previous graph based transformer model (MART), extending it to produce probabilities for each future mode. This multimodal output of the model is used to estimate a distribution, weighted by the probabilities, for each player and ball. The model is trained and evaluated on football tracking data, extracted from the $2023-2024$ seasons of Allsvenskan by Signality AB. Performance is measured with common distance based metrics such as ADE/FDE, as well as probability based metrics, intended to measure the predicted probability in a region close to the ground truth. Experiments comparing sMART to a naive MLP network and a velocity baseline show that sMART outperforms both these, for predicting the future 4 seconds of trajectories. Specifically, in the probability metric we see an improvement of 24 % for the players, and 11 % for the ball, compared to the velocity baseline, and approximately twice the probability compared to the MLP.