Less is More: End-to-end Learning of Insights from a Single Motion Sensor for Accurate and Explainable Soccer Goalkeeper Kinematics

On-field sensor-based soccer player tracking solutions are emerging and provide new insights into the dynamics of the player during training or a match. Yet, not all player positions are equally privileged. Goalkeepers’ training and performance assess- ment were for a long time ignored. Understanding what is ”the side of the post where most high dives were performed” provides valuable input for both the trainer and the athlete to improve perfor- mance or avoid injuries. In the current study, we focus on a practical methodology to extract insights from goalkeeper kinematics to inform such analytics. We demonstrate that information from a single motion sensor can be successfully used for learning patterns in goalkeeper’s motion and provide an explainable goalkeeper kine- matics assessment. We employed raw and quaternions data and we evaluated a series of machine learning algorithms that discriminate dive types (i.e. binary classification) and dives from other types of specific motions (i.e. multi-class classification) directly from the data. Our results demonstrate that XGBoost outperforms other approaches in terms of performance when considering both raw and quaternions, essentially benefiting from both types of data. Additionally, each prediction of the model is accompanied by an explanation of how each sensed motion component contributes to describing a specific goalkeeper’s action captured by the model. The explainable predictions along with the efficient deployment of XGboost were decisive in our applied study. We evaluated our methodology on a first batch of experiments using online available data from 7 goalkeepers during 30 minutes-long training sessions.