Imputing Multi-Agent Trajectories from Event and Snapshot Data in Soccer

Recent advances in wearable sensors and computer vision technologies have enabled the collection of tracking data in team sports, which has become a core resource for fine-grained analysis. However, the availability of tracking data remains constrained by high acquisition costs and technical limitations. Compared to tracking data, event data recording on-ball actions and snapshot data providing partial player positions are more widely accessible. To this end, this study proposes a novel approach to predict the positions of all players at each event timestamp in soccer matches, leveraging the limited information available from event and snapshot data. We propose an event-based imputation model that integrates spatial and temporal attention to capture the spatiotemporal multi-agent structure. In experiments, we evaluate our model on 13 soccer matches, achieving average position errors of 5.84 m. To assess the practical utility of our approach, we apply it to a downstream task called Pitch Control, which requires full tracking data. These results highlight the potential of event-based position imputation to expand access to fine-grained analysis in data-constrained settings.