Research on Match Event Recognition Method Based on LSTM and CNN Fusion

This conference paper introduces a combination of CNN and LSTM networks to improve automated recognition of match events in sports analytics. The proposed model is able to handle both spatial and temporal aspects from video, unlike previous approaches that deal with them separately. On the SoccerNet dataset and with sports footage, the fusion model managed to achieve 92.3% in classification accuracy, a 0.901 F1-score and showed a clear improvement in recalling rare events, like counterattacks (+31%) and penalties (+26%). By running operational assessments, it is proven that the service can be deployed in real time, offering sub-100ms latency and a 42% decrease in false alarms in use cases like broadcast overlays and coaching analytics. It is clear from the results that using spatiotemporal synergy improves the model's performance, making it suitable for high-speed sports intelligence systems.