Efficient Soccer Action Recognition with Motion Feature Integration in 3D CNNs

Action recognition technology is widely used in soccer match analysis, helping to evaluate player performance and game strategies. However, soccer actions vary significantly in speed and movement complexity, making it challenging for models to capture motion dynamics accurately. Optical Flow has been widely used to enhance motion representation, but its high computational cost makes it impractical for real-time applications. To address this issue, recent studies have proposed feature-level motion as an alternative, effectively reducing computational overhead while maintaining motion modeling capabilities. Building upon this approach, we extend motion features to 3D CNNs to further enhance spatio-temporal feature learning. In the soccer action recognition task, our method achieves an accuracy of 91.63%, outperforming the baseline by 4.42% while reducing computational cost by 18.9%. These results demonstrate the effectiveness of our approach as a more accurate and efficient solution for real-time applications such as sports broadcasting and match analysis.