Real-time Curling Trajectory Prediction via Attention-Enhanced Sequential Modeling and Multi-Dimensional Feature Fusion

Abstract

Trajectory prediction of curling stones has garnered increasing interest for its applications in sports analytics and robotic coaching; however, current methods typically rely on offline processing and fixed-length input windows, resulting in information loss and poor adaptability in real-time scenarios. To address this challenge, we propose a real-time adaptive LSTM-Attention model, operating solely on low-dimensional x, y coordinate streams-specifically designed for online trajectory prediction. Our method dynamically updates the prediction window by continuously incorporating new sensor measurements, leveraging an attention-augmented LSTM architecture to maintain historical motion context. We evaluate our method on real curling trace datasets and demonstrate a Median Absolute Error of 0.22 m. Experiments on real-world curling datasets show that our framework provides timely and reliable predictions suitable for live competitive environments. This research significantly enhances the practicality of trajectory forecasting tools, contributing to more informed tactical choices and improved competitive performance in curling matches.