Golf Ball Dynamic Motion Parameter Estimation based on Ball trajectory video and LSTM

Golf ball launch monitor is a system that helps individuals train golf. It provides key parameters such as the trajectory, speed, and spin of the ball. Due to the fast ball speed and the complicated dynamic motion of the golf ball with dimple structure, the highly accurate trajectory prediction system requires expensive motion sensors such as high-speed cameras or high-performance radar sensors, which makes the system expensive and consequently prevents many golf athletes from purchasing it to improve their skill. To solve this problem, this paper proposes a deep neural network-based golf dynamic motion parameter estimation method based on continuous ball images acquired through a commercially available smartphone-level camera. According to the previous study [1], golf ball motion dynamics is governed by a physics equation that is a function of initial differentiate ball speed, spin axis and amount of the spin around the axis. Once those key parameters are identified, golf ball trajectory can be predicted by physics equations with reasonable accuracy. Since the launch monitor is supposed to estimate the ball trajectory as early as possible even before the ball is at its peak height, it is important to estimate the parameters with a minimum number of video frame images. To achieve the goal, we have utilized the custom LSTM model, demonstrating its capacity to yield remarkably precise trajectory estimations even when fed 10 sets of initial coordinates. We achieved an exceptional MSE of 0.04, underscoring the high accuracy of our approach.