The neural network predicted point of fall of the ball in table tennis before hitting the racket

Japanese programmers have learned neural network in real-time to predict the crash site of the ball in table tennis at the body position of the player, including before the racket touched the ball. Experiments with professional players and Amateurs showed that the predicted point of the fall in 75 percent of cases fit into the diameter of the ball, say the authors of the articlepresented at the conference CHI 2020.

One of the main skills required in ball games, is the prediction of the behavior of the opponent and the trajectory on which it will send the ball. Engineers have long been trying to create robots capable of doing such predictions and play worse people. For example, several years ago Japanese company Omron has introduced a large robot to play table tennis, the ability to quickly track the trajectory of the ball and the player. As in almost all similar developments, to calculate the trajectory of the ball only used his previous position, and the calculations start after hitting the racket. But since when submitting a professional player already knows where he will send the ball, his opponent-a professional, usually tries during the backswing the position of the body about the feed to predict the direction of the shot and get ready.

From Erwin (Erwin Wu) and Hideki Koike (Koike Hideki) from Tokyo Institute of technology decided to use a computerized prediction of the trajectory of the ball the same approach. At the heart of their algorithm is based on two neural networks with different objectives and architecture. First, the data from the web camera installed on top of the host player get on convolutional neural network ResNet50 which lays out every frame the position of the major segments of the body feed of the player.

Leave a Reply

Your email address will not be published.