2023 Author: Bryan Walter | [email protected]. Last modified: 2023-05-21 22:24
Japanese programmers have taught a neural network in real time to predict the place of a ball in table tennis based on the position of the player's body, including even before the racket touches the ball. Experiments with professional and amateur players have shown that the predicted drop point is 75 percent of the ball's diameter, according to an article presented at the CHI 2020 conference.
One of the basic skills required in ball games is predicting the behavior of an opponent and the path they will take to send the ball. Engineers have long been trying to create robots that can make such predictions and play as well as humans. For example, a few years ago, the Japanese company Omron introduced a large table tennis robot that can quickly track the trajectory of the ball and the player. As with almost all similar designs, only its previous positions are used to calculate the trajectory of the ball, and the calculations begin after hitting the racket. But since, when serving, the professional player already knows where he will send the ball, his professional opponent, as a rule, tries to predict the direction of the hit and get ready during the swing by the server's body position.
Erwin Wu and Hideki Koike of Tokyo Institute of Technology decided to use the same approach to computer predict the trajectory of a ball. Their algorithm is based on two neural networks with different tasks and architecture. First, data from a webcam installed from above from the receiving player's side goes to the ResNet50 convolutional neural network, which marks the position of the serving player's main body segments on each frame.
Then the last ten frames with marked body positions are fed to a long short-term memory (LSTM) neural network, which is responsible for calculating the location of the ball. Together with them, this neural network is given data on the position of the gaming table, due to which the calibration takes place. The resulting drop point, if necessary, is displayed on the table in real time using a projector.
The authors collected the training data themselves, tracking the flight of the ball with a high-speed camera, but reduced the rate from 240 to 30 frames per second to match the webcam used outside of training. In addition, they recorded sound during the game to accurately determine when the ball touched the table. For preliminary training, the developers also used recordings of tennis pitches from the Internet. By comparing the exact trajectory data with the positions of the players' bodies during multiple serves, they trained the algorithm to work reasonably accurately.
After training the algorithm, the developers tested it on successful pitches of eight people, of whom four were frequent players and four were amateurs. The results showed that the predictions of the algorithm correctly fit into a circle with a diameter of four centimeters (the size of the ball) 81.25 percent of the time when serving as a professional and 68.75 percent when serving as an amateur (average 75 percent).
In another study, they tested the system's applicability for training by recruiting six people who had hardly ever played table tennis. They had to accept the serve of an experienced player in two conditions: with or without the projection of the calculated drop point onto the table. It turned out that with the projection, they successfully hit the ball to the opponent's side almost 11 percent more often.
Last year, American developers taught a neural network to create an animated three-dimensional model based on a video with the beginning of a person's movement, which accurately predicts further movement.