A Learning Framework Towards Real-time Detection and Localization of a Ball for Robotic Table Tennis System

As a real-time serving system interacting with a highly dynamic environment, robotic table tennis system has a high requirement against the accuracy and robustness of real-time detection and localization of a ping-pong ball. Relative to its size, the ball is a high speed flying-spinning object. The existing methods use general features such as color and shape to detect and localize the ball, which rigidly depends on the prior knowledge. Their performance is susceptible to the change of the environment, e.g., the light condition, the color of ball, and the disturbance of human players' presence in the image. In this paper, we propose a learning framework that trains a convolutional neural network to detect and localize a ball with high accuracy. It learns useful features from data directly without any prior knowledge. Therefore, the proposed method can effectively deal with the situation when the ball's color is changing in real-time. And it is more robust to the light condition and the disturbance of human players' presence. The effectiveness and accuracy of the method is verified using the collected data set, in comparison with the state-of-the-art method.