乒乓球机器人的视觉伺服系统

Abstract

As a typical real-time ‘eye-hand’ system, the Ping-Pong robot is an ideal platform for high-speed visual perception and fast motor control. It involves high-speed object identification and tracking, trajectory prediction, and fast arm servo, which are also very significant in industrial and military field. This paper proposes a whole solution of real-time vision system for a ping-pong robot, including target recognition algorithm based on feature histogram analysis and fast contour search, modeling algorithm based on dynamics analysis, learning and training algorithms of model parameters, state estimating and trajectory predicting algorithms based on model adaption. To get more robust and accurate prediction under various serving conditions, the proposed method establishes two equivalent forms of the dynamic model of flying ball, the discrete form for state estimation and the continuous form for trajectory prediction. These two forms share the same parameters. According to force analysis, the model parameters are deeply related to ball’s state. So we trained the model parameters offline respect to ball’s state, instead of setting them to constant values. Thus the model can be adapted accordingly online. Experimental results show the effectiveness and accuracy of the algorithms. It also works very well on Ping-Pong robots to rally with each other and with human.