The Research On The Trajectory Classification Of The Table Tennis Racket

Currently, there are two puzzles of Ping-Pong robot system:It’s hard for robot to response in a short time, when table tennis movement in high speed condition; Ping-Pong robot cannot judge the type of opponents hit the ball (e.g. Backspin, Topspin, Regular), can’t judge the rotation and the direction of rotation, which will cause the robot to return the ball in single strategy, poor adaptability. To solve this two problems, this paper launches the research work on the trajectory of the racket on Ping-Pong robot system, mainly including two parts:obtained the trajectory of the racket and classification.In order to research the trajectory of the racket, we must first obtain it. This paper presents an approach to detect the position and pose of the racket. The racket region is first extracted by the segmentation technique. We adopt the improved LSD algorithm to detect the line segments in the racket region image and then extract straight lines by cluster and obtain the image coordinates of the vertices of the rectangle and processed by Calman filter. Finally, the3D coordinates of the center of the racket and the racket pose are estimated via PnP positioning Method.After obtaining the trajectory of the racket, this paper classifies the trajectory of the racket on the basis of BP Neural Network and ELM.When using BP Neural Network to classify the trajectory, after a lot of experiments, ultimately select the characteristic, position, speed and the normal vector of the center point of the racket, describe the trajectory of the racket. First, we preprocess the input trajectory data, and then create a new BP network for classification, the input is the position, speed, the normal vector of the center point of the racket on each of the5frames before and after the hitting point, the output is the type of the trajectory, and the neural network learn a lot of historical data off line, and finally the trained network model can be used for the table tennis racket trajectory classification. At the same time, in order to verify the feasibility and robustness of the classification of BP neural network, this paper does another two exercises.When using ELM to classify the trajectory. Similarly select the characteristic, position, speed and the normal vector of the center point of the racket, describe the trajectory of the racket. Experiments show that compared with classification in use BP, ELM training time is shorter, but the accuracy will be relatively low.