Table Tennis Target Detection And Rotating Ball Trajectory Prediction Based On Deep Learning

In the robot vision system,how to detect and predict the moving object with rotation quickly and accurately has very important research significance and application value.Table tennis robot has attracted a large number of researchers to study deeply since 1980 s.Under the current technical background,with the development of AI technology,the research on table tennis robot has a broader prospect.This thesis focuses on the visual system of table tennis robot,and takes the rotating table tennis as the research object,discusses the feasibility of deep learning in the visual system of table tennis robot,and verifies it in the physical system of seven-degree-of-freedom KUKA manipulator.This thesis makes table tennis data set which can be used for deep learning network training.Under binocular vision system,table tennis data sets under different illumination backgrounds were collected,and some pre-processing work was done on some images to increase the richness of the data sets.For each picture,label information is manually labeled for network learning.Finally,more than 8000 table tennis image data are constructed,which lays a foundation for the follow-up work.This thesis designs a table tennis target detection network based on feature fusion network.In the feature extraction network,cross stage connection network(CSPNet)is used to enhance the learning ability of convolutional neural network and reduce the network parameters to improve the detection speed of the network.In view of the low detection accuracy and poor positioning ability of the existing network for small targets such as table tennis,this thesis adopts feature fusion network and adds a bottom-up network on the basis of feature pyramid network Finally,the selfadaptive pooling method is used to fuse the feature information of each feature graph.The feature layer with rich semantic information in the upper layer is connected with the feature layer with rich target location information in the lower layer for feature fusion,so as to enhance the positioning ability of the network for small targets.Because the network in this thesis only needs to detect the table tennis target,and the table tennis target in a single picture is small,resulting in the waste of training cost,this thesis proposes a new data enhancement method,which copies the table tennis in each picture in training,and enhances the generalization ability and learning efficiency of the network.After the adjustment and optimization of the network structure,the network in this thesis can complete the real-time tracking and accurate positioning of table tennis under different background and lighting conditions.This thesis presents a table tennis trajectory prediction network based on LSTM.By stacking the LSTM network,the table tennis trajectory prediction task can be realized,and the real-time and certain accuracy are satisfied.For different types of spinning ball,there are specific trajectory changes in its trajectory;therefore,this thesis attempts to deduce the general rotation type of table tennis according to the flight trajectory of table tennis.We decompose the table tennis movement into three coordinate directions,calculate the movement speed of the table tennis in the three coordinate directions by obtaining the three-dimensional coordinate information of the first five moments of the table tennis,judge the general rotation type by comparing the speed with the set threshold,and control the robot end joint to hit the ball at a specific angle for different rotation types of balls It can effectively improve the success rate of table tennis robot.Compared with the traditional physical model,this network has higher anti-interference ability and accuracy.Finally,the research method is verified in the 7-DOF KUKA manipulator system.