Research on Table Tennis Rotation Prediction Based on Deep Learning and Multi-scale Feature Fusion

Abstract

This paper designs a table tennis target detection network based on feature fusion network. In the feature extraction network, the cross-layer connection network (CSPNet) is used to strengthen the learning ability of the convolutional neural network and reduce the number of parameters of the network to improve the detection speed of the network; for the existing network, the detection accuracy of small targets such as table tennis is low and the positioning ability is low. For the problem of poor, this paper uses a feature fusion network, adds a bottom-up connection on the basis of the feature pyramid network, and finally uses an adaptive pooling method to fuse the feature information of each feature map, and the upper-level semantic information rich features the layer and the lower layer are connected with the feature layer rich in target location information to perform feature fusion to enhance the network's ability to locate small targets. Because the network in this paper only needs to detect the table tennis target, and the table tennis target in a single picture is small, resulting in a waste of training costs, this paper proposes a new data augmentation method, which combines the data in each picture during training. Multiple copies of Ping Pong, based on the existing data, further increase the richness of the data set. After adjusting and optimizing the network structure, the network of this paper can complete real-time tracking and accurate positioning of table tennis under different background and lighting conditions.

This paper builds a LSTM-based rotating table tennis trajectory prediction network. By stacking LSTM networks, the task of predicting the trajectory of table tennis can be realized, and real-time and certain accuracy can be met. For different types of spinning balls, there are specific trajectory changes in their trajectories to follow; therefore, this article attempts to infer the general type of ping pong ball rotation based on the flight trajectory of the ping pong ball. This article decomposes the table tennis movement into three coordinate movements, and calculates the movement speed of the table tennis in the three coordinate axis directions through the obtained three-dimensional coordinate information of the first five moments of the table tennis, passing the speed and the set threshold of Compare the type of rotation to judge the general type of rotation, and control the end joint of the robot to hit the ball at a specific angle for the ball of different rotation type, which effectively improves the success rate of the table tennis robot. Compared with the traditional physical model, the network in this paper has higher anti-interference ability and accuracy.