Detection And Trajectory Prediction Of Table Tennis Based On Lightweight Networks

With the development of the times,the integration of technology and sports has become increasingly close.As a real-time intelligent motion servo platform,table tennis robots cover various core technologies such as object detection,3D reconstruction,trajectory prediction,and have very important research significance and application value.Due to the small size,fast flight speed,and complex and variable trajectory of table tennis,it has always been difficult to accurately identify,locate,track,and predict its trajectory,which has attracted more and more scholars’ attention.Based on the above background,this article chooses binocular cameras as the research foundation,and focuses on solving the problem of poor accuracy and realtime performance of traditional table tennis object detection methods due to environmental and light interference by building a lightweight network.It achieves rapid detection,positioning,and 3D trajectory reconstruction of table tennis,and finally achieves accurate prediction of table tennis flight prediction by building a lightweight trajectory prediction network.The main work and innovations are:(1)This article first studies the camera imaging model and distortion model,deduces the transformation relationship of coordinate points in the four major coordinate systems,uses Zhang Zhengyou calibration method to calibrate the internal and external parameters of the binocular camera,and collects table tennis game images from different venues and backgrounds through network collection and on-site photography.In order to enhance the richness of the dataset and the generalization ability of the model,In the basic dataset produced,images are randomly selected with a 20% probability for 90 ° clockwise and counterclockwise rotation,brightness adjustment,adding random noise,and flipping left,right,up,and down to form a table tennis dataset that covers a total of 3506 images in various scenes of table tennis.Labels are manually annotated to lay the foundation for subsequent detection,positioning,and trajectory prediction.(2)In response to the poor accuracy and real-time performance of traditional detection methods for table tennis,this paper proposes a lightweight table tennis object detection algorithm based on the YOLOv5 s framework-Shuffle-YOLOv5 s.This algorithm uses an improved ShuffleNetv2 network unit combination to reconstruct the YOLOv5 s backbone network,accelerating feature extraction speed.Secondly,an efficient channel attention mechanism is introduced in the process of feature fusion to effectively improve the detection performance of the model.The convergence speed and positioning accuracy of the network are improved by using SIOU Loss as the positioning loss function.Finally,according to the small size characteristics of table tennis,dual scale object detection is adopted to further improve the inference speed of the model.The experimental results show that the SYOLO5 algorithm proposed in the paper reduces the parameter and computational complexity by 80% and 60%respectively compared to YOLOv5 s.It maintains a recognition accuracy of99.0% in table tennis detection tasks while achieving a detection speed of105 fps.This verifies the effectiveness of the algorithm proposed in this paper,which has important theoretical significance and application value for table tennis trajectory analysis and scientific training.(3)In response to the problem of poor accuracy in predicting trajectories through building traditional physical motion models for table tennis,this article constructed a dataset covering 3000 different types of table tennis trajectories,of which 2200 were derived from open-source table tennis trajectory data,and 800 were extracted using the detection algorithm proposed in this article through binocular cameras.And a rotating table tennis trajectory prediction network based on GRU was designed.By obtaining three-dimensional information of the top ten positions of the table tennis ball,long-term trajectory prediction for the future can be achieved.The experimental results show that the method in this paper meets the accuracy requirements,and the error of each axis is below 20 mm.