Research On Visual Detection And Tracking System Of Curling Robot

With the advancement of computer science,popularization and development of sports,computer technologies such as artificial intelligence and big data are more and more applied in the sports industry,and computer and digital technology are playing an increasingly important role in sports events and stadiums.The 2022 Winter Olympics will be held in Beijing,and the Olympic Organizing Committee has reached a consensus to develop intelligent curling robots for display and interaction during the Winter Olympics curling competitions.The curling robot combines artificial intelligence technology with ice and snow sports,which is a concentrated expression of the level of computer technology and sports in our country.Based on the 2022 Beijing Winter Olympics and the Longjiang Ice and Snow Industry’s Intelligent Curling Robot Project,this paper studies the curling robot detection and tracking system.First,a large number of real curling motion image sequences were collected,and a curling multi-target tracking data set was constructed to provide basic data support for subsequent work and algorithm research.Two vision-based multi-target detection and tracking methods are introduced,namely the Deep Sort method and the Fair MOT method.The Deep Sort method uses target detection and feature extraction as the two parts of the algorithm and uses two neural networks to calculate;and in the Fair MOT method,a multi-task model is trained to perform target detection and feature extraction tasks at the same time.Using these two methods,experiments are carried out on the constructed data set,and the advantages and disadvantages of the vision-based detection and tracking method used in curling tracking tasks are analyzed.Then,in view of the similar appearance of curling and the characteristics of collision and occlusion during the movement,a point target tracking method based on the movement process is proposed.A simulator was used to simulate the curling trajectory,and a large number of curling trajectories were generated.By analyzing the characteristics of curling movement,the Simi Net-CNN model based on the convolution structure and the Simi Net-LSTM model based on the long and short-term memory structure are designed and implemented.The two neural net work models perform feature extraction and similarity calculation in the time dimension.And then track the curling trajectory.The two network models have achieved good performance on the simulation trajectory test samples.Finally,the curling robot detection and tracking system is designed,including two parts: hardware platform and software system.At the same time,a curling position settlement method based on projective transformation and suitable for curling robots is proposed,which realizes the conversion from the curling pixel position to the actual curling track position.Combining the detection and tracking method based on vision and the tracking method based on the motion process,a tracking algorithm combining multiple features is designed,whi ch overcomes the problems of collision,occlusion and similar appearance in curling motion,reduces the number of ID switching during the tracking process,and improve the tracking accuracy.