A Research On Fast Moving Object Tracking Based On Binocular Vision

In the area of artificial intelligence,more and more mobile robots come into people's lives,and human-computer interaction becomes more frequent.At present,robots mainly acquire external information through vision.With the demand increasing,it is difficult for service robots with monocular vision to meet the demand of obtaining external information.Therefore,more and more robots equipped with binocular vision that imitate human eyes appear in various scenes.While executing the visual servo task,the most important part in the vision is to locate and track the target.This paper mainly starts from the visual requirements of the table tennis robot(the main task of the robot in this paper is to continuously bounce the ball),and studies the use of binocular vision to track and locate the fast-moving table tennis ball.The main challenges of the study are tracking fast-moving ping-pong ball precisely,locating the target position precisely and predicting the position and the velocity of drop point.The system needs high real-time performance.Firstly,this paper proposes a tracking algorithm that fuses the trajectory prediction information.The final target position is generated by fusing the position from tracking and the position from predicting.This article based on higher real-time algorithm---Siam RPN,improving feature extraction network,improving the real-time performance,and combining the predicted target position from kalman filter and the tracking result by the confidence of tracking.When the target is sheltered,the confidence level is low,the final result is mainly generated by the prediction information,and tracking box can move ahead by predicted information.When the target appears again,it can be captured with a higher probability.Then a fast moving target detection algorithm is proposed for proving the initial frame for tracker.Because the traditional detection algorithm is very hard to use in high real-time requirements project,this article uses the frame difference method to detect the moving object in the field of vision,in order to eliminate the interference of other elements,this article use color filter in the current frame again.Merging the two results can filter out most of the interference target,and detect the specific color of table tennis effectively.Then a binocular vision system for table tennis robot is designed.After study the binocular vision calibration and binocular positioning principle,this article also calculated position by mapping the target position in camera coordinate to the coordinate system of the robot,then analyzing the data and fitting the saddle-likely error.With the method of secondary calibration,the effects of camera calibration error can be reduced,and we can further reduce the target positioning error.In this paper,the experiments of testing tracking algorithm on OTB100 tracking data set and actual table tennis playing data set verified the real-time performance and accuracy of the tracking algorithm.The experiments on analysing the error between the truth position and calculating position after the second calibration verified the accuracy of the binocular vision positioning system.By combining detection,tracking and positioning,the method of trajectory fitting is used to predict the position and speed of the falling point of the table tennis.The experiments on error analysis proves that the binocular vision system is real-time and accurate,and can basically meet the visual needs of the table tennis robot.