Active Tracking Method And Its Application In Skiing

In recent years,our country's sports undertakings have continuously improved and the level of sports competition has been continuously improved.However,in terms of winter sports,there is still a large gap in the overall strength of our country's athletes.In the context of the Beijing Winter Olympics,it is of great significance and urgency to use sports science and technology to assist various trainings to quickly improve athletes' winter sports performance.Cross-country skiing is a major event in the Winter Olympics,and the performance of Chinese athletes has long been in the middle and lower reaches of the world.In view of the wide range and variety of ski tracks,trainers urgently need an automated video capture system that can capture images of the whole course of cross-country skiing for training and analysis.With the rapid development of computer computing ability,the introduction of neural networks has made significant progress in various application studies in computer vision.Object tracking methods that use neural networks and deep features have greatly improved accuracy and robustness.This topic uses a neural network-based object tracking algorithm combined with a PTZ camera to implement an active tracking method to solve the large field of view and high-speed object tracking of cross-country skiing.The active tracking in this paper belongs to a visual servo system.The target position information of the visual tracking will be used as the input of the camera control system.Therefore,the real-time collection of target position information is the key point.This paper introduces the latest twin area candidate network tracker,SiameseRPN for short,to detect the position of the player in the video.Siamese-RPN adopts a double-branch structure,and the network front-end calculates the template(frame)and the frame to be detected separately and simultaneously;in the end candidate region generation network,the classification of candidate regions and the regression of the location of the region are also carried out simultaneously,while achieving the object detection and tracking.This parallel processing structure allows the tracker to effectively use depth features and achieve processing speeds up to 160 FPS.The training and testing on real cross-country skiing videos show that the tracker can meet the real-time requirements of visual servo control.In active tracking,the rotation of the camera will superimpose the movement of the target in the image.In order to overcome the coupling effect when the camera is moving,this paper proposes a method for optimizing the tracker based on the motion prediction model.The motion prediction model uses Kalman filtering to capture the motion information of the target,predict the position of the center point,and uses the homography matrix to eliminate the effects of camera motion and optimize the delineation of the target area.The test results on the VOT2016 data set show that the optimized tracker is superior to Siamese-RPN and other trackers when processing moving targets with varying scales and movements.In addition,in order to make up for the shortcoming of Siamese-RPN that only conducts offline training,this paper further adopts the method of adding an online update structure,compressing the size of the update network and choosing a faster loss function optimization method than random gradient descent.Under the loss,the target tracking accuracy is significantly improved,which also lays a good foundation for the camera's high-precision servo control.The test results on the VOT2016 public data set show that the Siamese-RPN tracker with online update has an average overlap expectation(EAO)indicator that is better than other trackers in the experiment,reflecting the improved model's tracking accuracy and robustness.The balance is the most balanced.Finally,this paper uses the target position provided by the improved Siamese-RPN tracker to realize the tracking control of the Hikvision PTZ camera.In order to avoid the situation that the athlete's target may deviate from the camera's field of view when controlling the camera's movement,this paper uses a motion planning strategy based on the artificial potential field method.Through the mapping of image coordinates and PTZ coordinate system,the camera control parameters are finally obtained.The method in this paper has been tested and verified in Yabuli cross-country ski resort.The test results show that the location update rate of this method is improved by nearly 10 times compared with the tracking system based on GPS positioning,the tracking accuracy is greatly improved,and the tracking effect is significantly better GPS tracking system.