Target Tracking Method And Mobile Terminal Application Of Soccer Broadcast Video

Computer vision in sports video has become more and more mature,and in a very short period of time has gained considerable development.Target tracking is one of the hot research directions,and multi-target tracking is a difficult problem to be solved at this stage.The purpose of this thesis is to improve the existing target tracking algorithm and improve the tracking of small target players under the scene of soccer broadcast video.The algorithm is applied to mobile terminal for real-time tracking.In the pre-processing phase,this thesis proposes a video shot detection algorithm based on the double threshold limit of dominant color rate and frame difference threshold,for the phenomenon that there are a lot of shot switches in the original soccer broadcast video.The original soccer broadcast video is preprocessed so that the remote footage can be extracted completely and used to establish the data set needed for algorithm verification.In terms of small target detection,the YOLOv5 detection algorithm has shortcomings in the detection of small targets in the soccer scene.By adding a small target detection layer on the original model and giving up Mosaic data enhancement,this thesis improved the detection of small targets in the YOLOv5 s model.The experimental results show that the improved model can significantly improve the accuracy of small target detection compared with the original model.In terms of multi-target tracking,the Deep Sort multi-target tracking algorithm correlates the unreliable target detection results with the existing tracker output.Based on the Deep Sort algorithm framework,this thesis improves the problem of unreliability detection.Solve the problem of unreliable detection by collecting candidate targets from the output of detection and tracking.In order to make the optimal selection of a large number of candidate targets,a scoring mechanism based on full convolutional neural network is created to generate a unified standard confidence of detection box and prediction box.The goal is to make the longer the existence of the trajectory more reliable.The selected optimal candidate target is taken as the input of NMS.In this way,the long-term tracking drift caused by in-class occlusion can be reduced.Experimental results show that the improved algorithm has a certain inhibitory effect on ID increment in soccer scene,and also has a certain robustness and real-time performance for longterm tracking of targets.Finally,the improved target tracking algorithm is transplanted to the Android terminal to meet the needs of users for tracking specified targets.The application can display the running track of single target tracking,display the connection of station position of multi-target tracking,and provide basic video editing and other auxiliary functions.It provides simple and efficient tool support for user review.