Research On Object Tracking Technology For Soccer Match Video

In recent years,intelligent analysis technology for soccer match video has developed rapidly,covering a variety of applications such as video summary and retrieval,behavior recognition,event detection,game penalty assistance,athlete training assistance and tactics analysis.Player or soccer object tracking is one key technology of soccer video analysis and also is the basis of highlevel semantic analysis and video understanding.However,due to obstacles such as object occlusion,deformation,similar interference,low resolution,scale changes and out-of-view,the performance of current object trackers in soccer matches is not ideal enough.This thesis studied the single-object tracking technology of player in soccer match video of single-shot,aiming to design a practical algorithm to realize long-term and stable tracking for soccer player object in far shot.The research work and achievements included the following components:1.Two mainstream object tracking methods,Correlation Filter and Deep Learning,were studied and compared.Due to the low complexity and speed advantages of Correlation Filter methods,comparing to the high complexity and reliance on large scale training datasets and GPU of Deep Learning methods,the Kernelized Correlation Filter(KCF)algorithm was used as the basic tracking framework from the perspective of engineering practicability.To improve the algorithm’s robustness under various tracking difficulties,comprehensive improvements were then developed.2.Aiming at the insufficient characterization ability of object feature,a feature fusion method of dual-channel translation tracking using Fast Histogram of Oriented Gradient(FHOG),Color Name(CN)and grayscale features was proposed,which realized feature fusion both on the feature layer and decision layer,and enhanced the robustness of object model.3.To solve the problem that the object scale could not be adaptively changed,a scale filter was introduced to achieve up to 33-level scale estimation on each frame.Thus,the scale variation could be tracked in real time,which not only improved the tracking accuracy,but also strengthened the robustness of tracker.4.Aiming at a lack of quality judgment and the uncorrected tracking drift or errors in the tracking process,a tracking quality evaluation and object redetection mechanism was proposed.According to the evaluation results of the tracking quality at each frame,the multi-peak re-detection on response map or object re-detection function were respectively activated to correct the object tracking results or relocate object,which improving tracking accuracy and avoiding object loss.5.A long-term Kernelized Correlation Filter object tracking algorithm for soccer player was proposed based on these developments.Then,it’s fully tested and optimized on the BSPT(Benchmark for Soccer Player Tracking)dataset.Experiments show that the effectivity of the improved algorithm has been significantly enhanced in various scenes such as object occlusion,deformation,rotation,motion blur,out-of-view,etc.,and the performance is elevated by nearly 50% contrast to KCF.The algorithm is competitive to the deep learning object tracking methods also.Additionally,the algorithm has achieved an average running speed of 78.1FPS,which fully meets the real-time requirements.