Long Term Single Object Tracking Algorithm Base On Deep Learning

Object tracking as a practical and essential vision algorithmic analysis section of which the long-term,individual focus algorithm is of great importance,previously the solution of the relevant algorithm is mainly based on traditional filtering,in recent years,the performance of the long term single object tracking algorithm based on deep learning is gradually catching up with the traditional filtering method.However,the difficulties of the target tracking task have not been well solved,among which the tracking drift of similar targets,the flashing of targets into and out of the field of view,the blurring of target motion,and the constant change of target view and scale with the tracking time are the most prominent ones,and furthermore,the longtime tracking task requires a continuous tracking status of less than several hours and more than several days,which aggravates the complexity of the above problems and keeps the research difficulty of the long-time single target tracking algorithm in high level.(1)Propose a new algorithmic framework for long-term single-object tracking tasks,which provides a more unified perspective for studying long-term single-object tracking tasks by flowing the algorithm design from the task logic,defining more clearly the key moments in the task,and solving the corresponding problems.(2)propose a new validation network for tracking tasks to independently determine the target to be selected,thus achieving a more rigorous filtering of the target to be selected,unifying the recognition quality of each module in the algorithm framework,and optimizing both the discrimination accuracy and the processing speed,with a significant improvement in the performance of the overall algorithm framework.(3)Optimize the core short-time single-target tracking algorithm to significantly improve tracking accuracy without reducing processing speed.(4)This algorithm ranked second with F1-Score,Recall,and Accuracy scores of 0.6872,0.651,and 0.733,respectively,and was seven times faster than the first place in the VOT2020 long-term tracking challenge which subordinate competition of ECCV.