Research On Object Tracking Algorithm Based On Online Update In Complex Background

Visual object tracking requires tracking the specific target in a video automatically and accurately.It has been applied to security surveillance system,human-computer interaction,vision navigation system and so on.With the progressive development of computer vision,the requirement of stability of target tracking performance becomes higher.However,the target is often disturbed by various factors which are background clutters,occlusion,large scale deformation and so on,so it is crucial to the computer vision.In recent years,object tracking algorithms benefit from the latest advances in deep learning and the absolute advantages in neural network which are big data and rapid calculation.This paper does some research on the problem above,and put forward the online tracking by flow neural network.The contributions are as follows.A module of feature fusion by option flow is proposed to deal with the background clutters and make the best use of inter-frame information in object tracking.The current frame and the previous frame are put into the flow network and the object motion information in the two frames will be available.In order to enhance the feature representation of the target region,this paper fuses the current frame and the motion information.By way of the rich contextual information,the network has a strong target discrimination.A module of online update by meta-learning is proposed to deal with target drift caused by the target deformation or occlusion.This module introduces the average peak correlation energy as the tracking confidence index.Through the response volatile and the confidence of target detection,we could know the occlusion and whether to update the model in time.And it would avoid the accumulation error caused by the wrong update.Then,the module can adapt to the target appearance by itself and enhance the robustness of the network by updating the template rapidly in meta-learning.Finally,this paper does ablation experiments on OTB100 dataset,which support that the module of feature fusion by option flow could improve accuracy location in the cluttered background effectively and the module of online update by meta-learning could update the template in self-adaptive when the appearance of the target changes.At the same time,the horizontal comparison experiments on the public datasets of OTB100,La SOT,VOT2018 and Tracking Net show that the method is superior to the current object tracking algorithms.