Research On Anti-Background Cluttering Object Tracking Algorithm Based On Deep Learning

Object tracking is a very challenging topic in the field of computer vision,which has a very broad application background in intelligent video surveillance,robot navigation,virtual reality and other scenarios.In the process of object tracking,the scene and environment information around the target are not always monotonous.Due to the complexity of the background information during tracking,new requirements are put forward for the robustness of the tracking algorithm when dealing with multiple background alterations.In this paper,the popular correlation filtering in recent years and the target tracking algorithm with the introduction of deep learning method are thoroughly studied.Aiming at the issues that has mentioned above,improved algorithms based on the KCF tracking algorithm were been proposed.The main research and innovation work of this paper is as follows:(1)Aiming at the poor performance in the process of tracking suffering background cluttering and scale variation of KCF,a scale adaptive kernelized correlation tracking algorithm based on feature fusion of KCF tracker was proposed.Firstly,in order to improve the sensitivity of feature template to color information,a adaptive feature fusion method using HOG and CH was proposed,and the peak response after fusion was used as the basis to predict the target position of the next frame.Secondly,a scale pool containing seven fixed scale factors was used to compare the obtained target information with different scales,and the one with the largest response was used to update the current scale.Finally,APCE is used as the index of template update to control the response update of HOG feature,and the improved algorithm KSAFF is compared with the popular tracking algorithm in recent years in the OTB-100 video test set.(2)Aiming at the issue of tracking failure when suffering occlusion,based on KSAFF algorithm,the Deep KSAFF target tracking algorithm which integrates deeplearning features with spatial and temporal regularization is proposed.The algorithm was improved from three aspects: deep-learning feature fusion,learner optimization and solution.Firstly,Conv1 x,Res3d and Res4 f in pre-trained resnet-50 were used as deep-learning features,and adaptive fusion was performed for different tracking issues in order to obtain better tracking performance.Secondly,in order to solve the poor performance in handling occlusion of KSAFF algorithm,spatial and temporal regularization terms were introduced respectively on the basis of ridge regression trainer to enhance the discriminant ability between tracking target and background interference of tracker.Finally,the ADMM optimizer is used for a fast approximate solution of the response to accelerate the training of the filter.The Deep KSAFF improved algorithm is compared with the popular tracking algorithm and KSAFF in recent years on the OTB-100 dataset.