Learning background-aware and spatial-temporal regularized correlation filters for visual tracking

作     者：Zhang, Jianming He, Yaoqi Feng, Wenjun Wang, Jin Xiong, Neal N. 

作者机构：Changsha Univ Sci & Technol Sch Comp & Commun Engn Hunan Prov Key Lab Intelligent Proc Big Data Tran Changsha 410114 Peoples R China Northeastern State Univ Dept Math & Comp Sci Tahlequah OK 74464 USA 

出 版 物：《APPLIED INTELLIGENCE》 (Appl Intell)

年 卷 期：2023年第53卷第7期

页      面：7697-7712页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China Basic Research Fund of Zhongye Changtian International Engineering Co., Ltd. [2020JCYJ07] Postgraduate Training Innovation Base Construction Project of Hunan Province [2019-24851] "Double First-class" International Cooperation and Development Scientific Research Project of CSUST [2019IC34] Enterprise-University Joint Postgraduate Scientific Research Innovation Fund of Hunan Province [QL20210205] Postgraduate Scientific Research Innovation Fund of CSUST [CX2021SS70] 

主　　题：Convolutional neural network Correlation filter Background-aware Spatial-temporal ADMM 

摘      要：In visual tracking, correlation Filters (CFs) have attracted increasing research attention and achieved superior performance. However, owing to the larger search area, more background information is introduced to the shifted samples, meaning that tracking errors are prone to appear in the detection stage. Accordingly, in this work, firstly, hand-crafted features and deep features extracted from pre-trained convolutional networks are combined to improve the representation ability of object appearance. For deep features, we use two different VGG networks for extraction. Secondly, in an attempt to solve the problem of the object background of the traditional CF model not being modeled over time, and owing to the lack of spatial-temporal information of the image, we propose a new background-aware and spatial-temporal regularized correlation filters model (BSTCF) that introduces the background constraint and spatial-temporal regularization. The proposed BSTCF can effectively model not only the background but also variations in the background over time. Finally, we transform the objective function of BSTCF into an unconstrained Augmented Lagrange multiplier formular to promote convergence to the global optimum solution. Moreover, we adopt the alternating direction multiplier method (ADMM) to produce three sub-problems with closed-form solution, then propose a corresponding algorithm. Based on the above, we construct an intelligent tracking system and carry out extensive experiments to test its performance on OTB-2013, OTB-2015, TC128, UAV123, and VOT2016 public datasets. The experimental results demonstrate that the tracking algorithm achieves superior performance.