基于双核模型上下文的流形正则相关滤波跟踪算法  

Manifold Regular Correlation Filter Tracking Algorithm Based on Dual-Core Model Context

作　　者：崔丽群[1] 贺情杰 何牧泽 Cui Liqun;He Qingjie;He Muze(College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China;Beijing Chaoxing Company,Beijing 100000,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105 [2]北京超星公司,北京100000

出　　处：《激光与光电子学进展》2021年第12期147-158,共12页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(41701479);辽宁省自然科学基金(20180550529)。

摘　　要：为了兼顾算法的跟踪速度与精度,提出了一种基于双核模型上下文的流形正则相关滤波跟踪算法。其中,结合上下文相关框架与相关滤波算法的主模块承担主要跟踪任务,可弥补相关滤波学习模型中余弦窗过滤的背景信息。对上下文相关样本进行流形正则处理,可达到惩罚上下文相关框架、优化主模块模型的目的。辅助模块则结合了核相关滤波算法与卷积特征,当跟踪目标发生遮挡、形变或超出视距等情况,跟踪置信度低于经验阈值时启用辅助模块,防止主模块模型发生漂移。由于主模块的跟踪速度快、精度低,而辅助模块跟踪速度慢、精度高,两个模块可在速度和精度方面优势互补。在OTB2015和VOT2016数据集上的测试结果表明,本算法在精度和鲁棒性方面均超过了其他相关滤波算法。In this study,a manifold regular correlation filter tracking algorithm based on dual-core model context is proposed to balance the tracking speed and accuracy of the algorithm.The main module combines the context-related framework and the relevant filtering algorithm is responsible for the main tracking task,which can compensate for the background information filtered using the cosine window in the relevant filter-learning model.The manifold regular processing of context-related samples can achieve the purpose of penalizing context-related framework and optimizing the main module model.The auxiliary module combines kernel correlation filtering algorithms and convolution features.When the tracking target is occluded,deformed or exceeds the line of sight,the auxiliary module is activated when the tracking confidence is lower than the empirical threshold to prevent the drifting of main module model.The main module has fasttracking speed and low accuracy,whereas the auxiliary module has slow tracking speed and high accuracy.These modules can complement each other in terms of speed and accuracy.The test results on the OTB2015 and VOT2016 data sets show that the algorithm has better accuracy and robustness than other correlation filtering algorithms.

关 键 词：图像处理 目标跟踪 相关滤波 上下文 流形正则 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]