基于多阶段学习的相关滤波目标跟踪

摘要

Due to the target appearance changes and occlusion during tracking,the KCF algorithm using a single iterative update filter will accumulate much more noise information in the process of learning,which leads to the loss of the target.To solve this problem,we propose a correlation filtering target tracking algorithm based on multi-lifespan learning.We establish complementary relationship among global stage filter model,consistency stage filter model and initial stage filter model to parallel track the target.The experimental results achieved on the 51 video databases on benchmark show that our algorithm is superior to most existing methods in both overall accuracy and overall success rate with the scores of 77.6％ and 68.9％,respectively.%由于跟踪过程中目标外观变化和遮挡因素的影响,采用单一迭代更新滤波器的KCF算法在学习过程中会积累过多的噪声信息导致目标丢失.为解决该问题,本文提出一种基于多阶段学习的相关滤波跟踪算法.通过建立具有互补关系的全局阶段滤波器模型、一致性阶段滤波器模型以及初始阶段滤波器模型并行的对目标进行跟踪.在benchmark数据集的51个视频上的实验表明,本文算法取得的总体精度得分77.6％和总体成功率得分68.9％优于现有的大部分跟踪算法.