一种利用最大间隔相关滤波的鲁棒目标跟踪算法

摘要

Aimed at the problem that the Kernelized Correlation Filter position of the target is in the image through the correlation filter,such generative models are easily subject to the interference from the backgrounds similar to the targets,causing the failure of tracking,this paper strengthens the criterion of correlation filter based on the maximum margin.The similar backgrounds are updated as negative samples to improve the tracking robustness.First,the algorithm constructs a model of maximum margin correlation and distinguishes backgrounds similar to the targets by classifier.Second,the obtained similar backgrounds are taken as negative samples and the tracking model is updated online during the tracking process.So the algorithm can adapt to the various changes of the target movement and can achieve the goal of robust tracking by online updating the target tracking model.The paper selects seventeen typical image sequences of OTB2013 and VOT2014 database and compares the results of the six correlation tracking algorithm in the experiment.The experimental results show that the algorithm can improve by 8% and by 2% on precision and success rate compared with the suboptimal algorithm,the tracking efficiency is the best and the tracking real-time-ability is comparatively good.%核相关跟踪通过相关滤波定位目标在图像中的位置,这种生成式滤波器方法容易受到与目标相似背景的干扰,导致跟踪失败.针对这一问题,通过最大分类间隔增强相关滤波器的判别性,将相似背景作为负样本对模型进行更新来提高跟踪的鲁棒性.该算法首先建立了基于最大间隔相关滤波器的目标跟踪模型,通过分类判别出与目标相似的背景;然后在跟踪过程中,将获得的相似背景作为负样本并对跟踪模型进行在线更新,适应目标在运动中的各种变化,最终实现对目标的鲁棒跟踪.在OTB2013和VOT2014数据库中选取了17个典型的图像序列进行实验,同时与6种相关跟踪算法的结果进行比较.实验结果表明,该算法在精确度和成功率这2个性能指标上,相比于次优算法,在性能上分别提升8%和2%. 不仅取得了最好的跟踪效果.而且跟踪实时性较好.