Abstract: The Kernel Correlation Filter（KCF） tracking algorithm will suffer from the phenomenon of target loss when the target is affected by such factors as scale variation, severe occlusion, similar target interference, illumination variation and out of view. Once the target is lost, the KCF tracking algorithm itself is imperceptible, and the tracker will track the background information as a target to keep tracking, resulting in the complete loss of the real target. In order to solve this problem, based on the KCF tracking algorithm, this paper proposes a target loss early warning mechanism based on outlier detection method. The method detects the peak of the response of each frame using the mean and standard deviations of dynamic peak value data of fixed dimensions, and if the abnormal peak is found, the target is lost or will be lost, which solves the problem that loss target can not be perceived during the tracking process. The experimental results indicate that the proposed method can correctly predict the loss of target in the tracking process of KCF algorithm, and the success rate is 100%. The proposed method has high reliability and can provide a reliable basis for when to load target re detection mechanism, after the target is lost.

Key words: outlier detection, Kernel Correlation Filter（KCF）, target loss

摘要： 当目标受尺度变化、严重遮挡、相似目标干扰、光照变化和出视野等因素影响时，核相关滤波器（KCF）跟踪算法会出现目标丢失现象。目标一旦丢失，KCF跟踪算法本身是不能察觉的，并且跟踪器会将背景信息作为目标继续进行跟踪，导致目标彻底丢失。针对这一问题，在KCF跟踪算法的基础上，提出了一种基于异常值检测方法的目标丢失预警机制。该方法利用一组固定维数动态峰值数据的均值和标准差对每帧的响应峰值进行检测，如若发现异常峰值，则判定目标丢失或即将丢失，解决了KCF跟踪器在跟踪过程中目标丢失不能察觉的问题。实验结果表明，所提出的方法在KCF算法跟踪过程中目标丢失时，能够正确预警，成功率达到100%，具有很高的可靠性，为目标丢失后何时载入目标重检测定位提供可靠的依据。

关键词: 异常值检测, 核相关滤波器（KCF）, 目标丢失