结合尺度预测的核相关滤波器目标跟踪方法

Kernelized correlation filter based visual tracking with scale estimation

视觉目标跟踪问题中,被跟踪目标的尺度变化普遍存在。为解决这一难题,本文在核相关滤波器目标跟踪方法的基础上提出了一种能结合尺度预测的目标跟踪方法,简称为KCFSE。该方法使用两种更新策略不同的岭回归模型。在实际跟踪过程中,先采用可塑性强的模型跟踪目标的位置偏移。然后,以此位置为中心,构建图像金字塔,利用稳定相强的模型预测目标的尺度变化。对10组视频序列进行的实验测试表明,该方法在处理尺度变化的被跟踪目标时性能明显优于其他目标跟踪算法。

Scale variance of the object is universal in visual tracking applications. To solve this problem, we propose a novel KCF based tracking algorithm with scale estimation called KCFSE. In this algorithm, two regression model with different updating strategies are used. During the tracking procedure, the regression model with more plasticity is adopted at first to detect the spatially shift of the object. Afterwards, an image pyramid is built around the position detected and the regression model with more stability is adopted to estimate the the scale variance of the object. Experiments on 10 video sequence show that KCFSE outperforms other classic tracking algorithms as well as KCF when the scale of the tracked object is variant.