针对核相关跟踪算法(KCF)对特征敏感及无法跟踪尺度的问题, 本文从特征提取和尺度自适应两个方面对核相关滤波跟踪算法进行了研究。提出了一种基于色度饱和度-角度梯度直方图特征的自适应核相关跟踪算法来改善KCF算法的跟踪性能。首先, 研究了HSI颜色空间的特点, 基于颜色和梯度是互补的图像特征, 提出了一种融合了梯度和颜色的HHS-OG特征来有效提高原始KCF算法对目标和背景的判别力。其次, 针对KCF无法处理目标尺度变化的问题, 在跟踪的检测阶段采用一组固定的尺度因子进行图像块采样, 根据得到的滤波响应图估计目标的最优位置和尺度。将所提算法在大量视频序列上进行了跟踪实验, 结果显示其平均跟踪速度为37.5 frame/s, 跟踪精度和成功率分别提升了5.4%和10.1%。实验表明HHS-OG特征具有良好的目标-背景判别能力, 能够实现鲁棒跟踪, 而尺度自适应策略能较大程度地提高跟踪精度。

Since Kernelized Correlation Filters (KCF)tracking algorithm is sensitive to feature selecting and unable to estimate object scale, this paper researches the KCF tracking algorithm based on feature extraction and scale adapting. A scale adaptive KCF tracker by using HHS-OG (Histogram of Hue Saturation and Oriented Gradient, HHS-OG) feature was proposed to improve the tracking performance of the KCF tracker. Firstly, the HSI color space was studied. By taking the complementary of color and gradient in an image, a novel HHS-OG feature focused color and gradient features was proposed to improve the discrimination ability of the KCF algorithm to backgrounds and targets. As the KCF algorithm is unable to process the changed scale, a set of scale factors were used to sample image patches in the detection stage of tracking and the generated corresponding filter response maps were used to estimate the optimal target position and scale. The proposed tracker was tested on a large tracking benchmarks with 50 video sequences. Experimental results show that the tracker runs at a high speed of 37.5 frame per second and has a significantly improvement of 5.4% in representative precision score and 10.1% representative success score. The HHS-OG feature has good discrimination ability for backgrounds and targets and has robustness for target tracking. The scale adaptive strategy is effective for improving tracking performance.