Abstract: To improve the robustness of target tracking under complex background, an adaptive feature fusion based on correlation filter is proposed. The basis of HOG features, models based on HSV color statistics is as complementary to achieve the accuracy position prediction. Then, the color name and HOG features are trained respectively, and the fusion coefficients are adaptively assigned according to the peak value of the two response graphs. The multi-channel features are used to estimation target scale based on scale pool. Finally, the high-confidence updating of model is implemented by the average peak-to-correlation energy（APCE） of two response maps. Results are compared with 5 state-of-the-art trackers on two benchmark datasets. The experiment result shows that the proposed algorithm can achieve an accurate and robust tracking performance in fast moving, complex background changes, illumination, deformation and other complex tracking scenarios.

Key words: object tracking, correlation filter, feature fusion, adaptive weighting

摘要： 为提高复杂背景下目标跟踪的鲁棒性，提出一种基于相关滤波的自适应特征融合目标跟踪算法。在HOG特征基础上，增加HSV颜色概率直方图，以此获得准确的位置预测。然后分别训练颜色名和HOG特征，并根据两个响应图的峰值自适应地分配融合系数，进而基于尺度池方法，采用多通道特征实现目标的尺度估计。模型的高置信度更新由两个响应图的平均峰值相关能量（APCE）实现。与五种主流相关滤波跟踪算法在两个基准数据集上进行对比实验，实验结果表明在快速运动、复杂背景、光照变化、尺度变化等复杂跟踪场景下，该算法表现出较好的准确性和稳健性。

关键词: 目标跟踪, 相关滤波, 特征融合, 自适应加权